Framework Ga Conf
Content
Velammal Silver Jubilee Celebrations 1986‐2010
Proceedings of the International Conference On COMPUTERS, COMMUNICATION & INTELLIGENCE
22nd & 23rd July 2010 Organised by
Velammal College of Engineering & Technology
Viraganoor, Madurai – 625 009, India
International Conference on Computers, Communication & Intelligence 22nd & 23rd July 2010 CONFERENCE ORGANIZATION Chief Patron : Shri. M.V. Muthuramalingam, Chairman Organising Chair : Dr. N. Suresh Kumar, Principal Organising Secretaries : Dr. P. Alli and Dr. G. Manikandan
INTERNATIONAL ADVISORY COMMITTEE MEMBERS Dr. Kimmo Salmenjoki, Seinajoki University of Applied Sciences, Finland Dr. Fiorenzo Fianceschini, Polytechnico di Torino, Italy Dr. Namudri Kamesh, University of North Texas, USA Dr. Henry Selvaraj, University of Nevada, USA Dr. Kasim Mousa Al Aubidy, Philadelphia University, Jordan Dr. Jiju Antony, University of Strathclyde, UK Dr. Lorno Uden, University of Staffordshire, UK Dr. A. Vallavaraj, Caledonian College of Engineering, Sultanate of Oman Dr. Paulraj Murugesa Pandiyan, University of Malaysia, Perlis Dr. Arunagiri, Yanbu Industrial College, Kingdom of Saudi Arabia. Dr. Raja Sooriya Moorthi, Carnegie Melon University, USA Dr. Angappa Gunasekaran, University of Massachusetts, USA Dr. Sridhar Arjunan, RMIT University , Australia,
NATIONAL ADVISORY COMMITTEE MEMBERS Cmdr. Suresh Kumar Thakur, NRB, DRDO, India Dr. M. Mathirajan, Anna University, India Dr. Chitra T. Rajan, PSG Tech, India Dr. G. Arumugam, MKU, India Dr. S. Ibrahim Sadhar, Wins Infotech Pvt,. Ltd, India Dr. T. Devi, Bharathiar University, India Dr. L. Ganesan, AC Tech, India Dr. T. Purushothaman, GCT, India Dr. R. Murugesan, MKU, India Dr. B. Ramadoss, NIT, India Dr. S. Mercy Shalini, TCE, India Dr. Kannan Balasubramanian, MSEC, India Dr. K. Muneeswaran, MSEC, India Dr. K. Ramar, NEC, India Mr. Jegan Jothivel, Cisco Networking Academy, India
PAPER ID
PAPER TITLE
SESSION ID
PAGE NO. IN PROCEEDINGS
SESSION 1 AI002
Study of similarity metrics for genomic data using go-ontology
S1-01
87 - 94
AI005
Hybrid PSO based neural network classifier and decision tree for brain MRI mining Gap: genetic algorithm based power estimation technique for behavioral circuits Human action classification using 3d star skeletonization and rvm classifier Enhanced knowledge base representation technique for intelligent storage and efficient retrieval using knowledge based markup lFace detection using wavelet transform and rbf neural network
S1-02
95 - 100
S1-03
101 - 107
S1-04
108 - 115
S1-05
154 - 157
S1-06
303 - 306
Automated test case generation and performance analysis for GUI application The need of the hour – nosql technology for next generation data storage Intelligent Agent based Data Cleaning to improve the Accuracy of WiFi Positioning System Using Geographical Information System (GIS) Designing Health Care Forum using Semantic Search Engine & Diagnostic Ontology Optimization of Tool Wear in Shaping Process by Machine vision system Using Genetic Algorithm Framework for Comparison of Association Rule Mining using Genetic Algorithm A New Method For Solving Fuzzy Linear Programming With TORA
S1-07
178 - 187
S1-08
188 - 192
S1-09
24 - 30
S1-10
77 - 81
S1-11
453 - 456
S1-12
14 - 20
S1-13
237 - 239
S2-01
116 - 122
S2-02
316 - 319
COMN018
Relevance vector machine based gender classification using gait appearance features An energy efficient advanced data compression and decompression schemes for wsn Active noise control: a simulation study
S2-03
320 - 325
AI013
A survey on gait recognition using hmm model
S2-04
123 - 126
COMN022
Human motion tracking and behavior classification using multiple cameras
S2-05
131 - 134
AI006 AI007 COMP002 AI009 COMP016 COMP017 COMP115 COMP102 COMP135 COMP111 COMP114
SESSION 2 AI008 COMN013
COMP007
Adaptive visible watermarking and copy protection of reverted multimedia data
S2-06
168 - 173
COMP026
Hiding sensitive frequent item set by database extension
S2-07
357 - 362
COMP027
Integrated biometric authentication using finger print and iris matching Improvement towards efficient OPFET detector
S2-08
436 - 441
S2-09
417 - 420
S2-10
326 - 329
S2-11
369 - 376
COMN034
Texture segmentation method based on combinatorial of morphological and statistical operations using wavelets High Performance Evaluation of 600-1200V, 1-40A Silicon Carbide Schottky Barrier Diodes and Their Applications Using Mat L b based shape matching for trademarks retrieval Phase
S2-12
149 - 153
COMP146
The Medical Image segmentation
S2-13
212 - 215
S3-01
307 - 315
S3-02
158 - 167
S3-03
464 - 467
S3-04
174 - 177
S3-05
334 - 336
S3-06
442 - 452
Secure Multiparty Computation Based Privacy Preserving Collaborative Data Mining Towards Energy Efficient Protocols For Wireless Body Area Networks A cascade data mining approach for network anomaly Detection system Rule Analysis Based On Rough Set Data Mining Technique
S3-07
66 - 70
S3-08
207 - 211
S3-09
377 - 384
S3-10
291 - 296
On the Investigations of Design, Implementation, Performance and Evaluation issues of a Novel BD-SIIT Stateless IPv4/IPv6 T Role l of IPv6 over Fiber (FIPv6): Issues, Challenges and its The Impact on Hardware and Software. Entrustment based authentication protocol for mobile systems.
S3-11
260 - 269
S3-12
270 - 277
S3-13
389 - 392
COMP103 COMN028 COMP118
SESSION 3 AI014 COMP006 COMP142 COMP008 COMP013 COMP032 COMP038 COMP133 COMP119 COMP124 COMP128 COMP129 COMP137
An clustering approach based on functionality of genes for microarray data to find meaningful associations i i web based personalization of e-learning courseware using Semantic concept maps and clustering Modeling of Cutting Parameters for Surface Roughness in Machining A web personalization system for evolving user profiles in dynamic web sites based on web usage mining techniques and agent h l actionable knowledge within the organization using rough Creating set computing A new frame work for analyzing document clustering algorithms
SESSION 4 COMP022
Exploiting parallelism in bidirectional dijkstra for shortest-path computation Cld for improving overall throughput in wireless networks
S4-01
351 - 356
S4-02
46 - 49
S4-03
127 - 130
S4-04
6 - 13
S4-05
135 - 138
S4-06
139 - 148
S4-07
421 - 425
COMN027
Congestion management routing protocol in mobile adhoc networks Performance improvement in ad hoc networks using dynamic addressing Hierarchical zone based intrusion detection system for mobile adhoc networks. Implementing High Performance Hybrid Search Using CELL Processor Enhancing temporal privacy and source-location privacy in wsn routing by fft based data perturbation method Mixed-radix 4-2 butterfly fft/ifft for wireless communication
S4-08
203 - 206
COMP035
NTRU - public key cryptosystem for constrained memory devices
S4-09
55 - 59
COMP036
A novel randomized key multimedia encryption algorithm secure against several attacks Denial Of Service: New Metrics And Their Measurement
S4-10
60 - 65
S4-11
363 - 368
Fpga design of application specific routing algorithms for network on chip Selection of checkpoint interval in coordinated checkpointing protocol for fault tolerant open-mpi
S4-12
330 - 333
S4-13
216 - 223
S5-01
240 - 245
S5-02
343 - 350
S5-03
21 - 23
COMP024
Latest Trends and Technologies in Enterprise Resource Planning – ERP Integrating the static and dynamic processes in software development Content management through electronic document management system A Multi-Agent Based Personalized e-Learning Environment
S5-04
397 - 401
COMP109
Architecture Evaluation for Web Service Security Policy
S5-05
284 - 290
COMP110
Harmonics In Single Phase Motor Drives And Power Conservation.
S5-06
412 - 416
COMP030
Identification in the e-health information systems
S5-07
402 - 405
COMP138 COMN005 COMN020 COMN023 COMN024 COMN026
COMP037 COMP012 COMP019
SESSION 5 COMP116 COMP020 COMP023
COMP033
S5-08
297 - 302
S5-09
426 - 431
COMN033
A Robust Security metrics for the e-Healthcare Information Systems Theoretical Investigation Of Size Effect On The Thermal Properties Of Nanoparticles An efficient turbo coded ofdm system
S5-10
193 - 198
COMP018
Compval – a system to mitigate sqlia
S5-11
337 - 342
COMP126
Fault Prediction Using Conceptual Cohesion in Object Oriented System A Framework for Multiple Classifier Systems Comparison (MCSCF) A Comparative Study of Various Topologies and its performance analysis using WDM Networks
S5-12
256 - 259
S5-13
31 - 40
S5-14
457 - 463
S6-01
246 - 255
S6-02
224 - 229
S6-03
41 - 45
S6-04
1-5
COMP130
MRI Mammogram Image Segmentation using N Cut method and Genetic Algorithm with partial filters Localized Cbir for indexing image database
S6-05
278 - 283
COMP021
Particle swarm optimization algorithm in grid computing
S6-06
50 - 54
COMP149
Advancement in mobile technology Using BADA Enhancing the Life Time of Wireless Sensor Networks Using Mean Measure Mechanism Cloud Computing And Virtualization
S6-07
199 - 202
S6-08
82 - 86
S6-09
230 - 236
S6-10
432 - 435
COMP101
Dynamic Key Management to minimize communication latency for efficient group communication Towards Customer Churning Prevention through Class Imbalance
S6-11
71 - 76
COMP120
Membrane Computing - an Overview
S6-12
385 - 388
COMP148
Privacy Preserving Distributed Data Mining Using Elliptic Curve Cryptography Modeling A Frequency Selective Wall For Indoor Wireless Environment
S6-13
406 - 411
S6-14
393 - 396
COMP147
COMP127 COMP150
SESSION 6 COMP117 COMP112 COMP123 COMP125
COMP107 COMP113 COMP121
COMP149
A New Semantic Similarity Metric for Handling all Relations in WordNet Ontology Simplification of diagnosing disease through microscopic images of blood cells Efficient Apriori Hybrid Algorithm For Pattern Extraction Process
Paper Index Sl. No 1.
2. 3. 4. 5.
6. 7. 8. 9. 10. 11.
12. 13. 14. 15.
16. 17.
18. 19. 20.
Title MRI Mammogram Image Segmentation using NCut method and Genetic Algorithm with partial filters A.Pitchumani Angayarkanni Performance Improvement in Ad Hoc Networks Using Dynamic Addressing S.Jeyanthi & N.Uma Maheswari Framework for Comparison of Association Rule Mining using Genetic Algorithm K.Indira & S.Kanmani Content Management through Electronic Document Management System T.Vengattaraman, A.Ramalingam & P.Dhavachelvan Intelligent Agent based Data Cleaning to improve the Accuracy of WiFiPositioning System Using Geographical Information System (GIS) T.Joshva Devadas A Framework for Multiple Classifier Systems Comparison (MCSCF) P.Shanmugapriya & S.Kanmani Efficient Apriori Hybrid Algorithm For Pattern Extraction Process J.Kavitha, D.Magdalene Delighta Angeline & P.Ramasubramanian CLD for Improving Overall Throughput in Wireless Networks Dr. P. Seethalakshmi & Ms. A. Subasri Particle Swarm Optimization Algorithm In Grid Computing Mrs.R.Aghila, M.Harine & G.Priyadharshini NTRU - Public Key Cryptosystem For Constrained Memory Devices V.Pushparani & Kannan Balasubramaniam A Novel Randomized Key Multimedia Encryption Algorithm Security Against Several Attacks S. Arul Jothi Secure Multiparty Computation Based Privacy Preserving Collaborative Data Mining J.Bhuvana & Dr.T.Devi Towards Customer Churning Prevention through Class Imbalance M.Rajeswari & Dr.T.Devi Designing Health Care Forum Using Semantic Search Engine & Diagnostic Ontology Prof.Mr.V.Shunmughavel & Dr.P.Jaganathan An Enhancing the Life Time of Wireless Sensor Networks Using Mean Measure Mechanism P.Ponnu Rajan & D.Bommudurai Study of Similarity Metrics for Genomic Data Using GO-Ontology V.Annalakshmi,R. Priyadarshini &V. Bhuvaneshwari Hybrid PSO based neural network classifier and decision tree for brain MRI mining Dr.V.Saravanan & T.R.Sivapriya GAP: Genetic Algorithm based Power Estimation Technique for Behavioral Circuits Johnpaul C. I, Elson Paul & Dr. K. Najeeb Human Action Classification Using 3D Star Skeletonization and RVM Classifier Mrs. B. Yogameena, M. Archana & Dr. (Mrs) S. Raju Abhaikumar Relevance Vector Machine Based Gender Classification using Gait Appearance Features Mrs. B. Yogameena, M. Archana & Dr. (Mrs) S. Raju Abhaikumar
Page No. 1-5
6-13 14-20 21-23 24-30
31-40 41-45 46-49 50-54 55-59 60-65
66-70 71-76 77-81 82-86
87-94 95-100
101-107 108-115 116-122
21. A Survey on Gait Recognition Using HMM Model M.Siva Sangari & M.Yuvaraju 22. Congestion Management Routing Protocol In Mobile ADHOC Networks A. Valarmathi1 & RM. Chandrasekaran 23. Human Motion Tracking And Behaviour Classification Using Multiple Cameras M.P.Jancy & B.Yogameena 24. Hierarchical Zone Based Intrusion Detection System for Mobile Adhoc Networks. D G Jyothi & S.N Chandra shekara 25. Implementing High Performance Hybrid Search Using CELL Processor Mrs.Umarani Srikanth 26. Phase Based Shape Matching For Trademarks Retrieval B.Sathya Bama, M.Anitha & Dr.S.Raju 27. Enhanced Knowledge Base Representation Technique for Intelligent Storage and Efficient Retrieval Using Knowledge Based Markup Language A. Meenakshi, V.Thirunageswaran & M.G. Avenash 28. Semantic Web Based Personalization Of E-Learning Courseware Using Concept Maps And Clustering D.Anitha 29. Adaptive visible watermarking and copy protection of reverted multimedia data S.T.Veena & Dr.K.Muneeswaran 30. A Web Personalization System for evolving user profiles in Dynamic Web Sites based on Web Usage Mining Techniques and Agent Technology G.Karthik, R.Vivekanandam & P.Rupa Ezhil Arasi 31. Automated Test Case Generation and Performance Analysis for GUI Application Ms. A.Askarunisa & Ms. D. Thangamari 32. The Need Of The Hour - NOSQL Technology for Next Generation Data Storage
123-126 127-130 131-134 135-138 139-148 149-153 154-157
158-167
168-173 174-177
178-187 188-192
K.Chitra & Sherin M John
33. An Efficient Turbo Coded ofdm system Prof. Vikas Dhere 34. Advancement In Mobile Technologyusing Bada V.Aishwarya, J.Manibharathi & Dr.S.Durai Raj 35. Mixed-Radix 4-2 Butterfly FFT/IFFT For Wireless communication A.Umasankar & S.Vinayagakarthikeyan 36. Towards Energy Efficient Protocols For Wireless Body Area Networks Shajahan Kutty & J.A. Laxminarayana 37. The Medical Image Segmentation Hemalatha & R.Kalaivani 38. Selection of a Checkpoint Interval in Coordinated Checkpointing Protocol for Fault TolerantOpen MPI P.M.Mallikarjuna Shastry & K. Venkatesh 39. Simplification Of Diagnosing Disease Through Microscopic Images Of Blood Cells Benazir Fathima, K.V.Gayathiri Devi, M.Arunachalam & M.K.Hema 40. Cloud Computing And Virtualization R. Nilesh Madhukar Patil & Mr. Shailesh Somnath Sangle 41. A New Method For Solving Fuzzy Linear Programming With TORA S. Sagaya Roseline , A. Faritha Asma & E.C. Henry Amirtharaj 42. Latest Trends And Technologies In Enterprise Resource Planning – Erp B.S.Dakshayani
193-198 199-202 203-206 207-211 212-215 216-223
224-229 230-236 237-239 240-245
43. A New Semantic Similarity Metric for Handling all Relations in WordNet Ontology K.Saruladha, Dr.G.Aghila & Sajina Raj 44. Fault Prediction Using Conceptual Cohesion in Object Oriented System V.Lakshmi, P.V.Eswaripriya, C.Kiruthika & M.Shanmugapriya 45. On the Investigations of Design,Implementation, Performance and Evaluation issues of a Novel BD-SIIT Stateless IPv4/IPv6 Translator J.Hanumanthappa, D.H.Manjaiah & C.V.Aravinda 46. The Role of IPv6 over Fiber (FIPv6): Issues, Challenges and its Impact on Hardware and Software. J.Hanumanthappa, D.H.Manjaiah & C.V.Aravinda 47. Localized CBIR for Indexing Image Databases D.Vijayalakshmi & P. Vijayalakshmi 48. Architecture Evaluation for Web Service Security Policy B.Joshi.vinayak ,Dr.D.H. Manjaiah ,J. Hanumathappa & Nayak.Ramesh.Sunder 49. Rule Analysis Based On Rough Set Data Mining Technique P.Ramasubramanian, V.Sureshkumar & P.Alli 50. A Robust Security metrics for the e-Healthcare Information Systems Said Jafari, Fredrick Mtenzi, Ronan Fitzpatrick & Brendan O’Shea 51. Face Detection Using Wavelet Transform And Rbf Neural Network M.Madhu, M.Moorthi, S.Sathish Kumar & Dr.R.Amutha 52. An Clustering approach based on Functionality of Genes for Microarray data to find meaningful associations M.Selvanayaki & V.Bhuvaneshwari 53. An Energy Efficient Adavanced Data Compression And Decompression Schemes For Wsn G.Mohanbabu#1, Dr.P.Renuga#2 54. Active Noise Control: A Simulation Study Sivadasan Kottayi & N.K. Narayanan 55. Texture Segmentation Method Based On Combinatorial Of Morphological And Statistical Operations Using Wavelets V.Vijayapriya & Prof.K.R.Krishnamoorthy 56. FPGA Design Of Routing Algorithms For Network On Chip R.Anitha & Dr.P.Renuga 57. Creating Actionable Knowledge within the Organization using Rough set computing Mr.R.Rameshkumar, Dr.A.Arunagiri, Dr.V.Khanaa & Mr.C.Poornachandran 58. COMPVAL – A system to mitigate SQLIA S. Fouzul Hidhaya & Dr. Angelina Geetha 59. Integrating the Static and Dynamic Processes in Software Development V. Hepsiba Mabel, K. Alagarsamy & S. Justus 60. Exploiting Parallelism in Bidirectional Dijkstra for Shortest-Path Computation R.Kalpana, Dr. P.Thambidurai, R. Arvind kumar, S. Parthasarathi & Praful Ravi 61. Hiding Sensitive Frequent Item Set by Database Extension B. Mullaikodi & Dr. S.Sujatha 62. Denial Of Service:New Metrics And Their Measurement Dr.KannanBalasubramanian & P.Kavithapandian 63. High Performance Evaluation of 600-1200V, 1-40A Silicon Carbide Schottky Barrier Diodes and Their Applications Using Mat Lab K.Manickavasagan 64. A Cascade Data Mining Approach for Network Anomaly Detection System C. Seelammal
246-255 256-259 260-269
270-277
278-283 284-290 291-296 297-302 303-306 307-315
316-319 320-325 326-329
330-333 334-336 337-342 343-350 351-356 357-362 363-368 369-376
377-384
65. Membrane Computing - an Overview R.Raja Rajeswari & Devi Thirupathi 66. Entrustment Based Authentication Protocol For Mobile Systems. R.Rajalakshmi & R.S.Ponmagal 67. Modeling A Frequency Selective Wall For Indoor Wireless Environment. Mrs. K.Suganya, Dr.N.Suresh Kumar & P.Senthil Kumar 68. A Multi-Agent Based Personalized e-Learning Environment T. Vengattaraman, A. Ramalingam, P. Dhavachelvan & R.Baskaran 69. Identification in the E-Health Information Systems Ales Zivkovic 70. Privacy Preserving Distributed Data Mining Using Elliptic Curve Cryptography M.Rajalakshmi & T.Purusothaman 71. Harmonics In Single Phase Motor Drives And Energy Conservation Mustajab Ahmed Khan & Dr.A.Arunagiri 72. Improvement towards efficient OPFET detector Jaya V. Gaitonde & Rajesh B. Lohani 73. Enhancing Temporal Privacy and Source-Location Privacy in WSN Routing by FFT Based Data Perturbation Method R.Prasanna Kumar & T.Ravi 74. Theoretical nvestigation of size effect on the thermal properties of nanoparticles K.Sadaiyandi & M.A.Zafrulla Khan 75. Dynamic Key Management to minimize communication latency for efficient group communication Dr.P.Alli ,G.Vinoth Chakkaravarthy & R.Deepalakshmi 76. Integrated Biometric AuthenticationUsing Fingerprint and IRIS Matching A.Muthukumar & S.Kannan 77. New Framework for Analyzing Document Clustering Algorithms Mrs. J. Jayabharathy & Dr. S. Kanmani 78. Optimization of Tool Wear in Shaping Process by Machine vision system Using Genetic Algorithm S.Palani, G.Senthilkumar, S.Saravanan & J.Ragunesan A Comparative Study of Various Topologies and its performance analysis using WDM Networks 79. P. Poothathan, S. Devipriya & S. John Ethilton 80. Modeling of Cutting Parameters for Surface Roughness in Machining M. Aruna & P. Ramesh Kumar
385-388 389-392 393-396 397-401 402-405 406-411 412-416 417-420 421-425
426-431 432-435
436-441 442-452 453-456
457-463 464-467
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
MRI Mammogram Image Segmentation Using Ncut Method And Genetic Algorithm With Partial Filters (1)
S.Pitchumani Angayarkanni M.C.A,M.Phil,Ph.d Lecturer,Department of Computer Science, Lady Doak College, Madurai [email protected]
ABSTRACT: Cancer is one of the most common leading deadly diseases which affect men and women around the world. Among the cancer diseases, breast cancer is especially a concern in women. It has become a major health problem in developed and developing countries over the past 50 years and the incidence has increased in recent years. Recent trends in digital image processing are CAD systems, which are computerized tools designed to assist radiologists. Most of these systems are used for automatic detection of abnormalities. However, recent studies have shown that their sensitivity is significantly decreased as the density of breast increases. In this paper , the proposed algorithm uses partial filters to enhance the images and the Ncut method is applied to segment the malignant and benign regions , futher genetic algorithm is applied to identify the nipple position followed by bilateral subtraction of the left and the right breast image to cluster the cancerous and non cancerous regions. The system is trained using Back Propagation Neural Network algorithm. Computational efficiency and accuracy of the proposed system are evaluated based on the Frequency Receiver Operating Characteristic curve(FROC). The algorithm are tested on 161 pairs of digitized mammograms from MIAS database. The Receiver Operating Characteristic curve leads to 99.987% accuracy in detection of cancerous masses. Keywords: Filters, Normalized Cut, Segmentation, BPN, Genetic Algorithm and FROC.
INTRODUCTION: Breast cancer is one of the major causes for the increased mortality among women especially in developed countries. It is second most common cancer in women. The World Health Organization’s International estimated that more than 1,50,000 women worldwide die of breast cancer in year. In India, breast cancer accounts for 23% of all the female cancer death followed by cervical cancer which accounts to 17.5% in India. Early detection of cancer leads to significant improvements in conservation treatment. However, recent studies have shown that the sensitivity of these systems is significantly decreased as the density of the breast increased
Velammal College of Engineering and Technology, Madurai
while the specificity of the systems remained relatively constant. In this work we have developed automatic neuron genetic algorithmic approach to automatically detect the suspicious regions on digital mammograms based on asymmetries between left and right breast image. One of the major tool used for early detection of breast cancer is mammography. Mammography offers high quality images at low radiation doses and is the only widely accepted imaging method for routine breast cancer screening. Although mammography is widely used around the world for breast cancer detection, there are some difficulties when mammography is used for diagnosing breast cancer. One of the difficulties with mammography is that mammograms generally have low contrast compared with normal breast structure, and thus make it difficult for radiologists to interpret them. Studies show that the interpretation of mmaograms by radiologists could result in high rate of falsepositive and false-negative. This difficulty has caused high proportion of women without cancers to undergo breast biopsies and miss the breast treatment time. Several solutions were proposed in the past to increase accuracy and sensitivity of mammography and reduce unnecessary biopsies. Double reading of mmamograms is one of the solutions and has been advocated to reduce the proportion of missed cancers. The basic idea for double reading is to read the mammograms by two radiologists. However this solution is both costly and time consuming.Instead CAD has drawn attention from both computer scientists and radiologists in the interpretation of mammograms. CAD which integrates computer science, image processing , pattern recognition and artificial intelligence technologies can be defined as a diagnosis that is made by a radiologist who uses the output from a computerized analysis of medical images as a “second opinion” in detecting lesions and in making diagnostic decisions. It has been proven that this kind of system can improve the accuracy of breast diagnosis for early prediction of breast cancer. Computer aided breast cancer detection system is especially useful
Page 1
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
when the radiologist become tired of screening mammograms. In the CAD System for breast cancer, the detection of abnormal regions , such as calcification, mass and architectural distortion is the central task and the performance of a CAD system will depend on the performance of the detection of these abnormalities. There have been many proposed algorithms for detection of these abnormalities. In this paper we have introduced the detection of microcalcifications. As one of the early signs of breast cancer , mirocalcifications are tiny granule like deposits of calcium, which appear as small bright spots of mmaograms. Their size varies from 0.1 mm to 1mm. “Cluster: of MCs is defines as a group of three to five MCs within regions. Generally microcalcification clusters are important indication of possible cancer. This algorithm effectively and automatically detect MCs . 2. ALGORITHM DESIGN: There are four steps involved in the algorithm for the detection MCCs which is shown in the figure.
Thus, the product of the image matrix, which is usually very large because it represents the initial image (pixel table), by the filter yields a matrix corresponding to the processed image. 2.1.1 HIGH PASS FILTER: It allow high frequency areas to pass with the resulting image having greater detail resulting in a sharpened image. The boundary information of the enhanced image was extracted for visual evaluation. A high-pass (laplacian) filter was used for this purpose.
Figure 2: Mammogram Image enhanced using high pass filter
The table coefficients determine the properties of the filter. The following is an example of a 3 X 3 filter:
2.1.2) LOW PASS FILTER: Low pass filtering, otherwise known as "smoothing", is employed to remove high noise from a digital image. Noise is often introduced during the analog-to-digital conversion process as a side-effect of the physical conversion of patterns of light energy into electrical patterns . There are several common approaches to removing this noise: • If several copies of an image have been obtained from the source, some static image, then it may be possible to sum the values for each pixel from each image and compute an average. This is not possible, however, if the image is from a moving source or there are other time or size restrictions. • If such averaging is not possible, or if it is insufficient, some form of low pass spatial filtering may be required. There are two main types: • reconstruction filtering, where an image is restored based on some knowledge of the type of degradation it has undergone. Filters that do this are often called "optimal filters". • enhancement filtering, which attempts to improve the (subjectively measured) quality of an image for human or machine interpretability. Enhancement filters are generally heuristic and problem oriented One of the most important problems in image processing is denoising. Usually the procedure used for denoising, is dependent on the features of the image, aim of processing and also post-processing algorithms [5]. Denoising by lowpass filtering not only reduces the noise but also blurs the edges.
Velammal College of Engineering and Technology, Madurai
Page 2
Parti al Filter
Feature Extraction using NCut Segmentation
Genetic Algorith m
Multilay ered BPN
Fig 1: Flow Chart of Algorithm
2.1 PARTIAL FILTER FOR IMAGE ENHANCEMENT: A filter is a mathematical transformation (called a convolution product) which allows the value of a pixel to be modified according to the values of neighbouring pixels, with coefficients, for each pixel of the region to which it is applied. The filter is represented by a table (matrix), which is characterized by its dimensions and its coefficients, whose centre corresponds to the pixel concerned. 1
1
1
1
4
1
1
1
1
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Spatial and frequency domain filters are widely used as tools for image enhancement. Low pass filters smooth the image by blocking detail information. Mass detection aims to extract the edge of the tumor from surrounding normal tissues and background, high pass filters (sharpening filters) could be used to enhance the details of images.
method
the
Microcalcifications
are
clustered. Figure 4: After Normalized Cut Segmentation
Figure 3: Mammogram Image Enhanced Using Low Pass filter
2.2 IMAGE SEGMENTATION: The goal of image segmentation is to cluster pixels into salient image regions, i.e., regions corresponding to individual surfaces, objects, or natural parts of objects. In this we apply Normalized Cut method of segmentation to cluster microcalcification regions. Finally we outline the normalized cut approach of Shi and Malik [13].Here we seek a partition F and G = V − F of the affinity weighted,undirected graph (without source and sink nodes). In order to avoid partitions where one of F or G is a tiny region, Shi and Malik propose the normalized cut criterion, namely that F and G should minimize.
The computational efficiency 12.563 seconds on the 160x160 image. 2.3 GENETIC ALGORITHM: A partial filtering absed normalized cut method is used to generate a image to separate the breast and the non breast region . The GA enhances the breast border . Border detector detects the edges in the binary images , where each pixel takes on either the intensity value of zero for a non border pixel or one for border pixel. Each pixel in the binary map corresponds to an underlying pixel in the original image . In this proposed system , kernel is extracted from border points as a neighborhood array of pixels of the size 3*3 window of binary image. The binary kernels are considered population strings for GA. The corresponding kernels are extracted from gray level mammogram image using spatial coordinate points and the sum of the intensity values are considered as the fitness value . After identifying initial population and the fitness value , the genetic operator can be applied to generate a new population. Reproduction operator produces new string for crossover. Reproduction is implemented as linear search through roulette wheel with slots weighted in proportion to kernel fitness values. In this function, a random number multiplies the sum of population fitness called as stopping point.
Figure 5: GA
Note any segmentation technique can be used for generating proposals for suitable regions F, for which N(F, V − F) could be evaluated. Indeed, the SMC approach above can be viewed as using S and T to provide lower bounds on the terms L(F, V ) and L(G, V ) (namely L(S, V ) and L(T, V ), respectively), and then using the S-T min cut to globally minimize L(F,G) subject to S C F and T C G. Using this
Velammal College of Engineering and Technology, Madurai
Page 3
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.4 GENERATING THE ASYMMETRIC IMAGE: After the images were aligned, bilateral subtraction was performed [47,48] by subtraction was performed by subtracting the digital matrix of the left breast image from the digital matrix of the right breast image. Microcalcification in the right breast image have positive pixel values in the image obtained after subtraction, while microcalcification in the left breast image have negative pixel values in the subtracted image. As a result, two new images were generated: one with positive values and the other with negative values. The most common gray level was zero, which indicated no difference between the left and right images. Simple linear stretching of the two generated images to cover the entire available range of 1024 gray levels was then calculated. The difference between corresponding pixels contains important information that can be used to discriminate between normal and abnormal tissue. The asymmetry image can be thresholded to extract suspicious regions. To generate FROC curve, the asymmetry image is thresholded using ten different intensity values ranges from 50-150. Figure 6 shows a asymmetry image and connected regions extracted based on thresholding to obtain a progressively larger number of high difference pixels.
Figure 6
Figure 7a) Steps involved in automated Classification using Ant Colony Optimization
2.6. ROC CURVE: Finally the technique was evaluated on the mammograms randomly selected from the non-suspicious section of the data base. The method outlined small regions in 5 out of the 15 non suspicious mammograms. The areas identified were generally very small compared to those in abnormal mammograms
Asymmetric images
Two different techniques are used in the interpretation of mammogram. The first technique consists of systematic search of each mammogram for visual pattern symptomatic tumors. Such as, a bright, approximately circular blob with hazy boundary might indicate the presence of a circumscribed mass. The second technique, the asymmetric approach , consists of systematic comparison of corresponding regions in the left and the right breast. 2.5. BPN TRAINING: In addition, a backpropagation artificial neural network (BPANN) was also developed and evaluated on the same data. The parameters for ANN training were published before. Figure 5 compare the ROC curves for the LGP and the BPANN algorithms respectively. The BP-ANN yielded an ROC area index of Az=0.88±0.01. Our GP approach achieved a statistically significantly better performance with Az=0.91±0.01.
Velammal College of Engineering and Technology, Madurai
Figure 8 Lesion Areas detected for Abnormal and Non-Suspicious cases (large image extracts). [Figures (a) and (b) are presented at different ordinate scales]
Fig 8(a) shows the extracted areas for the abnormal lesions. (Image sequence 54 - 87 are stellate lesions and 74 to 100 are regular masses). We first establish whether these represent two different populations, by applying a Mann-Whitney (Wilcoxon rank sum) non-parametric test, since it is unrealistic to presume any specific underlying distribution. Median values are 450 and 1450 pixels respectively which produce a confidence level of 85% that the two data sequences emanate from distinct populations. Since this is not significant at normally acceptable levels we can compare the abnormals as a single distribution against the nonsuspicious set, Fig 8(b). Using the same test, median values of 5500 and 10 pixels for the two distributions are established, giving a confidence level of greater than 97.5%
Page 4
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
that the two distributions are different, suggesting that our PROTOCOLS ARE AN EFFECTIVE METHOD OF AREA DETECTION. CONCLUSION: The proposed algorithms are tested on 161 pairs of digitized mammograms from Mammographic Image analysis Society(MIAS) database. A free response receiver operating characteristic (FROC) curve is generated for the mean value of the detection rate for all the 161 pairs of mammograms in the MIAS database, to evaluate the performance of the proposed method. There is no doubt that for the immediate future mammography will continue to play a major role in the detection of breast cancer. The ultimate objective of this thesis was to identify tumor or masses in breast tissue. Since hamartomas consists of normal breast tissue with abnormal proportions and the first step was try to identify the different tissue type in mammography with normal breast tissue. The important features have been extracted from the Normalized cut method of the each sub image using various statistical techniques. The Genetic algorithm has been implemented and the breast border was identified from the clustered image. The tests that were carried out using a set of 117 tissues samples, 67 benign and 50 malignant. The result analysis has given a sensitivity of 99.8%, a specificity of 99.9% and an accuracy above 99.9%, which means encouraging results. The preliminary results of this approach are very promising in characterizing breast tissue. REFERENCES: [1] Bosch. A.; Munoz, X.; Oliver.A.; Marti. J., Modeling and Classifying Breast Tissue Density in Mammograms, Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on Volume 2, Issue , 2006 Page(s): 1552 – 15582. [2] Dar-Ren Chena, Ruey-Feng Changb, ChiiJen Chenb, Ming-Feng Hob, Shou-Jen Kuoa, ShouTung Chena, Shin-Jer Hungc, Woo Kyung Moond, Classification of breast ultrasound images using fractal feature, ClinicalImage, Volume 29, Issue4, Pages 234-245. [3] Suri, J.S., Rangayyan, R.M.: Recent Advances in Breast Imaging, Mammography,and Computer-Aided Diagnosis of Breast Cancer. 1st edn. SPIE (2006) [4] Hoos, A., Cordon-Cardo, C.: Tissue microarray pro.ling of cancer specimens and cell lines: Opportunities and limitations. Mod. Pathol. 81(10), 1331–1338 (2001) [5] Lekadir, K., Elson, D.S., Requejo-Isidro, J., Dunsby, C., McGinty, J., Galletly, N.,Stamp, G., French, P.M., Yang, G.Z.: Tissue characterization using dimensionality reduction and .uorescence imaging. In: Larsen, R., Nielsen, M., Sporring, J. (eds.) MICCAI 2006. LNCS, vol. 4191, pp. 586–593. Springer, Heidelberg (2006).
[6] A. Papadopoulos, D. I. Fotiadis, and A. Likas,“An Automatic microcacalcification Detection System Based On a Hybrid Neural Network Classifier,” Artificial Intelligence in Medicine,vol. 25, pp. 149-167, 2002. [7] A. Papadopoulos, D. I. Fotiadis, and A. Likas, “Characterization of Clustered microcalcifications in Digitized Mammograms Using Neural Networks and Support Vector Machine,” Artificial Intelligence in Medicine,vol. 34, pp. 141-150, 2005. [8] R. Mousa, Q. Munib, and A. Moussa, “Breast Cancer Diagnosis System based in Wavelet Analysis and Fuzzy-Neural,” Expert Systems with Applications, vol. 28, pp. 713-723, 2005.
Velammal College of Engineering and Technology, Madurai
Page 5
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Performance Improvement in Ad Hoc Networks Using Dynamic Addressing S.Jeyanthi#1, N.Uma Maheswari*2 #
Lecturer, Computer Science Department PSNA College of Engg & Tech,Dindigul,Tamilnadu,India 1
[email protected]
*
Assistant Professor PSNA College of Engg & Tech,Dindigul,Tamilnadu,India 2
[email protected]
Abstract Dynamic addressing refers to the assignment of IP addresses automatically. In this paper we propose the scalable routing in ad hoc networks. It is well known that the current ad hoc protocol do not scale to work efficiently in networks of more than a few hundred nodes. Most current adhoc routing architectures use flat static addressing and thus, need to keep track of each node individually, creating a massive overhead problem as the network grows. In this paper, we propose that the use of dynamic addressing can enable scalable routing in adhoc networks. We provide an initial design of a routing layer based on dynamic addressing, and evaluate its performance. Each node has a unique permanent identifier and a transient routing address, which indicates its location in the network at any given time. The main challenge is dynamic address allocation in the face of node mobility. We propose mechanisms to implement dynamic addressing efficiently. Our initial evaluation suggests that dynamic addressing is a promising approach for achieving scalable routing in large adhoc and mesh networks.
Keywords — Adhoc networks, Flat static addressing, Dynamic addressing, Unique permanent identifier. I. Introduction
Adhoc
networking technology has advanced tremendously but it has yet to become a widely deployed technology. Ad hoc networks research seems to have downplayed the importance of scalability. In fact, current ad hoc architectures do not scale well beyond a few hundred nodes. Existing Ad Hoc Routing Layers do not support several hundred nodes and lack of scalability. It uses flat static addressing. It creates a massive Routing overhead. It increases searching time (not optimal solution). The easy-touse, self-organizing nature of ad hoc networks make them attractive to a diverse set of applications. Today, these are usually limited to smaller deployments, but if we can solve
Velammal College of Engineering and Technology, Madurai
the scalability problem, and provide support for heterogeneous means of connectivity, including directional antennas, communication lasers, even satellites and wires, ad hoc and mesh-style networking is likely to see adoption in very large networks as well. Large-scale events such as disaster relief or rescue efforts are highly dependent on effective communication capabilities. Such efforts could benefit tremendously from the use of self-organizing networks to improve the communications and monitoring capabilities available. Other interesting candidate scenarios are community networks in dense residential areas, large scale, long-range networks in developing regions, and others, where no central administrator exists, or where administration would prove too costly. The current routing protocols and architectures work well only up to a few hundred nodes. Most current research in ad hoc networks focus more on performance and power consumption related issues in relatively small networks, and less on scalability. The main reason behind the lack of scalability is that these protocols rely on flat and static addressing. With scalability as a partial goal, some efforts have been made in the direction of hierarchical routing and clustering [1] [2] [3]. These approaches do hold promise, but they do not seem to be actively pursued. It appears to us as if these protocols would work well in scenarios with group mobility [4], which is also a common assumption among cluster based routing protocols. We examine that whether dynamic addressing is a feasible way to achieve scalable adhoc routing. By”scalable” we mean thousands up to millions of nodes in an ad hoc or mesh network. With dynamic addressing, nodes change addresses as they move, so that their addresses have a topological meaning. Dynamic addressing simplifies routing but introduces two new problems: address allocation, and address lookup. As a guideline, we identify a set of properties that a scalable and efficient solution must have: Localization of overhead: a local change should affect only the immediate neighborhood, thus limiting the overall overhead incurred due to the change.
Page 6
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Lightweight, decentralized protocols: To avoid the responsibility at any individual node, and to keep the necessary state to be maintained at each node as small as possible. Zero-configuration: To remove the need for manual configuration beyond what can be done at the time of manufacture. Minimal restrictions on hardware: Omni directional link layers do not scale to large networks. Localization technologies, such as GPS, may limit protocol applicability. We present a complete design including address allocation, routing and address lookup mechanisms, and provide thorough evaluation results for the address allocation and routing components. First, we develop a dynamic addressing scheme, which has the necessary properties mentioned above. Our scheme separates node identity from node address, and uses the address to indicate the node’s current location in the network. Second, we study the performance of a new routing protocol, based on dynamic addressing, through analysis and simulations. The address allocation scheme uses the address space efficiently on topologies of randomly and uniformly distributed nodes, empirically resulting in Average routing table size< 2 log2 n Where n is the number of nodes in the network. We describe a new approach to routing in ad hoc networks, and compare it to the current routing architectures. However, the goal is to show the potential of this approach and not to provide an optimized protocol. We believe that the dynamic addressing approach is a viable strategy for scalable routing in ad hoc networks. II. Related Work In most common IP-based ad hoc routing protocols [5] [7] [8], addresses are used as pure identifiers. Without any structure in the address space, there are two choices: either keep routing entries for every node in the network, or resort to flooding route requests throughout the network upon connection setup. However, this approach can be severely limiting as location information is not always available and can be misleading in, among others, non-planar networks. For a survey of ad hoc routing, see [9]. In the Zone Routing Protocol (ZRP) [10] and Fisheye State Routing (FSR) [11], nodes are treated differently depending on their distance from the destination. In FSR, link updates are propagated more slowly the further away they travel from their origin, with the motivation that changes far away are unlikely to affect local routing decisions. In ZRP is a hybrid reactive/ proactive protocol, where a technique called border casting is used to limit the damaging effects of global broadcasts. In multilevel-clustering approaches such as Landmark [12], LANMAR [3], L+ [13], MMWN [1] and Hierarchical State Routing (HSR) [2], certain nodes are elected as cluster
Velammal College of Engineering and Technology, Madurai
heads. These cluster heads in turn select higher level cluster heads, up to some desired level. A node’s address is defined as a sequence of cluster head identifiers, one per level, allowing the size of routing tables to be logarithmic in the size of the network, but easily resulting in long hierarchical addresses. In HSR, for example, the hierarchical address is a sequence of MAC addresses, each of which is 6 bytes long. A problem with having explicit cluster heads is that routing through cluster heads creates traffic bottlenecks. In Landmark, LANMAR and L+, this is partially solved by allowing nearby nodes route packets instead of the cluster head, if they know a route to the destination. Our work is, as far as we know, the first attempt to use this type of addressing in ad hoc networks. Tribe [14] is similar to DART at a high level, in that it uses a two phase process for routing: first address lookup, and then routing to the address discovered. However, the tree-based routing strategy used in Tribe bears little or no resemblance to the area based approach in DART. Tree-based routing may under many circumstances suffer from severe traffic concentration at nodes high up in the tree, and a high sensitivity to node failure. III. Overview of Network Architecture In this section, we present sketch of network architecture shown in figure 1, which could utilize the new addressing scheme effectively. In our approach, we separate the routing address and the identity of a node. The routing address of a node is dynamic and changes with node movement to reflect the node’s location in the network topology. DART
Cluster creation
Address allocation Mapping Distributed lookup table
Routing Figure.1 Overall system design
Page 7
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The identifier is a globally unique number that stays the same throughout the lifetime of the node. For ease of presentation, we can assume for now that each node has a single identifier. When a node joins the network, it listens to the periodic routing updates of its neighboring nodes, and uses these to identify an unoccupied address. The joining node registers its unique identifier and the newly obtained address in the distributed node lookup table. Due to mobility, the address may subsequently be changed and then the lookup table needs to be updated. When a node wants to send packets to a node known only by its identifier, it will use the lookup table to find its current address. Once the destination address is known the routing function takes care of the communication. The routing function should make use of the topological meaning that our routing addresses possess. We start by presenting two views of the network that we use to describe our approach: a) the address tree, and b) the network topology. Address Tree: In this abstraction, we visualize the network from the address space point of view. Addresses are l bit binary numbers, al-1, . . . , a0. The address space can be thought of as a binary address tree of l + 1 level, as shown in figure 2. The leaves of the address tree represent actual node addresses; each inner node represents an address sub tree a range of addresses with a common prefix. • Level 0 sub tree is a single address. • Level 1 sub tree has a 2 bit prefix and can contain up to two leaf nodes. • Level 2 sub tree containing addresses [100] through [111].
each sub tree of the address tree are enclosed with dotted lines. Note that the set of nodes from any sub tree in figure 2 induces a connected sub graph in the network topology in figure 3.
Figure.3 A network topology with node addresses assigned.
The nodes that are close to each other in the address space should be relatively close in the network topology. More formally, we can state the following constraint. Prefix Sub graph Constraint: The set of nodes that share a given address prefix form a connected sub graph in the network topology. This constraint is fundamental to the scalability of our approach. Intuitively, this constraint helps us map the virtual hierarchy of the address space onto the network topology. The longer the shared address prefix between two nodes, the shorter the expected distance in the network topology. Finally, let us define two new terms that will facilitate the discussion in the following sections. A Level-k sub tree of the address tree is defined by an address prefix of (l-k) bits, as shown in figure 2. For example, a Level0 sub tree is a single address or one leaf node in the address tree. A Level-1 sub tree has a (l-1)-bit prefix and can contain up to two leaf nodes. In figure 1, [0xx] is a Level-2 sub tree containing addresses [000] through [011]. Note that every Level-k sub tree consists of exactly two Level-(k - 1) sub trees. We define the term Level-k sibling of a given address to be the sibling of the Level-k sub tree to which a given address belongs. By drawing entire sibling sub trees as triangles, we can create abstracted views of the address tree, as shown in figure 4.
Figure.2 Address tree of 3-bit binary address space.
For presentation purposes, nodes are sorted in increasing address order, from left to right. The actual physical links are represented by dotted lines connecting leaves in fig 3. Network Topology: This view represents the connectivity between nodes. In figure 3, the network from figure 2 is presented as a set of nodes and the physical connections between them. Each solid line is an actual physical connection, wired or wireless, and the sets of nodes from
Velammal College of Engineering and Technology, Madurai
Figure.4 Routing entries corresponding to figure 2. Node 100 has entries for sub trees 0xx, 11x (null entry) and 101.
Here, we show the siblings of all levels for the address [100] as triangles: the Level-0 sibling is [101], Level-1 is [11x], and the Level-2 sibling is [0xx]. Note that each address has exactly one Level-k sibling, and thus at most l siblings in total.
Page 8
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Finally, we define the identifier of a sub tree to be the min of the identifiers of all nodes that have addresses from that sub tree. In cases where the prefix sub graph constraint is temporarily violated, two disconnected instances of the address sub tree exist in the network. In this case, each instance is uniquely identified by the min of the subset of identifiers that belongs to its connected sub graph. Our addressing and routing schemes have several attractive properties. First, they can work with omni directional and directional antennas as well as wires. Second, we do not need to assume the existence of central servers or any other infrastructure, nor do we need to assume any geographical location information, such as GPS coordinates. However, if infrastructure and wires exist, they can, and will, be used to improve the performance. Third, we make no assumptions about mobility patterns, although high mobility will certainly lead to increased overhead and decreased throughput. Finally, since our approach was designed primarily for scalability, we do not need to limit the size of the network; most popular ad hoc routing protocols today implicitly impose network size restrictions. IV. Routing In this work, we use a form of proactive distance-vector routing, made scalable due to the hierarchical nature of the address space. Although we have chosen to use distance vector routing, we would like to point out that many of the advantages of dynamic addressing can be utilized by a linkstate protocol as well. Each node keeps some routing state, routing state about a node’s Level-i sibling is stored at position i in each of the respective arrays. The routing state for a sibling contains the information necessary to maintain a route toward a node (any node) in that sub tree. The address field contains the current address of the node, and bit i of the address is referred to as address[i], where i = 0 for the least significant bit of the address. Arrays next hop and cost are self-explanatory. The id array contains the identifier of the sub tree in question. As described earlier, the identifier of a sub tree is equal to the lowest out of all the identifiers of the nodes that constitute that sub tree. Finally, route log[i] contains the log of the current route to the sibling at level i, where bit b of log i is referenced by the syntax route log[i][b]. To identify the most significant bit that differs between the current node’s address and the destination address. In this case, the most significant differing bit is bit number 2. The node then looks up the entry with index two in the next hop table, and then sends the packet there. In our example, this is the neighbor with address [011]. The process is repeated until the packet has reached the given destination address. The hierarchical technique of only keeping track of sibling sub trees rather than complete addresses has three immediate benefits. One,
Velammal College of Engineering and Technology, Madurai
the amount of routing state kept at each node is drastically reduced. Two, the size of the routing updates is similarly reduced. Three, it provides an efficient routing abstraction such that routing entries for distant nodes can remain valid despite local topology changes in the vicinity of these nodes. A. Loop Avoidance DART uses a novel scheme for detecting and avoiding routing loops, which leverages the hierarchical nature of the address space to improve scalability. A simple way of implementing this is to concatenate a list of all visited nodes in the routing update, and to have nodes check this list before accepting an update. However, this approach has a scalability problem, in that routing updates will quickly grow to unwieldy sizes. Instead, DART makes use of the structured address space to create a new kind of loop avoidance scheme. In order to preserve scalability, we generalize the loop freedom rule above. For each sub tree, once a routing entry has left the sub tree, it is not allowed to re-enter. This effectively prevents loops, and can be implemented in a highly scalable manner. V. Node Lookup We propose to use a distributed node lookup table, which maps each identifier to an address, similar to what we proposed in [5]. Here, we assume that all nodes take part in the lookup table, each storing a few7 <identifier, address> entries. However, this node lookup scheme is only one possibility among many, and more work is needed to determine the best lookup scheme to deploy. For our proposed distributed lookup table, the question now arises: which node stores a given <identifier, address> entry? Let us call this node the anchor node of the identifier. We use a globally, and a priori, known hash function that takes an identifier as argument and returns an address where the entry can be found. If there exists a node that occupies this address, then that node is the anchor node. If there is no node with that address, then the node with the least edit distance between its own address and the destination address is the anchor node. To route packets to an anchor node, we use a slightly modified routing algorithm: If no route can be found to a sibling sub tree indicated by a bit in the address, that bit of the address is ignored, and the packet is routed to the sub tree indicated by the next (less significant) bit. When the last bit has been processed, the packet has reached its destination. This method effectively finds the node with the address minimum edit distance to the address returned by the hash function. For example, using figure 3 for reference, let’s assume a node with identifier ID1 has a current routing address of [010]. This node will periodically send an updated entry to the lookup table, namely <ID1, 010>. To figure out where to send the entry, the node uses the hash function to calculate an address, like so: hash (ID1). If the returned address is [100], the packet will simply be routed to the node with that address. However, if the returned address
Page 9
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
was instead [111], the packet could not be routed to the node with address [111] because there is no such node. In such a situation, the packet gets automatically routed to the node with the most similar address, which in this case would be [101] A. Improved Scalability We would like to stress that all node lookup operations use unicast only: no broadcasting or flooding is required. This maintains the advantage of proactive and distance vector based protocols over on-demand protocols: the routing overhead is independent of how many connections are active. When compared with other distance vector protocols, our scheme provides improved scalability by drastically reducing the size of the routing tables, as we described earlier. In addition, updates due to a topology change are in most cases contained within a lower level sub tree and do not affect distant nodes. This is efficient in terms of routing overhead. To further improve the performance of our node lookup operations, we envision using the locality optimization technique described in [5]. Here, each lookup entry is stored in several locations, at increasing distance from the node in question. By starting with a small, local lookup and gradually going to further away locations, we can avoid sending lookup requests across long distances to find a node that is nearby. B. Coping with Temporary Route Failures On occasion, due to link or node failure, a node will not have a completely accurate routing table. This could potentially lead to lookup packets, both updates and requests, terminating at the wrong node. The end result of this is that requests cannot be promptly served. In an effort to reduce the effect of such intermittent errors, a node can periodically check the lookup entries it stores, to see if a route to a more suitable host has been found. If this should be the case, the entry is forwarded in the direction of this more suitable host. Requests are handled in a similar manner: if the request could not be answered with an address, it is kept in a buffer awaiting either the arrival of the requested information, or the appearance of a route to a node which more closely matches the key requested. This way, even if a request packet arrives at the anchor node before the update has reached it; the request will be buffered and served as soon as the update information is available. C. Practical Considerations Due to the possibility of network partitioning and node failure, it is necessary to have some sort of redundancy mechanism built-in. We have opted for a method of periodic refresh, where every node periodically sends its information to its anchor node. By doing so, the node ensures that if its anchor node should become unavailable, the lookup information will be available once again within one refresh period. Similarly, without a mechanism of expiry, outdated
Velammal College of Engineering and Technology, Madurai
information may linger even after a node has left the network. Therefore, we set all lookup table entries to expire automatically after a period twice as long as the periodic refresh interval. VI. Dynamic Address Allocation To assess the feasibility of dynamic addressing, we develop a suite of protocols that implement such an approach. Our work effectively solves the main algorithmic problems, and forms a stable framework for further dynamic addressing research. When a node joins an existing network, it uses the periodic routing updates of its neighbors to identify and select an unoccupied and legitimate address. It starts out by selecting which neighbor to get an address from the neighbor with the highest level insertion point is selected as the best neighbor. The insertion point is defined as the highest level for which no routing entry exists in a given neighbor’s routing table. However, the fact that a routing entry happens to be unoccupied in one neighbor’s routing table does not guarantee that it represents a valid address choice. We discuss how the validity of an address is verified in the next subsection. The new node picks an address out of a possibly large set of available addresses. In our current implementation, we make nodes pick an address in the largest unoccupied address block. For example, in figure 4, a joining node connecting to the node with address [100] will pick an address in the [11x] sub tree. Figure 5 illustrates the address allocation procedure for a 3-bit address space.
Figure.5 Address allocation procedure for a 3-bit address space
There are several ways to choose among the available addresses, and we have presented only one such method. However, it has turned out that this method of address selection works well in simulation trials. Under steady-state, and discounting concurrency, the presented address selection technique leads to a legitimate address allocation: the joining node is by definition connected to neighbor it got its new address from, and the new address is taken from one of the neighbors’ empty sibling sub trees, so the prefix sub graph constraint is satisfied. Node A starts out alone with address [000]. When node B joins the network, it observes that A has a null routing entry corresponding to the sub tree [1xx], and
Page 10
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
picks the address [100]. Similarly when C joins the network by connecting to B, C picks the address [110]. Finally, when D joins via A, A’s [1xx] routing entry is now occupied. However, the entry corresponding to sibling [01x] is still empty, and so, D takes the address [010]. Merging Networks Efficiently - DART handles the merging of two initially separate networks as part of normal operations. In a nutshell, the nodes in the network with the higher identifier join the other network one by one8. The lower id network absorbs the other network slowly: the nodes at the border will first join the other network, and then their neighbors join them recursively. Dealing with Split Networks - Here, we describe how we deal with network partitioning. Intuitively, each partition can keep its addresses, but one of the partitions will need to change its network identifier. In this situation, there are generally no constraint violations. This reduces to the case where the node with the lowest identifier leaves the network. Since the previous lowest identifier node is no longer part of the network, the routing update from the new lowest identifier node can propagate through the network until all nodes are aware of the new network identifier. VII. Maintaining the Dynamic Routing Table While packet forwarding is a simple matter of looking up a next hop in a routing table, maintaining a consistent routing state does involve a moderate amount of sophistication. In addition to address allocation, loop detection and avoidance is crucial to correct protocol operation. DART nodes use periodic routing updates to notify their neighbors of the current state of their routing table. If, within a constant number of update periods, a node does not hear an update from a neighbor, it is removed from the list of neighbors, and its last update discarded. Every period, each node executes Refresh() function. Refresh() checks the validity of its current address, populates a routing table using the information received from its neighbors, and broadcasts a routing update. When populating the routing table, the entry for each level, i, in the received routing update is inspected is sequence, starting at the top level. For neighbors where the address prefix differs at bit i, we create a new routing entry, with a one-hop distance. It also has an empty route log, with the exception of bit i, which represents the level-i sub tree boundary that was just crossed. The sub tree identifier is computed using the id array. After this, the procedure returns, as the remaining routing information is internal to the neighbor’s sub tree, and irrelevant to the current node. For nodes with the same address prefix as the current node, we go on to inspect their routing entry for level i. First, we ensure that the entry is loop free. If so, then keep the routing entry as long as the identifier of the entry is the same or
Velammal College of Engineering and Technology, Madurai
smaller than what is already in the routing table, and as long as the distance is smaller. VIII. Simulation Results We conduct our experiments using two simulators. One is the well known ns-2 network simulator. In ns-2, we used the standard distribution; version 2.26 used the standard values for the Lucent Wave LAN physical layer, and the IEEE 802.11 MAC layer code, together with patch for a retry counter bug recently identified by Dan Berger at UC Riverside9. For all of the ns-2 simulations, we used the Random Waypoint mobility model with up to 800 nodes and a maximum speed of 5 m/s, a minimum speed of 0.5 m/ s, a maximum pause time of 100 seconds and a warm-up period of 3600 seconds10. The duration of all the ns-2 simulations was 300 seconds11, wherein the first 60 seconds are free of data traffic, allowing the initial address allocation to take place and for the network to thereby organize itself. The size of the simulation area was chosen to keep average node degree close to 9. For example, for a 400-node network, the size of the simulation area was 2800x2800 meters. This was done in order to maintain a mostly connected topology. Mobility parameters were chosen to simulate a moderately mobile network. DART is not suitable for networks with very high levels of mobility, as little route aggregation benefits are to be had when the current location of most nodes bear little relation to where these nodes were a few seconds ago. Our simulations focus on the address allocation and routing aspects of our protocol, not including the node lookup layer, which is replaced by a global lookup table accessible by all nodes in the simulation. The choice of lookup mechanism (for example distributed, hierarchical, replicated, centralized, or out-of-band) should be determined by network characteristics, and performance may vary depending on what mechanism is used. Here follows a summary of our findings. DSDV, due to its periodic updates and flat routing tables, experiences very high overhead growth as the network grows beyond 100 nodes, but nevertheless performs well in comparison with other protocols in the size ranges studied. AODV, due to its reactive nature, suffers from high overhead growth both as the size of the network, and the number of flows, grows. While AODV performs very well in small networks, the trend suggests that it is not recommendable for larger networks .DSR, in our simulations, performed well in small networks, and never experienced high overhead growth, likely due to its route caching functionality. However, due to excessive routing failures, DSR demonstrated unacceptable performance in larger networks. Finally, DART, demonstrated its scalability benefits in terms of no overhead growth with the number of flows, and logarithmic overhead growth with network size. The simulation results shown in Figure 6, 7, 8.
Page 11
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IX. Performance Evaluation We compare our protocol to reactive protocols (AODV, DSR) and a proactive protocol (DSDV). Our results shown in table.1 suggest that dynamic addressing and proactive routing together provide significant scalability advantages and high level addressing. DART reduces the overhead growth with the number of flows, and logarithmic overhead growth with network size shown in figure 1. No of Transmitted packets No of Lost packets Figure.6 Communication between source and destination
DSDV DSR AODV DART 10726 5138 23349 42661 82
234
1003
1654
Table.1 DART performance
Figure.7 Destination became member in group 3 Figure.9 Throughput vs. Network Size (Nodes)
Figure.8 Destination leaving group3 and move towards group4
Velammal College of Engineering and Technology, Madurai
X. Conclusion We proposed Dynamic address routing, an initial design toward scalable ad hoc routing. We outline the novel challenges involved in a dynamic addressing scheme, and proceeded to describe efficient algorithmic solutions. We show how our dynamic addressing can support scalable routing. We demonstrate, through simulation and analysis, that our approach has promising scalability properties and is a viable alternative to current ad hoc routing protocols. First, we qualitatively compare proactive and reactive overhead and determine the regime in which proactive routing exhibits less overhead that its reactive counterpart. Large scale simulations show that the average routing table size with DART grows logarithmically with the size of the network. Second, using the ns-2 simulator, we compare our routing scheme to AODV,
Page 12
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
DSR and DSDV, and observe that our approach achieves superior throughput, and with considerably smaller overhead, in networks larger than 400 nodes. The trend in simulated overhead, together with the analysis provided, strongly indicate that DART is the only feasible routing protocol for large networks. We believe that dynamic addressing can be the basis for ad hoc routing protocols that for massive adhoc and mesh networks. REFERENCES [1] Ram Ramanathan and Martha Steenstrup, “Hierarchically-organized, multihop mobile wireless networks for quality-of-service support,” Mobile Networks and Applications, vol. 3, no. 1, pp. 101–119, 1998. [2] Guangyu Pei, Mario Gerla, Xiaoyan Hong, and Ching-Chuan Chiang, “A wireless hierarchical routing protocol with group mobility,” in WCNC, 1999. [3] G. Pei, M. Gerla, and X. Hong, “Lanmar: Landmark routing for large scale wireless ad hoc networks with group mobility,” in ACM MobiHOC’00, 2000. [4] X. Hong, M. Gerla, G. Pei, and C. Chiang, “A group mobility model for ad hoc wireless networks,” 1999. [5] J. Eriksson, M. Faloutsos, and S. Krishnamurthy, “Peernet: Pushing peer-2-peer down the stack.,” in IPTPS, 2003. [6] C. Perkins, “Ad hoc on demand distance vector routing,” 1997. [7] Charles Perkins and Pravin Bhagwat, “Highly dynamic destinationsequenced distance-vector routing (DSDV) for mobile computers,” in ACM SIGCOMM’94, 1994. [8] David B Johnson and David A Maltz, “Dynamic source routing in ad hoc wireless networks,” in Mobile Computing, vol. 353. Kluwer Academic Publishers, 1996. [9] Xiaoyan Hong, Kaixin Xu, and Mario Gerla, “Scalable routing protocols for mobile ad hoc networks,” IEEE NETWORK, vol. 16, no. 4, 2002. [10] Z. Haas, “A new routing protocol for the reconfigurable wireless networks,” 1997. [11] Guangyu Pei, Mario Gerla, and Tsu-Wei Chen, “Fisheye state routing: A routing scheme for ad hoc wireless networks,” in ICC (1), 2000, pp. 70– 74. [12] Paul F. Tsuchiya, “The landmark hierarchy : A new hierarchy for routing in very large networks,” in SIGCOMM. 1988, ACM. [13] Benjie Chen and Robert Morris, “L+: Scalable landmark routing and address lookup for multi-hop wireless networks,” 2002. [14] Aline C. Viana, Marcelo D. de Amorim, Serge Fdida, and Jos F. de Rezende, “Indirect routing using distributed location information,” ACM Mobile Networks Applications, Special Issue on Mobile and Pervasive Computing, 2003. [15] Jakob Eriksson, Michalis Faloutsos, and Srikanth Krishnamurthy, “Scalable ad hoc routing: The case for dynamic addressing,” in IEEE InfoCom, 2004.
Velammal College of Engineering and Technology, Madurai
Page 13
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Framework for Comparison of Association Rule Mining using Genetic Algorithm K.Indira, Research Scholar, Department of CSE, Pondicherry Engineering College, Pondicherry, India
S.Kanmani Professor & Head Department of IT Pondicherry Engineering College, Pondicherry, India
[email protected]
[email protected]
Abstract – A new framework for comparing the literature on Genetic Algorithm for Association Rule Mining is proposed in this paper. Genetic Algorithms have emerged as practical, robust optimization and search methods to generate accurate and reliable Association Rules. The main motivation for using GAs in the discovery of high-level prediction rules is that they perform a global search and cope better with attribute interaction than the greedy rule induction algorithms often used in data mining. The objective of the paper is to compare the performance of different methods based on the methodology, datasets used and results achieved. It is shown that the modification introduced in GAs increases the prediction accuracy and also reduces the error rate in mining effective association rules. The time required for mining is also reduced. Keywords - Data Mining, Genetic Algorithm, Association Rule Mining,
I.
INTRODUCTION
• Data selection: at this step, the data relevant to the analysis is decided on and retrieved from the data collection. • Data transformation: also known as data consolidation, it is a phase in which the selected data is transformed into forms appropriate for the mining procedure. • Data mining: it is the crucial step in which clever techniques are applied to extract patterns potentially useful. • Pattern evaluation: in this step, strictly interesting patterns representing knowledge are identified based on given measures. • Knowledge representation: is the final phase in which the discovered knowledge is visually represented to the user. This essential step uses visualization techniques to help users understand and interpret the data mining results. The architecture of Data mining system is depicted in figure.1
In today’s jargon enormous amount of data are stored in files, databases, and other repositories. Hence it becomes necessary, to develop powerful means for analysis and interpretation of such data and for the extraction of interesting knowledge to help in decision-making. Thus, there is a clear need for (semi-) automatic methods for extracting knowledge from data. This need has led to the emergence of a field called data mining and knowledge discovery. Data Mining, also popularly known as Knowledge Discovery in Databases (KDD), refers to the nontrivial extraction of implicit, previously unknown and potentially useful information from data in databases. The Knowledge Discovery in Databases process comprises of a few steps starting from raw data collections to formation of new knowledge. The iterative process consists of the following steps: • Data cleaning: also known as data cleansing, is a phase in which noise data and irrelevant data are removed from the collection. • Data integration: at this stage, multiple data sources, often heterogeneous, may be combined in a common source.
Velammal College of Engineering and Technology, Madurai
Figure 1. Architecture of typical Data mining System.
Page 14
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
This paper reviews the works published in the literature, where basic Genetic Algorithm is modified in some form to address Association Rule Mining. The rest of the paper is organized as follows. Section II briefly explains Association Analysis. Section III gives a preliminary overview of Genetic Algorithm for Rule Mining. Section IV Reviews the different approaches reported in the literature based on Genetic Algorithm for Mining Association Rules. Section V lists the inferences attained from the comparison. Section VI presents the concluding remarks and suggestions for further research. I.
new population. The flowchart of the Basic Genetic Algorithm is given in figure 2.
ASSOCIATION ANALYSIS
Association analysis is the discovery of what are commonly called association rules. It studies the frequency of items occurring together in transactional databases, and based on a threshold called support, identifies the frequent item sets. Another threshold, confidence, which is the conditional probability that an item appears in a transaction when another item appears, is used to pinpoint association rules. The discovered association rules are of the form: P Q [s,c], where P and Q are conjunctions of attribute value-pairs, and s (for support) is the probability that P and Q appear together in a transaction and c (for confidence) is the conditional probability that Q appears in a transaction when P is present. III. GENETIC ALGORITHM A Genetic Algorithm (GA) is a procedure used to find approximate solutions to search problems through the application of the principles of evolutionary biology. Genetic algorithms use biologically inspired techniques such as genetic inheritance, natural selection, mutation, and sexual reproduction (recombination, or crossover). Genetic algorithms are typically implemented using computer simulations in which an optimization problem is specified. For this problem, members of a space of candidate solutions, called individuals, are represented using abstract representations called chromosomes. The GA consists of an iterative process that evolves a working set of individuals called a population toward an objective function, or fitness function. Traditionally, solutions are represented using fixed length strings, especially binary strings, but alternative encodings have been developed. The evolutionary process of a GA is a highly simplified and stylized simulation of the biological version. It starts from a population of individuals randomly generated according to some probability distribution, usually uniform and updates this population in steps called generations. Each generation, multiple individuals are randomly selected from the current population based upon some application of fitness, bred using crossover, and modified through mutation to form a
Velammal College of Engineering and Technology, Madurai
Figure 2. Basic Genetic Algorithm.
A. [Start] Generate random population of n chromosomes (suitable solutions for the problem) B. [Fitness] Evaluate the fitness f(x) of each chromosome x in the population C. [New population] Create a new population by repeating the following steps until the new population is complete i. [Selection] Select two parent chromosomes from a population according to their fitness (the better fitness, the bigger chance to be selected) ii. [Crossover] With a crossover probability cross over the parents to form a new offspring (children). If no crossover was performed, offspring is an exact copy of parents. iii. [Mutation] With a mutation probability mutate new offspring at each locus (position in chromosome). iv. [Accepting] Place new offspring in a new population D. [Replace] Use new generated population for a further run of algorithm E. [Test] If the end condition is satisfied, stop, and return the best solution in current population F. [Loop] Go to step 2
Page 15
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IV. ANALYSIS ON GENETIC ALGORITHM FOR MINING ASSOCIATION RULES Among the Genetic algorithms designed for the purpose of Association rule mining is discussed based on the following criteria 1. Genetic Operations 2. Encoding I. Initial Population II. Crossover III. Mutation IV. Fitness Threshold 3. Methodology. 4. Application areas. 5. Evaluation Parameters The various methodologies are listed in Table A1. given in Annexure. 1. Genetic Operations The basic steps in the traditional Genetic algorithm implementations are discussed in the previous section. Modifications are carried out in the traditional GA to increase the prediction accuracy thereby reducing error rate in mining association rules. The variations have been carried out in various steps of GA. 2. Encoding : Encoding is the process of representing the entities in datasets for mining. Rules or chromosomes can be represented either with fixed length data [2..18] or by varying length chromosomes[1], Fuzzy rules are implemented for encoding data in [10], In [4] and [6] coding is done using natural numbers, In [17] Gene expressions are used for representation of chromosomes and encoding is carried out using arrays in [7] . Initial Population: The initial population could be generated by random selection, seeded by users [1], single rule set generation [5] and Fuzzy Rules [10]. Crossover: The Crossover operator which produces new offspring and hence new population plays a vital role in enhancing the efficiency of the algorithms. The changes are carried out as discussed in Table A1. Saggar. M et. Al [2] describes whether crossover is to be performed or not and if required the locus point where the crossover begins is of prime importance. Crossover on same attributes of both offspring if present or random attributes in absence of similar attributes is carried out in [11]. In [12] setting the crossover rate dynamically for each generations are presented. The concept of Symbiotic combination, where instead of crossover the combination of chromosomes to generate a new chromosome based on Symbiogenesis in Ramin Halavathy et. Al [5] has proved to increase the speed of the rule generation system.
Velammal College of Engineering and Technology, Madurai
Mutation: Mutation is the process where attributes of selected entities are changed to from a new offspring. The Mutation Operator is altered based on the application domain into macro mutation in [1], changing locus points of mutation in [2]. The weight factor is taken into consideration for locus point of mutation in [5] so as to generate a better offspring, Dynamic mutation where the mutation point is decided on the particular entity and generation selected enhances the diversity of colony is introduced in [12], mutation 1 & mutation 2 where mutation is performed twice to generate offspring is performed in [16]. Adaptive mutation where the mutation rate differs for each generation is found to produce better offspring’s in [17]. Fitness Threshold: The passing of chromosomes from a population to new population for the succeeding generation depends on the fitness criteria. Changes to the fitness functions or threshold values alter the population and hence the effective fitness values lead to the efficiency of the rules generated. The negation of the attributes are taken into consideration while generating rules by including criteria like True Positive, True Negative, False Positive and False Negative[2]. By these criteria rules with negated conditions of attributes are also generated. By varying the fitness values dynamically in each generation, the speed of the system can be improved [4]. Factors like strength of implication of rules when considered while calculating fitness threshold proves to generate more interesting rules [6]. The Sustainability index, creditable index and inclusive index when considered for fitness threshold results in better predictive accuracy [7]. When the real values of confidence and support derived and applied in threshold rate found to generate faster than traditional methods [8]. The predictability and comprehensibility factors of rules tends to provide better classification performance[11]. 3. Methodology Rather than altering the operations in basic GA algorithm the changes made in the methodology has also proved to increase the performance. In [5] the crossover operation is replaced by symbiotic recombination techniques. Wenxiang Dou et. Al [8] describes the generation of rules and displayed to the user. The user decides the interesting rules thereby seeding the population by the user. Here instead of searching the whole database for support the system searches K- itemset alone and hence is faster than other methodologies. The real values of support and confidence are taken into account while generating threshold values. If user is not satisfied then more rules with next level of support could be generated. The response time of the system is found to be increased dramatically. Antonio Peregrin et. Al [9] states the distributed genetic algorithm where the Elit data pool is the central node. Each node connected are considered as Intercommunicating
Page 16
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Subpopulations. The Data Learning Flow (DLF) Copies the Training Example with Low Fitness to neighborhood and each node is assigned with different partitions of learning data. In [18] the concept of Dynamic Immune Evolution and Biometric Mechanism is introduced. The immune recognition, immune memory and immune regulation is applied to GA and thereby making the system faster and discovering interesting rules though support might be low.
• Predictive Accuracy of more than 95 percentage was seen in 55 % of datasets. • Predictive Accuracy between 80 - 94 percentage was seen in 15 % of datasets. • Predictive Accuracy between 75 – 79 percentage was seen in 10 % of datasets. • Predictive Accuracy of less than 75 percentage was seen in 20 % of datasets.
4. Application Areas
Predictive accuracy of 100 % is achieved with methodology [1] and [14]. The error rate is seen maximum in Pima dataset under UCI repository by SGERD method [10].
From the dataset in Table A1 it is derived that the Genetic Algorithm is flexible in nature and could be applied to varying data sets. Different areas of applications as Biology, biometrics, Education, Manufacturing Information System, Application Protocol interface records from Computers for Intrusion Detection, Software Engineering, Virus information from Computer data, Image data base, Finance information, Students Information etc. are taken up for study. It is seen that by altering representations and operators the Genetic algorithm could be applied for any fields without compromising the efficiency. 5. Evaluation Parameters To achieve the best performance from any system setting of the evaluation parameters are vital. For mining efficient association rules using the evolutionary Genetic Algorithm the parameters that affects the functioning of the system are representation of chromosomes and their length, population Size, Mutation probability, Crossover probability and the Fitness threshold values. The use of support and confidence factor in deciding the threshold values also increases the performance of the system. It is seen form the Table A1 that parameters don’t have predefined boundaries. The values of the parameters are based upon the methodology applied and the dataset. It is found that the mutation rate differs from 0.005 to 0.6. Similarly the Crossover probability ranges from 0.1 to 0.9. The fitness function is found to be the most crucial parameter. The optimum method for the fitness threshold evaluation has to be arrived in order to achieve the highest performance. The carrying over of offspring from one generation to other is based on fitness function the population. Different factors of the system are takem into consideration while generating fitness function. Efficient the fitness function, efficient and accurate are the association rule that are generated by the system V.
INFERENCES
The study of the different system using Genetic Algorithm for Association rule mining was carried out. The predictive accuracy based on the methods under comparison is noted that
Velammal College of Engineering and Technology, Madurai
Based on the time for mining association rules as low as 5 and 9.81 seconds were achieved in SGERD [10] and Quick response data mining model [8] respectively and maximum of 7012 seconds is taken up for mining KDDCUP99 dataset using Genetic algorithm with symbiogenesis [5]. The tools used for mining are GA Toolbox from Mat lab, Java, C++ and Weka for training the datasets. From the Study it could be inferred that • The system is flexible and could be applied to varying datasets and applications. • The performance of the system is improved when compared to other existing methods like Apriori, C4.5, REGAL, Data-miner etc. • The Genetic operators are the key in achieving increased performance. • The speed of the system increases rapidly when compared with other methods. • The optimum values for the mutation, crossover and threshold decide on the performance of the system. • Better and precise rules are generated. • Classification error rates are reduced in comparison.. • Negated attributes could be derived in rules. • Critical rules could be generated even if support is low. VI. CONCLUSION The framework for comparing the effectiveness of Genetic Algorithm in mining Association rules is taken up for study. Various systems applying GA for mining Association rules are considered and the tabulation was done with the data chosen. From the analysis it could be derived that the classification performance of the system was found to be robust. Genetic Algorithm based implementation Outperforms Conventional Methods. The speedup achieved is remarkable. For future work the combination of one or two methodologies could be carried out for basic GA operators. The given system for a particular domain can be further modified for other domains.
Page 17
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Table A1 : Genetic Algorithms For Association Rule Mining: A Comparative Study Data Set Ref .No
Methodology Steps In Genetic Algorithm Represen -tation
[1]
Varying Length
Selectio n
Seeded Populati on
Mutation
Macro Mutations
Cross Over
As GA
As GA
[2]
[3]
[4]
[5]
[6]
Binary Coding
As GA
Using Natural Numbers
As GA
Using Natural Numbers
Roulette Wheel Selection
As GA
Roulette Wheel Selection
As GA
Optimum Value
Whether Required Or Not And If Point Of Crossove r
As GA
Optimum Value
Changes In Weight Symbiotic Generate Of Combinati Single Membersh on Rule Sets ip Operator Function
As GA
As GA
As GA
TP,TN,FP, FN Used
As GA
As GA
Grilled Mushrooms in Agaricus and Lepiota Family.
Synthetic Database for the Selection of Electives for a Course
Data Set from MIS Synthetic Image Database KDDCUP99 Classes 2 CRX Classes 2
As GA
Individual Evaluation Using Strength Of Implication
[7]
Array Represent a -Tion
As GA
As GA
As GA
Based On Sustaining, Creditable And Inclusive Index
[8]
As GA
Done By
As GA
As GA
Based On Real
Evaluated By / Parameters Size / Pc/ Suppor Confi t dence
Fitness
Vehicle Silhouette Dataset Whether Required Or Not And If Point Of Mutation
Sample Size
Fitness
Gen er atio ns
8128 Of 23 Species 22 Attributes 846 Records 18 Attributes
*
100
*
*
0.1
0.005
5%
User define d
*
5% *
*
.25% To 2%
*
4 To 14
*
Completen ess consider -red
User defined
*
*
-
*
*
Features 15 Size 690 Features 4 Size 150
Vote Classes 2
Features 16 Size 435
Wine Classes 3
Features 13 Size 178
10000 0.4
10
0.85
0.4
0.2
0.8
0.6
6
0.6
8
50
*
Six Datasets from Irvine Repository
*
S.I 1.0 C I 1.0 I.I 1.0
10
* Single Table
Velammal College of Engineering and Technology, Madurai
Accuracy On Training Dataset Between 95 To 100. On Test Data Between 62 And 71 Rules With Negation Of Attributes As Well As General Rules Generated Based On Time And Fetching Knowledge GA Is Faster Than Apriori. Runs 2 To 5 Times Faster Than Apriori
Features 41 Size 494021
IRIS Classes 3
Car Test Results Dataset
Pm
Results
40
10%
50%
50
SEA Has Better Or Similar Results When Compared With GA SEA Much Faster Than GA
During Generation Between 1 To 200 Interesting Rules For Whatever Be The Threshold. Predictive Accuracy Better Than CN2 And AntMiner Methods Response Got
Page 18
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Users From Rules Generate d •
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Support And Confidence
Produced Randomly with 100 Transactions
Multiple Intercommunicating Subpopulations Distributed Data And DLF • Central Elite Pool UCI : Nursery • The Data Learning Flow (DLF) Copies The Training Example With Low Fitness To Neighborhood • Each Node Is Assigned With Different Partitions Of Learning Data Fuzzy Rules
As GA
As GA
As GA
As GA
Gene String Struct ure •
0 *
*
0.01
0.9
•
As GA
Elitist As GA Recombinat ion Method
As GA
Attributes
As GA
As GA
As GA
As GA
As GA
As GA
As GA
As GA
Dynamic
Adaptive
As GA
As GA
Mutation1 & Mutation 2
Adaptive
As GA
Done On Same Attributes If Present Or Random
Dynamic
As GA
As GA
As GA
As GA
Predictive Accuracy
11 Data Sets from Irvine Machine Learning Repository Adult
Nursery Datasets from UCI
Based On Last Generation
Finance Service Data of Certain City
Individual Based
Database Of Student Achievement in Schools in Recent Years
Modified To Decide KDD CUP99 Whether A Dataset Chromosome Is Right Or Not
Feature Selection Is Applied
CM1, KC1, KC2, PC1 From UCI Repository Vehicle Dataset And
As GA
Adaptive
Based On Distance Between Rules Lympography Dataset From UCI ML Measure Of Overall Performance
Dynamic Immune Evolution And Biometric
Real Case Data
Computers
Velammal College of Engineering and Technology, Madurai
12960 Instances
*
*
optimu m 48842 Instances 15attributes
40
Optimu m
0.3
200
Optimu m
0.6
0.9
*
*
*
0.01
0.7
*
*
148 Records 18 Attributes 4 Classes
Varie s
240
Length Of Chromosome 41 • Generations 100 • Generation Gap 0.9
Varie s
Varie s
0.1
0.01 0.01
846 Records, 18 Attributes 4 Classes
*
*
50
*
•
22
*
0.05 *
12960 With 9 Attributes
2050 Groups
optimu m
50 0
600
0.6
0.8
0.005
*
Varies *
*
*
In Ten Seconds Whereas For Apriori It Is More Than 3000 Seconds Faster And Better Behavior Number Of Rules Generated Is Between 60%80% Smaller Classification Error Rates Are Low Outperforms C4.5 Better Classification Performance Produces Partial Association Rules After 252 Generations Whereas It Is 850 In traditional GA The Algorithm Based On 0.1 Support And 0.7 Confidence Is Close To Actual Situation Rules Generated Are Useful In Detecting Intrusion Generated Rules That Provide Better Estimation And Explanation Of Defective Modules
10 0
GRA Outperforms Conventional Methods
50
Performance And Effectiveness Of Proposed Model Is Close With Real World Analysis Faster &
Page 19
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
•
Mechanism Is Introduced Immune Recognition, Immune Memory And Immune Regulation Is Applied To GA
Daily Records Of API
0.26 *
*
*
*
Discovers New Critical Rules Though Support Not High
Note : *- not defined in literature REFERENCES [1]. Cattral, R., Oppacher, F., Deugo, D.,”Rule Acquisition with a Genetic Algorithm”, Proceedings of the 1999 Congress on Evolutionary Computation,. CEC 99, 1999. [2]. Saggar, M., Agrawal, A.K., Lad, A., “Optimization of Association Rule Mining”, IEEE International Conference on Systems, Man and Cybernetics, Vol. 4, Page(s): 3725 – 3729, 2004 [3]. Cunrong li, Mingzhong Yang, “Association Rules Data mining in Manufacturing”, 3rd International Conference on Computational Electromagnetics and Its Applications, Page(s): 153 – 156, 2004. [4]. Shangping Dai, Li Gao, Qiang Zhu, Changwu Zhu, “A Novel Genetic Algorithm Based on Image Databases for Mining Association Rules”, 6th IEEE/ACIS International Conference on Computer and Information Science, Page(s): 977 – 980, 2007 [5]. Halavati, R., Shouraki, S.B., Esfandiar, P., Lotfi, S., “Rule Based Classifier Generation Using Symbiotic Evolutionary Algorithm” , 19th IEEE International Conference on Tools with Artificial Intelligence, Volume: 1, Page(s): 458 – 464, 2007. [6]. Zhou Jun, Li Shu-you, Mei Hong-yan, Liu Haixia, “A Method for Finding Implicating Rules Based on the Genetic Algorithm”, Third International Conference on Natural Computation, Volume: 3, Page(s): 400 – 405, 2007. [7]. Hua Tang, Jun Lu, “A Hybrid Algorithm Combined Genetic Algorithm with Information Entropy for Data Mining”, 2nd IEEE Conference on Industrial Electronics and Applications, Page(s): 753 – 757, 2007. [8]. Wenxiang Dou, Jinglu Hu, Hirasawa, K., Gengfeng Wu, “Quick Response Data Mining Model using Genetic Algorithm”, SICE Annual Conference, Page(s): 1214 – 1219, 2008 [9]. Peregrin, A., Rodriguez, M.A., “Efficient Distributed Genetic Algorithm for Rule Extraction”,. Eighth International Conference on Hybrid Intelligent Systems, HIS '08. Page(s): 531 – 536, 2008 [10]. Mansoori, E.G., Zolghadri, M.J., Katebi, S.D., “SGERD: A Steady-State Genetic Algorithm for Extracting Fuzzy Classification Rules From Data”, IEEE Transactions on Fuzzy Systems, Volume: 16 , Issue: 4 , Page(s): 1061 – 1071, 2008. [11]. Xian-Jun Shi, Hong Lei, “A Genetic AlgorithmBased Approach for Classification Rule Discovery”,
Velammal College of Engineering and Technology, Madurai
International Conference on Information Management, Innovation Management and Industrial Engineering, ICIII '08, Volume: 1 , Page(s): 175 – 178, 2008. [12]. Xiaoyuan Zhu, Yongquan Yu, Xueyan Guo, “Genetic Algorithm Based on Evolution Strategy and the Application in Data Mining”, First International Workshop on Education Technology and Computer Science, ETCS '09, Volume: 1 , Page(s): 848 – 852, 2009 [13]. Hong Guo, Ya Zhou, “An Algorithm for Mining Association Rules Based on Improved Genetic Algorithm and its Application”, 3rd International Conference on Genetic and Evolutionary Computing, WGEC '09, Page(s): 117 – 120, 2009 [14]. Yong Wang, Dawu Gu, Xiuxia Tian, Jing Li, “Genetic Algorithm Rule Definition for Denial of Services Network Intrusion Detection”, International Conference on Computational Intelligence and Natural Computing, CINC '09, Volume: 1 , Page(s): 99 – 102, 2009 [15]. Rodriguez, D., Riquelme, J.C., Ruiz, R., AguilarRuiz, J.S., “Searching for Rules to find Defective Modules in Unbalanced Data Sets”, 1st International Symposium on Search Based Software Engineering, Page(s): 89 – 92, 2009 [16]. Gonzales, E., Mabu, S., Taboada, K., Shimada, K., Hirasawa, K., “Mining Multi-class Datasets using Genetic Relation Algorithm for Rule Reduction”, IEEE Congress on Evolutionary Computation, CEC '09, Page(s): 3249 – 3255, 2009 [17]. Haiying Ma, Xin Li, “Application of Data Mining in Preventing Credit Card Fraud”, International Conference on Management and Service Science, MASS '09, Page(s): 1 – 6, 2009 [18]. Genxiang Zhang, Haishan Chen, “Immune Optimization Based Genetic Algorithm for Incremental Association Rules Mining”, International Conference on Artificial Intelligence and Computational Intelligence, AICI '09, Volume: 4, Page(s): 341 – 345, 2009
Page 20
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Content Management through Electronic Document Management System T.Vengattaraman#1, A.Ramalingam*2, P.Dhavachelvan#3 #
Department of Computer Science, Pondicherry University
Puducherry, India. 1
[email protected] 3 [email protected]
*
Department of Master in Computer Applications, Sri Manakula Vinayagar Engineering College
Puducherry, India. 2
[email protected]
Abstract— This paper presents “Electronic Document Management System” which is to develop a web based application which can be used to manage the contents hosted on a web site. Electronic document management system provides services in the areas of providing information to the different areas such as Development, Business, Management, Technology etc. The main functionality of this system is Content Creation/Hosting, which is the administration section of the system in which the content is created by a group of professional possessing different specialization. This section is divided into four stages such as authoring, editing, approving and deploying. After going through these four phases of content management the content is available to the users. Each content is placed in one of the three different categories. And content View is the one where the content to be viewed is divided into three halves as developer Content, manager Content and protected Content. The Protected content is login protected where as the other two are open to all. The developer content is used by the all for content/information modification that helps in the dynamic display of different information for providing services to the Business, Management, etc whereas the manager content is also used by all but the contents are used for customization purpose. Keywords— Electronic document management system, Content management, Web publishing.
I. INTRODUCTION To develop a web based interface for creating new content for the site or manage existing content. The system should provide help in managing different personnel capable of working in different areas of content creation. To make the best possible content available as an end result. Divide the complex task of content creating into no of specialists. To allow the floating of content with the system before being hosted on site. Different levels of approval to make the content very precise as per the requirements. To speed up the processing with in the
Velammal College of Engineering and Technology, Madurai
Content Creation. The current system although semi atomized required manual processing of approval and editing before being approved for the deployment. Transfer of information content between different sections of the current system is in the form of documents. Selection of a person for a task is done by manually approaching the person and confirming the availability of the person. Due to mismanagement the work is delayed to later date than the due date. Although the current system is confined to only one type of a website management. It can be generalized a General Electronic Document Management system with which any type of website can be managed. The current system is interactive with the database provides efforts can be made so that the system can adopt the available database features of a new site to make is as a part of content management. II. LITERATURE SURVEY As organizations increasingly move towards a paperless environment due to an increasing need in storing more and more documents, an Electronic Document Management System (EDMS) must be in place to cope not only with the increasing volume of these documents but also to make sure that these documents are searchable for use in the future. EDMS is proven to be an appropriate tool for handling necessary documents of organizations around the world. To aid the researchers pursue this study, they have gathered some studies related to their proposed system. Electronic Document Management System (EDMS) seemed to be the answer in organizing business information for easy access and quick retrieval and archiving important electronic records in a secure and stable environment. EDMS might just provide the tool to effectively manage electronic documents and records throughout its life cycle. After implementing the EDMS with Accounts Payable application and Electronic Reports Management technology to AEC, access and retrieval of documents is more efficient and there have been no reported “lost” or “misfiled” documents. Employees in
Page 21
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
remote locations can quickly retrieve and view records not requiring much time and effort, and more paper. Since EDMS provides a secure and stable environment, records cannot be altered, and any document access or action can be tracked through audit history. Vital records stored in EDMS are secured and backed up daily[1-3]. To implement an automated document analyzer that will provide the file organization, searching, retrieval and segmentation of files is the main objective of the thesis by Revano [7]. To achieve his goal, he used three data mining algorithms such as the Aho-Corasick algorithm for exact set matching, the Segmentation algorithm which generates a series of cluster based on the identified relationships in a dataset, and the Text mining algorithm or automatic text categorization. Based from findings of the study, the researchers had proven that there was a significant difference between the proposed system and the existing system being implemented in CITE department. Based from the gathered related studies, the researchers have come up with the idea of developing an electronic document management system. Organizing documents, paper and electronic, is now a concern in any organization. In relation with the studies mentioned, the researchers agreed that exhibiting the basic components of electronic document management systems such as capturing, storage and archiving, indexing and retrieval, distribution, and security will produce a good EDMS. III. SYSTEM OVERVIEW The complete system can be divided into six halves on basis of access levels. The system architecture is given in the figure 1. A. Account management: Using this part of an application the administrator can view the list of users and their area of specialization. The administrator can create a new users, modify existing user. An administrator provides permission to the newly created user by placing the new user into set of roles such as an author, approver, and editor or deploy. This part of the application is only accessible to the administrator. Login Area
Processing Area Database Approving
User Name
Account Login
Deploy
Database
Editing
Password
Authorizing
B. Utilities: Utilities section of the application is used to shut down the site for the normal person to browse as well as to up the site back for its use. C. Authoring: An administrator or a person with the author privileges can access this part of the application. This part of the application includes creating new content in the form of stories which is normally done by the developers or content writers. The newly created content may include no of notes which will guide the editor at the time of editing the content. The newly created content then can be posted to editor for editing. D. Editor: An editor receives the content posted by the author. An editor can view the content and later post the content to a new revision or to an existing revision. If content is found unsuitable to the cause the content is returned back to the author. This part of the application can be explored only by an administrator or the users who possess an editor privilege. The editor can withdraw the content from being hosted if found unfit for hosting. E. Approver: An approver is a person who will approve the contents to be hosted on the site. An approver can approve the content to the deploy section or Discontinue the content usage or return the content back to the editor for revision. The returned content should accompany with a message to the editor regarding the revision in the content. This part of the application can be accessed by the administrator of the person who possesses an Approver privilege. F. Deploy: This area of the application includes the deployment part of an application. A deploy person can view the content before deploying it. The person can also return the content if found unfit to be hosted on the site. The returned content is sent back to the approver. The deployment of the content includes the content to be placed in specific area of the hosting environment. The hosting environment is divided into three categories. The Deploy content, the manager content, the protected content. These categories are subdivided into no of sections. G. Administrator: An administrator has all the privileges that of the guest as well as the normal registered user. Along with these common features an administrator has the administrator related features such as creating new users and granting roles to those newly created users. The roles granted by the administrator cannot be changes by the user. An administrator can create new user as a guest or as a user
Fig 1 General Architecture Diagram
Velammal College of Engineering and Technology, Madurai
Page 22
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
or an administrator. The access levels are as per the grants done by the administrator. An administrator can also be part of a team and could lead a project team this is possible only if administrator when building a team includes himself in the team section. If included as a manager he is not a part of the team but supervisor of the team. The register option on the homepage of the application is provided only to register a new user as a guest. The information of the entire system will be maintained at a centralized data base. Notifications between sections is provides in terms of content list notification in the users area. Provide Interactive interface through which a user can interact with different areas of application easily. Disapproved content is returned back to the lower for redesign. Approved content is removed from the user list and made as part of the user’s list to which the content is being notified. Deploy the application on a single system and make is available on all the systems within the network, thereby reducing the maintenance cost of software. The system screen shots are shown in the figure 2 and figure 3.
Fig. 2 The Contents are hosted in this area
IV. CONCLUSION This system has been appreciated by all the users and is easy to use, since it uses the GUI provided in the user dialog. User friendly screens are provided. The usage of software increases the efficiency, decreases the effort. It has been efficiently employed as a Content management mechanism. It has been thoroughly tested and implemented. The application is capable of managing only the current site. It can be converted into general Electronic Document Management system by providing a tight integration of the existing database with the one the site holds. The current database should be redesigned so that it can adapt to the changes with respect to the site. The current Electronic Document Management system manages only the articles and there exists no provision for the new categories to be added and placed into the content section. This can be brought into the current on user request. The application can be converted into a Mobile based using ASP.net with which the deployment of application will be done only on enterprises server and is accessible to all other departments of the organization. The current application is confined to only one enterprise. REFERENCES [1] Adam, “Fundamentals of EDRMS. Implementing Electronic Document and Record Management Systems”, CRC Press. Madison Ave. New York, 2007 [2] J. Feldman and E. Freuder, “Integrating Business Rules and Constraint Programming Technologies for EDM”, Business Rules Forum 2008. [3] Georgia Archives, “Electronic Document Management System Technologies” ,November 2008. [4] P. Immanuel, “Basic Components of Electronic Document Management System – Essential Characteristics of an Effective Document Management System”, November 2008 [5] Laserfiche, “A guide to the benefits, technology, and implementation of electronic document management solutions”, Archival and scanning services, Nov 2008. [6] D.P. Quiambao, “Document Management System at the Property Management Division of Subic Bay Metropolitan Authority”, 2004 [7] T.J. Revano, “Development of Automated Document Analyzer using Text Mining and Aho-Corasick Algorithm” . TIP Manila: Unpublished Thesis, 2008 [8] Sire Technologies, “A Guide to Understanding Electronic Document Management Systems: The Future File Cabinets”, 2006.
Fig. 3 To deploy the content.
Velammal College of Engineering and Technology, Madurai
Page 23
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Intelligent Agent based Data Cleaning to improve the Accuracy of WiFi Positioning System Using Geographical Information System (GIS) T.JOSHVA DEVADAS Lecturer, Department of MCA, The American College, Madurai [email protected] Abstract — Wifi positioning system uses GIS to achieve higher accuracy by means of comparing the error distance. The objective of this paper is to minimize the distance error generated during the process of positioning. Wifi positioning system needs to have a proper placement of those access points which are free from error distance. Data cleaning is introduced in this concept to clean the errors that are generated from the experimental results. An agent is a Intelligent Processing system that is situated in the environment and perform autonomous action in order to meet the design objectives. Agents are introduced in this system uses its intelligence during the data cleaning task by using its dynamic and flexible characteristics. This paper aims at describing the functionalities of such a data cleaning agent so as to improve the performance of Wifi positioning system. Keywords— Wireless positioning, WiFi, GIS, Constraint optimization, Data cleaning, Agents
1.
Introduction
WiFi stands for wireless fidelity which is one of the Institute of Electrical and Electronics Engineers or IEEE standards for WLAN. Wifi is truly a freedom to make a connection to Internet without the old-fashioned network cable. WiFi or 802.11b is one of 802.11 specifications family and it is also the first dominating standard in the market. Nowadays new wireless networking devices are adopting the new standard of IEEE 802.11g, which offers higher data speed, namely driving from 11 Mbps to 54 Mbps, while sharing common operating frequency of 2.4 GHz.
WiFi technology is evolving technically and practically in past four years leading WLAN to be a common sight at universities, airports, coffee shops, offices and organizations. That service point or access point is often referred to as “wireless hotspot” or “hotspot” in short. Incentives for developing and standardizing WLAN are definitely mobility and flexibility. This phenomenon stimulates the creation of supportive features for new-age laptops. Data preprocessing is a procedural activity which is to be done at the first step of data mining process. Data preprocessing is a process of requesting the data to be
Velammal College of Engineering and Technology, Madurai
cleaned. Data mining process need to analyze the possible occurrence of the noisy data[14]. The noise will occur due to faults generated during the data transmission. To resolve this we need to develop a system which can produce an error free data by identifying or removing the outliers identified in the system. In this paper we consider only those error positions which are to be eliminated during the wifi positioning process. Agents are intelligent autonomous computer systems, which are capable of functioning autonomously and react to the environment[1]. The intelligent agent doesn’t just react to the changes to its environment but itself takes the initiative under specific circumstance. This characteristics request the agent should have a well defined goal. Agents incorporated in the data cleaning task have a specific goal of removing the errors by identifying the outliers. In this paper section 2 deals about related work and background details of wifi positioning system. Section 3 describes the positioning computation by describing radio propagation model, free-space model and wall-attenuation method. Section 4 explains data cleaning process along with agent participation, section 5 describes performance evaluation.
2. Related works and Background Now days all the WLAN is the emerging research area that varies from real-world application to pure theoretical aspect. Major issue in wireless security is the wireless positioning application. WEP and WPA are encryption algorithms that are to be realized by the realworld globally. Hence the development for wireless security turns to be the positioning technique. Many researchers use the positioning technique to improve the accuracy of their need with various platform and architecture. The objective of this paper is to improve the accuracy by the means of positioning the access points in two areas with the aid of GIS. The experiment considers the indoor environment with two access points.
Page 24
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
There are two types of computation to perform this positioning process. One is location fingerprinting model and the other one is the propagation based computation model[8]. Location fingerprinting requires site survey and to collect signal strength samples at every single equally divided area of that building to build up a signal strength database. Propagation based computation is used to get the strength of signal but it don’t have the facility to store the database. These are the some of the computation to do the mapping process. The following is the test site for the system where we need to position the access points
the corridor. Those fifteen markers are points of interest in our experiment to be used for learning and testing purposes. They are 1.25 meters apart from one another.
FIG. 2.PROCESS OF NETSTUMBLER Multipath Fading Minimization:
Fig .1. Q5 floor plan with markers and access points locations In the test site we need to place the access points in two different areas and the corridor is selected for our test process and get the signal from two access points namely NAL-Q51 and DSP-Q52. Triangulation is a basic geometric concept for finding crossing point on the basis of distance or angle [6][7]. Therefore it is radiating circularly in horizontal direction and so the access point is in the pattern of circle. The existing system needs to be improved in order to decrease the error distance between the two access points and therefore it needs to perform the operation to obtain the calculated signal strength and error distance[2]. The next step in this process is to place the access point in an error free area. The existing system makes use of the Netstumbler software to detect the signal strength[10]. Signals are received from the two access points with the help of PDA or laptop and that are detected using the Netstumbler software. Netstumbler imports the signal strengths and are mapped into excel sheet to calculate the signal average. MATLAB is used to find the erroneous data and later the positioning is done by using the floor position and calculated positions. Predetermined location of the access points named DSPQ52 and NAL-Q51 is included in Fig 1 with markers along
Velammal College of Engineering and Technology, Madurai
From all the four directions, the signal strength is collected to minimize the multipath fading. The Netstumbler software is used to detect the signal strength. Signal strength samples directly reflect the multipath effect in terms of direction sensitivity and angle of reception. In principle, we should achieve higher resolution of sampling and higher accuracy. Experiment Methodology: Equipments required for capturing signal strength, raw data processing and position determination are listed in categories of Hardware and Software accordingly. i) Hardware Access Point or Hotspot is functioning as a signal transmitter. Wireless Network Interface Card (Wireless NIC) or WiFi card is considered as wireless LAN signal receiver in this application. Wireless NIC is available in various interface types and it can operate on both laptop or desktop computer and Personal Digital Assistant (PDA) correspondingly. Latest laptop computer and PDA are equipped with internal Wireless NIC. Desktop computer is not suitable due to obvious reason of mobility from point to point around the floor. ii) Software NetStumbler is a freeware program and it is one of the most widely used wireless network auditing tools allowing the detection of 802.11b, 802.11a, and 802.11g networks. This software offers two versions for both Windows and WindowsCE platform, which are Netstumbler 0.4.0 and
Page 25
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Ministumbler 0.4.0 correspondingly. Signal strength samples are saved into NS1 file and it can be further be exported into Microsoft EXCEL for calculation, but Ministumbler does not have that feature. EXCEL is used to calculate the signal strength average for each particular point on the floor plan. MATLAB Characterizes wireless propagation model and minimize the error distance.
transmission model is described as PR = PTGTGR { λ / 4пd }2 where PR and PT are receiving and transmitting power in unit of watts. GT and GR are receiving and transmitting antenna gain. λ is a wavelength and d is a distance between transmitter and receiver.
3. Position Computation
Our experiment considers this model because it is considered to be more suitable for indoor environment. This model represents a common and reliable in building characteristics. Attenuation due to an intervening wall or partition is described as PR[dBm] = Po[dBm]- 10 log{d/d0}n –WAF where n is a path loss exponent and WAF is a Wall Attenuation Factor. Building environment may have the different electrical and physical characteristics which have obstruction that varies from the place to place. These values can be obtained by experiments.
Conceptual Overview Our experiment considers first floor plan found in Fig 1. To locate the access point’s position x-y coordinates is chosen. Distance error is calculated with respect to the exact location on the floor plan in hand. Signal strength should be collected from the NAL-Q51 and DSP-Q52 will be interpreted to any of the two models namely free-space model and wall attenuation model.
Wall Attenuation Factor Model:
Learning Process Position determination is based on free-space model and it does not require learning process as universal relationship is implied every environment considered. WAF model is used to achieve more accurate result and to represent more realistic indoor environment[9]. Before positioning service WAF and n have been computed from experiment. Once the measurement of signal strength at marking points is done, linear regression is applied to those data sets resulting in WAF and n parameters. Fig.3. Position determination overview Radio Propagation models Basically the radio channel associates with reflection, refraction, diffraction and scattering of radio waves that influence propagating paths. Transmitted signal from direct and indirect propagation path are combined either constructively or destructively causing variation of received signal strength at the receiving station. The situation is even more severe for indoors communication. The building may have different architectures, construction material which results in receiving challenging and unpredictable signal. Free-space Model This model is used for the worst case consideration. This model can be implemented with GIS but it unveils a trend for further improvement. This model is not suitable to implement because this may disturb the other signals. Also, this model is not appropriate for the indoor environment to do the positioning system. Free-space model or Friis
Velammal College of Engineering and Technology, Madurai
Fig. 4. Line of sight for NAL access point Line joining between the two access points to every single point of interest on the floor plan can at least differentiate amount of signal attenuation due to variety of obstructing objects. Longer the range from a transmitter, the lower signal strength from a receiver can perceive. Shorter range from transmitter can yields lower signal strength only in the case of presence of a large barrier between transmitter and receiver. Regional division is essential for the case of
Page 26
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
NAL-Q51 access point as received signal strength does not undergo common obstruction as the case of DSP-Q52 access point. Essentially, Geographical information is now used to separate region under consideration. System Simulation Signal strength collected is translated into transmitterreceiver distance with the aid of two radio propagation models. Transmitter-receiver distance is the separation distance of access point from user and the distance are visualized as a radius of antenna radiation circle with center at that access point of signal reception. User can be at any position on that antenna radiation circle. Signal strength received from DSP-Q52 is translated into radius of antenna radiation circle with the center at the access point itself. 4. Data cleaning process Data cleaning is a process of identifying and removing the erroneous data[4]. Since the data is received with poor quality, it is necessary to clean the data before using it for processing. Poor quality of data is obtained due to some noise. The noisy data may have erroneous data, missing values, duplicate/ redundant data, and useless information. The error occur in the collected data are more common. One way of addressing this problem is to painstakingly check through the data with the help of data cleaning methods. Data Cleaning Methods Data cleaning is done by using either decision trees, or filling the missing values, or removal of noisy data and eliminate the redundancies. Decisions trees induced from training data can often be simplified, without loss of accuracy, by discarding misclassified instances from the training set. Filling of the missing values is done by replacing with a global constant, such as attribute mean value, string constant, and numeric constants. A noise is nothing but a random error which can be removed either by binning method or clustering method or with the help of human inspection [9]. We consider the clustering method is more suitable for cleaning the data of a wifi positioning system, because the formation of clusters clearly identifies the outliers. Identification of the outlier will eliminate irrelevant information by forming various clusters. Need and strength of cleaning
Erroneous data is assumed to be the access points coordinates which causes intersection with the signal strength. Since the data collected may have many such erroneous data items and the errors are removed by a data cleaning process or method. Agents Agents are intelligent processing elements which can work on behalf of others[13]. An intelligent agent may be categorized as a human agent, or a hardware agent or a software agent. A software agent may in turn be categorized into information agent, cooperation agent and transaction agent. The agents are categorized into two major categories as internal and external properties. Internal properties, describes the actions within the agent, which includes the ability to learn, reactivity, autonomy and goal-orientedness[11]. The external properties include all those characteristics which are associated with other agents for communication I n t e r a c t i o n
I n t e r a c t i o n
Action Information fusion
Information processing
Fig. 5.Workflow of the agent Work process of intelligent agent The functionality of the agent can be described as a black box. Input is received through perceptions by the agent. After receiving input, agent use the information received for its intelligent processing and send the output as action performed over the given input[5]. Intelligent processing of the agent is done as a series of actions performed over the input. Interaction component found in this process is responsible of sending and receiving data from or to the external components. Information fusion is responsible for collecting the inputs from the interaction component. The information processing component is responsible of handling the input by carefully analyzing the data to perform action on it. The action component is capable of performing action on the data received and sends the output to the interaction component.
In wifi positioning system access points PDA or laptop detect the signal strength and export into excel sheet is consider as primary data for data cleaning. At this stage the data collected may have erroneous information.
Velammal College of Engineering and Technology, Madurai
Page 27
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Need and the Role of Agents in Data Cleaning Agents are introduced in the data cleaning process are capable of reacting with the environment with their intelligent behavior and is used to improve performance of wifi positioning system. The Intelligent behavior is achieved by systematically analyzing the data and store the useful information alone in the knowledgebase. Intelligent agents incorporated in the preprocessing phase of the data cleaning task uses its flexible and dynamic characteristics to clean the data. Agent uses its knowledge base in the cleaning task reacts with the system by verifying the existence of the required information. Use of agents in wifi positioning system will improve the performance by not repeating the same procedure for the access positions which has intersection in signal strength. Intersection of such signal strengths causes an error and the errors positions are stored in the knowledge base of an agent. When the agents receive the data, it checks its knowledgebase to verify the existence of erroneous data. If an erroneous position was found then the agent filters this information otherwise it sends the details to perform action. Outliers are identified by using the domain of the erroneous data. Also, Agent based cleaning process distinguishes the access points which do not have the crossing point in their signal strength. Functionality of Intelligent Agents in WiFi positioning system DSP NAL
Excel sheet
Data cleaning statergies
Cleaned data
Fig.6.Data cleaning process Agents collect data from excel data sheet through the interaction component and send it to the information fusion component which is responsible of receiving data. Data collected through excel data sheet may have some erroneous signal coordinates where we cannot locate either the DSP or NAL. These signal coordinates collected through information fusion component are transferred to information processing component. Information processing component is responsible of computing the error distance between access points. If there is an error between the access points then those access points are
Velammal College of Engineering and Technology, Madurai
transferred to the action component of the agent, which in turn stores the details of the access point coordinates in its knowledge base as action. The error free access point coordinates alone be transferred as output of the action component through the interaction component. At this stage the data processing handles only those access points which are error free causes improvement in average signal strength computation. Ultimately the computation time required to accomplish the MATLAB simulation will also be greatly be reduced. Hence, the use of agents in the wifi positioning system improves the performance in one or more aspects. Data cleaning Agent Procedure PDA receives signals from the access points. The netstumbler software detects the signal strength from the PDA and explores the signals into the excel data sheet. This data sheet is given as an input to the Data Preprocessing component. The preprocessing component is to clean the given data sheet with the help of intelligent agents. Agents are incorporated in the intelligence component to improve the performance of the system. Agents analyze the data access points carefully and compute the error distance to each of the access points. If agents encounters an erroneous data then agents learns to update its knowledge base. Clusters are formed using the erroneous data that are stored in the knowledge base. Before sending the details of the access points to the action component, agent stores all such erroneous coordinates of the access points in its knowledge base. Data cleaning is a treatment for decreasing possible negative influence[3]. An outlier is an observation which deviates so much from other observations as to arouse suspicions that was generated by a different mechanism. The goal of outlier detection is to uncover the “different mechanism”. If samples of the different mechanism exist in knowledge base, then build a classifier to learn from samples. For fine classification, a function that designates the distribution of samples is necessary. Agents are intelligent systems that verify the input data (access points) from its knowledge base to identify the outlier. If the access point coordinates are error free then the agent uses the access point coordinates to position the laptop. Otherwise the agent filters the details of erroneous access points by not sending them to the next layer. Use of agents in this process eliminates the computation time required to find the erroneous data using MATLAB. Agent based data cleaning eliminates the computation time associated with the detection of erroneous data which causes performance improvement in the system.
Page 28
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
What makes the difference in the performance? In the existing system the work process is done manually and the errors are detected individually until the positioning is done. Though the manual work progress successfully, the work process is repeated for all the access point signal strength. To overcome this repetition problem and not go beyond the cross point’s intelligent agents are introduced in the cleaning system. Agent makes use of the error distances stored in its knowledge base to clean the erroneous signals.
Fig.7. Agents (Intelligent Process) in positioning process 5. Results and Performance evaluation What the existing system will do? In the existing system, access points are positioned with the help of GIS and the errors positions are to be ignored are calculated with the help of MATLAB. All the operations taken individually and the signals are transformed into the excel sheet with the help of the software Netstumbler. Error distance is calculated for each of the floor positions using MATLAB and the error signal strengths are stored in the form of matrix. The next step is to position the access point according to the erroneous distance. This process to done manually and every signal strength are stored separately requires more time and makes the work process as a long one.
The process is done for all the areas and the agent stores only the erroneous coordinates in the knowledgebase. Time is saved by the use of agent while experimenting this process. The accuracy of the positioning is improved when agents are incorporated in the system with is the aid of geographical information system. In the existing system performance is measured by means of error distance which requires the help of GIS but for the agent based system error distance is calculated and stored in the knowledgebase.
Fig. 8. Floor plan with corridor dimension What Agent based system is doing? The main objective to introduce the agents in the positioning system improves the performance by cleaning the erroneous data and to store it in the knowledge base for further processing. The accuracy and the time required to position the lap top is reduced. The signals are broadcasted from the access point and the signals make use of the Netstumbler to export the signal into excel sheet. Agent reads the excel sheet and store only the error details in its knowledge base. Agent makes use of its knowledge base to verify the error details and quit the work process for erroneous data. Reduction of processing time by the use of agents not only improves the performance but also not repeating the operations which are already done during the cleaning task will also be considered as one of the performance improvement factor. Thus Agents present the data cleaning system enable the system with intelligent behavior and improves positioning strategies.
Velammal College of Engineering and Technology, Madurai
The above Fig 7 displays a floor plan of our test site along with specific x and y dimensions of corridor section. Now the X and Y ranges from 0 and 19.125 and from 9.25 to 11.125 meters respectively. Use of agent in positioning system improves the accuracy and reduces the working time of the process. Also it scopes down and eliminates unlikely intersection area from positioning algorithm in the optimization phase.
6. Conclusion In this paper, Intelligent Agents are introduced to clean the erroneous access point coordinates for the wifi positioning system that uses the knowledge of geographical Information system. The performance variation of the wifi positioning system is described in terms of agent based and normal positioning. Agents are introduced between the excel data sheet and the MATLAB simulation data
Page 29
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
processing component. Future enhancement of this paper may introduce Multi-Agents to perform the cleaning task with distinct functionality.
7.References
[13] Walter Brener,Rudiger Zarnekow, “Software Agents Foundation & Application”, Morgan publishers,Elsevier,2002 [14]William E Winkler Conference SIGKDD’03
“Data
Cleaning
Methods:,
[1] Ayse Yasemin SEYDIM “Intelligent Agents: A Data Mining Perpespective”,CiteSeer .IST, Literature Digital Library. 1999. [2] Bahi,P , Padmanabhan V RADAR ; An In-building RF based User Location and Tracking System, IEEE Infocom ,March 2000 [3] S.D.Bay and M.Schwabacher. Mining distance-based Outliers in near linear time with randomization and a simple pruning rule. In Proceedings of Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Washington, D.C.USA, pages29–38, 2003. [4]Erhard Rahm, Hong Hai Do, “ Data Cleaning : Problems and Current approaches” [5] Gerhard Weiss , “ Multiagent Systems, “A Modern Approach to Distributed Artificial Intelligence”, The MIT Press, , 1999 [6]Hightower,J and Borriello G “Location Systems for Ubiquitous Computing”, IEEE Communic. Mag. August 2001, 57-66 [7] Hightower J and Borriello G, “Location Sensing Techniques”, IEEE Communic. Mag , August 2001 Technical Report [8] Jan R H, Lee Y.R “ An Indoor Geolocation System for wireless Lans” Proceedings of the International Conference Parallel Process Workshop (ICPPW’03) 2003. [9] Jiawei Han , Micheline Kamber “ Data mining: Concepts and Techiniques “ , Morgan Kaufmann Publishers, Elsevier, 2001 [10] http://www.netstumbler.com [11] S.Russell, P. Norvig, Artificial Intelligence, “ A Modern Approach, Printice-Hall,1995. [12]Tussanai Parthornratt , Kittiphan Techakittiroj, “ Improving Accuracy of WiFi positioning system by using Geographical Information System “, AU J.T. 10(1) : 38-44, Jul 2006
Velammal College of Engineering and Technology, Madurai
Page 30
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Framework for Multiple Classifier Systems Comparison (MCSCF) P.Shunmugapriya ,
S.Kanmani
Research Scholar, Department of CSE, Pondicherry Engineering College, Pondicherry, India
Professor & Head Department of IT Pondicherry Engineering College, Pondicherry, India
[email protected] Abstract In this Paper, we propose a new framework for comparing Multiple Classifier Systems in the literature. A Classifier Ensemble or Multiple Classifier System combines a finite number of classifiers of same kind or different types, which are trained simultaneously for a common classification task. The Ensemble is able to efficiently improve the generalization ability of the classifier when compared with a single classifier. The objective of this paper is to introduce a framework MCSCF that analyses the existing research work on Classifier Ensembles. Our framework compares the classifier ensembles on the basis of Constituent classifiers of an ensemble and Combination methods, Classifier selection basis, Standard datasets, Evaluation criteria and Behavior of classifier ensemble outcomes. It is observed that, different types of classifiers are combined using a number of different fusion methods and classification accuracy is highly improved in the ensembles irrespective of the application domain. Keywords - Data Mining, Pattern Classification, Classifier Ensemble, Multiple classifier Systems
I. INTRODUCTION Combining classifiers to get higher classification accuracy is rapidly growing and enjoying a lot of attention from pattern recognition and machine learning communities. For ensembling, the classification capabilities of a single classifier are not required to be very strong. What is important is to use suitable combinative strategies to improve the generalization of classifier ensembles. In order to speed up the convergences and simplify structures, the combinative components are often “weak” or simple [1]. Classifiers like Region Based Classifier ,Contour Based Classifier, Enhanced Loci classifier, Histogram-based classifier , Crossing Based Classifier, Neural Networks classifier, K-nearest neighbor classifier, SVM, Anorm, KDE, KMP(Kernel Matching Pursuit), Minimum Distance Classifier, Maximum Likelihood Classifier, Mahalanobis Classifier, Naïve Bayesian, Decision tree, Fisher classifier, Nearest Means classifier are often combined to form classifier fusion or classifier ensemble as shown in [Table I].
Velammal College of Engineering and Technology, Madurai
[email protected]
Also there are many classifier combination methods available. By using different combination methods to a number of various classifiers, classifier ensembles result with variations in their performance and classification accuracy. So far, different methods of classifier ensembles have been experimented on different application domain. The important objective of this paper is to provide a comparison framework of works on classifier combination .i.e. what classifiers are used in ensembles, what are the methods of combination that have been applied, what application domain and datasets are considered. [2] and [9] are classifier ensemble works carried out in context aware framework using evolutionary algorithm and [2] on line proposal. [3] uses voting method to form an ensemble that classifies Hyper spectral data. [5] considers distribution characteristic and diversity as the base features for classifier selection and uses k-means algorithm, a new kernel based probability c-means (KPCM) algorithm is designed for classifier ensemble. [6] is double bagging, a variable method of bagging .This uses SVM with 4 kernels and has shown enhanced classification accuracy than bagging and boosting. A-priori knowledge about classifiers is considered as an important criterion in selecting the classifiers [7] and uses a number of classifiers and ensemble methods to form an ensemble that has always 90% accuracy of classification. Boosting method is employed to obtain diversity among the classifiers and uses very effective method of ensembling - stacked generalization [8]. In [10] feature selection and GA are used and the ensemble performs well in accurate prediction of cancer data. For the area of Remote sensing, [11] considers diversity as an important feature and works on satellite image. [12], [14], [18] and [19] are also works carried out giving importance to diversity of classifiers. [16] is based on feature selection that uses product rule for the ensemble. Certain works are carried out in interest on looking at the performance of classifier ensembles, by using several combinations of classifiers and then look for the best combination [17], [24] and [25].
Page 31
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The text is organized as follows: Section II is about data mining, pattern classifier and classifier ensemble. Section III presents the comparative framework of works on classifier ensemble. In section IV, we discuss the comparison of ensemble results and our inference from the framework. Section V is our conclusion. II. DATA MINING Data mining or Knowledge Discovery in Databases (KDD) is a method for data analysis developed in recent years. It is an interdisciplinary research focusing upon methodologies for discovering and extracting implicit, previously unknown and potentially useful knowledge (rules, patterns, regularities as well as constraints) from data. The mining task is mainly divided into 4 types: Class/Concept description, Association analysis, Classification or Prediction and Clustering analysis. Among them, the classification task has received more attention. In the task of classification, the mining algorithm will generate the exact class description for the classification of unknown data by analyzing the existing data. A. Pattern Classification Classification is the action of assigning an object to a category according to the characteristics of the object. In Data Mining, Classification refers to a form of Data analysis which can be used to extract models describing important data classes or to predict future data trends. i.e. To classify objects(or patterns) into categories(or classes). Classification has 2 stages: • Model is determined from a set of data called training data. i.e. classes have been established before hand. The model is known as rules, decision trees or mathematical formula. • Correctness of the evolved model is estimated. This is done by studying the results of the evolved model’s function on a set of data called test data. Classes of the test data are determined before hand. Classification techniques that are available are Decision Tree Learning, Decision Rule Learning, Naive Bayesian Classifier, Bayesian Belief Networks, Neural Networks, kNearest Neighbor Classifier, Support Vector classifier. These techniques differ in the learning mechanism and in the representation of the learned model. Classification algorithms are Decision tree, Correlation analysis, Bayesian classification, Neural networks, Nearest neighbor, Genetic Algorithms, Rough sets and Fuzzy
Velammal College of Engineering and Technology, Madurai
technology. Also methods like Support Vector Machine (SVM), Neural Networks, K Nearest Neighbor, K Means, and Boosting and Quadratic classifiers exist.. SVM is the most preferred method in many classification and machine learning tasks. Also many standard tools like WEKA, KNIME, Rapid Miner, R Project, MATLAB tools are available for classification. Several frameworks like LIBLINEAR and LIBSVM are useful resources available on the web for classification. Many people and industries are interested in the decision support system and prediction systems for the better choice and risk. Classification forms the basis of these systems. B. Multiple Classifier Combination Though many classifiers exist and are widely used for classification in different application areas, the classifiers are not able to give a good accuracy of classification (Performance accuracy). To enhance the performance of classification algorithms, many works were carried out and there comes the idea of combining different classifiers to perform the classification task. Such a combination is called by names like Combination of Multiple Classifiers, Committee Machine, Classifier Ensemble and Classifier Fusion. The most popular usage is Classifier Ensemble. Combining classifiers to get a higher accuracy is an important research topic in the recent years. Classifier Ensemble is a learning paradigm where several classifiers are combined in order to generate a single classification result with improved performance accuracy. There are many methods available for combining the classifiers. Some of the popular ensembling methods are Majority voting, Decision templates, naïve Bayesian, Boosting, Bagging etc. There are many proposals for classifier ensembles and also works are still under pursue. To have an idea about what are the types of ensembles that have been proposed so far, we have given a comparative framework of existing ensemble methods in section III. [1].3
COMPARATIVE FRAMEWORK OF WORKS ON CLASSIFIER ENSEMBLE
The framework for comparing the existing classifier ensembles is given in Table I. Features for classifier selection, types of classifiers, combination methods, datasets, evaluation criteria are the important features discussed in the framework. Ensemble works are carried out in 2 perspectives: a. To look for the performance of ensemble, by combining a number of different classifiers using various combination methods. In certain works, the classifiers that are to be put into combination are selected on the basis of several features like,
Page 32
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
diversity, accuracy, feature selection, feature subspaces, random sampling etc.. b. Application of different ensemble works to a number of application domain and several standard datasets and conclude the performance of the ensemble. A. Classifier Selection Basis Classifiers are selected for a particular ensemble, based on several features as specified in the 3rd column of Table I. The features are Prior knowledge about classifiers, Diversity of classifiers, Feature selection, Reject option and Random subspaces. L.R.B.Schomaker et al. [7] considers Prior knowledge about the behavior of classifiers as an important factor. A prior knowledge about classifiers performing classification in single or when put into combination the way they behave, plays an important role in his work. Another criterion for selection of classifiers is Diversity of classifiers. The diversity of the classifier outputs is therefore a vital requirement for the success of the ensemble. [1]. Diversity indirectly talks about the correctness of the individual classifiers that are to be combined. [5], [8], [11], [12], [14], [19] are some samples that say about ensemble works in which, classifiers are selected on the basis of diversity of classifiers. The diversity among the combination of classifiers is defined as: if one classifier has some errors, then for combination, we look for classifiers which have errors on different objects [1]. There are two kinds of diversity measures: pairwise and nonpairwise. Pairwise measures are taken ,considering two classifiers at a time and Non pairwise measures are taken considering all the classifiers at one time [23]. Both these measures have different evaluation criteria. Diverse classifiers for any work are obtained by bootstrapping of the original dataset, resampling of training set and adoption of different feature set with Bagging [20], Boosting [21] and Random Subspace method [22]. When classifiers are more diverse, it is likely to expect very good classification accuracy from them. This is because the more, the diversity of the classifiers, they are very good finding different error conditions. Feature Selection is another criterion for selecting the classifiers. Data objects contain a number of features. There may be some irrelevant features that are not helpful for classification. In that case, features that are important in identifying the uniqueness of an object are selected by using several feature selection algorithms. After deciding the features, classifiers that can work on this particular feature subset are selected and combined [10], [15] discusses the ensemble proposals on the basis of feature selection
Velammal College of Engineering and Technology, Madurai
Classifiers of an ensemble are also selected based on the assumption that they are independent of each other and also has some diversity [5]. Distribution characteristic is difference in performance of classifiers in different region of characteristic space. Apart from performing the task of classification correctly, classifiers should also be capable of identifying data objects that do not belong to a particular class and capable of rejecting them. When such classifiers are selected for the ensemble, it is called as classifiers with Reject option. Apart from the above features, classifiers are also selected related to the type of application domain and the method of ensemble to be applied. Certain ensembles are fit into context aware framework and in most of such cases evolutionary algorithm is used for classification leading to a very good classification performance. [2] and [9] discusses such type of methodology. GA plays a very important role in combining classifiers [11]. When GA is the method of ensemble to be applied, classifiers that can work well in GA framework are selected for the ensemble. B. Classifiers for Ensemble Different kinds of standard classifiers exist and most of the classifiers perform well when put into an ensemble. Region Based Classifier, Contour Based Classifier, Enhanced Loci classifier, Histogram-based classifier, Crossing based classifier, Nearest Neighbor, K- Nearest Neighbor ( K taking values 1,2,3 etc), Anorm, KDE, Minimum Distance classifier, Maximum Likelihood Classifier, Mahalanobis Classifier , SVM, KMP(Kernel Matching Pursuit), Decision Tree, Binary Decision Tree, Nearest Mean Classifier, Maximum Entropy Model, Heterogeneous Base Learner, Naïve Bayesian classifier, Linear Discriminant Analysis, Non-Linear Discriminant Analysis and Rough Set Theory are the classifiers used in most of the ensemble applications. When diversity is the base feature considered, then by using methods like Bagging, Boosting, Random Sampling, the classifiers are trained and thereby the classifiers are diversified. The diverse classifiers are then put into ensemble. In case of Context Aware Framework, the classifiers are trained with different context conditions before put into ensemble [2], [9]. Cost sensitivity is considered as the important factor for a classifier and in such cases the classifier is trained to possess cost sensitivity [18]. SVM is one such classifier trained to have cost sensitivity and it is CS-SVC.
Page 33
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
D. Datasets for Classifier Ensemble SVM is used as such in many ensembles. There are 4 kernel functions with which SVM can combine and yield very good classification accuracy: Linear Kernel, Polynomial Kernel, Radial Basis Kernel and Sigmoid kernel. SVM with Radial Basis Kernel yields a very good performance.
C. Classifier Combination Methods Majority Voting, Weighted Majority Voting, Naïve Bayesian Combination, Multinomial Methods, Probabilistic Approximation, Singular Value Decomposition, Trainable Combiners, Non Trainable Combiners, Decision Templates, Dempster-Shafer method, Bagging, Boosting and Stacking(Stacked Generalization) are the methods available for combining the classifiers according to Ludmila I Kuncheva [1].
Datasets that are used for an individual classifier can serve as the datasets for ensemble too. Same as classifiers, ensembles have their usage in all most all the areas like Machine Learning, Image processing, Bioinformatics, Geology, Chemistry, Medicine, Robotics, Expert and Decision making systems, Diagnosis systems, Remote sensing etc. According to the type of application domain, datasets are taken from various available standard Databases. UCI datasets [12],[14],[16] and [17], ELENA database[5] and [8], EFERET, EYALE, EINHA [2] and [9] are some of the benchmark datasets that are used for experiments in most of the ensemble works. For remote sensing and land cover classification, satellite images are used. Some Researchers collect datasets on their own for their works. E. Evaluation Methods
Out of the available methods, Majority Voting, Dempster-Shafer method, Boosting, Stacking and Bagging are the popular methods which are used in most of the ensemble works [6],[7],[8],[12],[14],[17],[18],[19]. Bayesian decision rules are another popular method of ensembling. The rules are Product Rule, Sum rule, max rule, min rule and median rule.[11]. Product Rule is the rule which is most commonly used among the 5 rules. [16]. A multinomial method has two types: Behavior Knowledge Space (BKS) Method and Wernecke’s Method. Out of these 2 methods BKS is widely used [7]. In Boosting, Adaboost algorithm is the most commonly used ensemble algorithm [12]. Apart from this, GA is used for forming the classifier ensemble [2], [5], [9] and [10]. Classifiers are also fit into context aware framework for the task of ensembling [2] and [9]. In several cases, a method which is capable of integrating the advantages of existing classifiers is used for combining the classifiers. Another way is initially after classification is done by some classifiers, an additional classifier either of the same type or of a different kind is used finally for classification [3].
The performance of a classifier is a compound characteristic, whose most important component is the classification accuracy [1]. Some of the standard measures that are very often used for evaluating the performance of classifiers and ensembles are Cross validation(confusion matrix as a result of Cross validation), Kappa statistics, McNemar’s test, Cochran’s Q test, F-test, OAO(One-Against-One) approach, OAA(One-Against-All) approach, ROC(Receiver Operating Characteristic), AUC(Area Under the ROC Curve), FAR, FRR, Multinomial Selection Procedure for Comparing Classifiers [9], [10], [11]. [14], [16], [17], [19],[24]. Cross-Validation is the most popular method for estimating the classification accuracy. Some of the variants of Cross-Validation obtained by varying the training and testing data are: K-Hold-Out Paired t-Test, K-Fold Cross-Validation Paired t-Test, Diettterich’s 5 x 2-Fold Cross-Validation Paired t-Test (5x2cv). IV. RESULTS AND DISCUSSION Most of the works discussed here are carried out in order to enhance the performance of multiple classifiers ensemble and simplify the system design. Different classifiers are combined using a number of different classifier combination methods and has been experimented with standard datasets.
A new ensemble method is designed, a new ensemble algorithm is proposed or modifications are made to existing ensemble methods and algorithms in certain cases without using the existing methods [12], [16] and [24].
It is seen from [2], [3],[5],[6],[7],[8],[10],[12],[14],[16],[17][19] and [24] the new methods designed are highly robust and show improved classification accuracy than the existing ensemble methods and the individual constituent classifier. Most of the new methods show a better performance compared to individual classifiers and
Velammal College of Engineering and Technology, Madurai
Page 34
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the ensembles Bagging, Boosting, Rotation Forest and Random Forest. Not only that, the running time is improved in [14] and the training time for the ensemble is reduced by 82% without any performance degradation [18]. Also the new methods designed using Evolutionary algorithm in Context aware framework has resulted in ensembles with highest recognition rate and reduced error rates [2] and [9]. It could be inferred from the framework that: • Feature Selection and Diversity of classifiers play an important role in selecting the classifiers for an ensemble. • 75% of the ensemble works uses Cross-Validation for evaluating the performance of classifier ensemble. • Bayesian Decision Rules, Bagging, Boosting, Stacking and voting are the methods most commonly used for combination of classifiers. • UCI datasets are taken for most of the applications. • 80% of the works are on two-class classification problem. • WEKA and MATLAB are the tools used for the implementation of most of the classifier works. • 100% Classification accuracy has not been achieved in any ensemble work.
V. CONCLUSION AND FUTURE WORK A classifier ensemble is a very successful technique where the outputs of a set of separately trained classifiers are combined to form one unified prediction. First it improves the generalization performance of a classification system greatly. Second, it can be viewed as an effective approach for classification as a result of its variety of potential applications and validity. Although classifier ensembles have been used widely, the key problem for researchers is to effectively design the individual classifiers that are highly correct, with diversities between them and thereby the ensemble is also highly accurate. We have given a framework that consolidates the types of existing ensemble methods. It uses classifier selection features, types of classifiers and combination methods for comparison. Any new work can easily fit into our framework and will be useful to the researchers for making an analysis of existing work. As a Future work, Ensembles can be tried for 100% classification accuracy and for many unexplored application domain. Most of the works on binary classification can be extended for multi class classification.
Table I: Comparative Framework of Works on Classifier Ensemble Ref. No
[7]
On the Basis of
A-Priori Knowledge
Classifiers Used
Method of Ensembling
1.Region Based Classifier 2.Contour Based Classifier 3.Enhanced Loci 4.Histogrambased 5.Crossing based
1.Majority Voting(MV) 2.Dempster Shafer (DS) 3.Behavioral Knowledge Space(BKS)
Dataset
Hand Written Numerals
4 UCI datasets : 1.Pima [12]
Diversity of Classifiers
1.Knn (k=1) 2.Knn (k=3) 3.Svm (r=1) 4.Anorm 5.Kde.
Sample Size
*
764 patterns 8 features 2 classes
AdaBoosting 2.Spam
4016 patterns 54 features 2 classes
3.Haberman
2 classes
Velammal College of Engineering and Technology, Madurai
Evaluated By Comparing the classification results with that of constituent classifiers
Comparing the classification results with that of constituent classifiers
Results 1.Classificati on Accuracy ≈ 90% 2.Better Accuracy than the constituent classifiers Better Performance of the new method than the individual classifiers of the ensemble
Page 35
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
4.Horse-colic
[8]
[11]
[14]
[19]
Diversity of Classifiers
3 different classifiers diversified by Boosting
1.Boosting 2.Stacked Generalizati on
SATIMAGE Dataset from ELENA Database
Diversity of Classifiers
1.Minimum Distance Classifier 2.Maximum Likelihood Classifier 3.Mahalanobis Classifier 4.K nearest Neighbor Classifier
Five Bayesian Decision rules: 1.Product rule 2. sum rule 3. max rule 4. min rule, 5. median rule
Remote Sensing Image –SPOT IV Satellite Image
Diversity of Classifiers
1.Diversity of Classifiers
SVM KMP(Kernel Matching Pursuit)
1.Maximum Entropy Model 2.Heterogeneous
Bagging and Majority voting
L1 Regularized maximum Entropy Model
I.UCI Datasets 1.Waveform 2.Shuttle 3.Sat II. Image Recognition6 Plane Class Images
Customer Relationship Management ORANGE – A French Telecom
Velammal College of Engineering and Technology, Madurai
6435 pixels 36 attributes 6 classes
Comparing the classification results with that of constituent classifiers
6 Land Cover classes
1.Overall Accuracy 2.Kappa statistics 3.McNemar’s Test 4.Cochran’s Q Test 5.F-Test
614 Sheets 128 X 128 pixels
1.OAO Approach 2. Comparing the classification results with that of constituent classifiers
2 Versions 1.Larger Version15,000 Feature Variables 50,000
1.By Cross Validation 2. Overall Accuracy
1.Boosting generates more diverse classifiers than Cross validation. 2.Highly Robust compared to original Boosting and Stacking 1. Diversity is not always beneficial. 2.Increasing the number of Base classifiers in the ensemble will not increase the Classificati on accuracy. 1. A New Ensemble of SVM and KMP is designed. 2. High Classificati on Accuracy of SVM. 3. Quick Running Time of KMP. 1. Good Classificati on Accuracy. 2.Has won 3rd place in
Page 36
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Base Learner 2.Best Ensemble Proposal KDD Cup 2009
3.Naïve Bayesian
Boosting
MODL CriterionSelective Naïve Bayesian
Company’s Dataset 3 tasks (Churn, Appetency, Upselling)
examples
Ensemble proposal of KDD Cup 2009
2.Smaller Version: 230 features 50,000 examples
Post processing the results of 3 methods with SVM
[5]
[18]
[10]
[16]
1.Distrib ution Character istics 2.Diversi ty
1.Feature Subspace s 2.Diversi ty
Feature Selection
Feature Selection
Classifiers with Diversity
Cost Sensitive SVC s
1.Fisher classifier 2.Binary Decision Tree 3. Nearest mean Classifier 4.SVM 5.Nearest Neighbor(1-nn)
‘n’ number of SVM s
1. Kernel Clustering 2. New KPCM Algorithm
Phenome Dataset from ELENA Database
2 classes 5000 samples 5 features
Tr.Set 7931 positive 7869negative
In comparison with Bagging applied to the same dataset.
Comparison with Conventional SVCs in terms of Detection Rate and Cost Expectations
1.Random Subspace method 2.Bagging
Hidden Signal DetectionHidden Signal Dataset
Ensemble Feature Selection based on GA
1.Colon cancer data
Tr.Set – 40 Tst.Set– 22
1.10- Fold Cross Validation and
2.Hepato Cellular Carcinoma Data
Tr.Set – 33 Tst.Set– 27
2. Assigning Weights
Tst.Set 9426 positive 179,528 negative
Tr.Set – 21 Tst.Set– 29
Better Performance than bagging and any other constituent ensemble classifier 1. SVC parameter optimizations reduced by 89%. 2. Reduction in overall Training time by 82% without performance degradation.
Better prediction Accuracy
3.High grade glioma dataset
Product Rule
12 UCI benchmark Datasets
Velammal College of Engineering and Technology, Madurai
*
4- Fold CrossValidation
1.Simplifie d Datasets 2.Reduced time complexity, 3.A new FSCE
Page 37
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[9]
[2]
[17]
1.Evoluti onary Fusion 2.Context awarenes s 1.Evoluti onary Fusion 2.Context awarenes s
*
Best classifiers in each ensemble obtained by GA
1.K-Means Algorithm 2.GA
‘n’ number of different classifiers trained with different context conditions
Embedding Classifier Ensemble into Context Aware Framework
1.Naïve Bayesian 2.K-Nearest Neighbor
1.Static Majority Voting(SMV ) 2.Wighted Majority Voting(WM V) 3.Dynamic WMV(DW MV)
1.Decision Tree
[6]
*
2. SVM With 4 kernels: 1. linear, 2.polynomial 3.radial basis 4.sigmoid
1.Bagging 2. Double Bagging
Face Recognition FERET Dataset 4 Face Recognition Systems 1.E- FERET 2.E-YALE 3.E-INHA 4.IT (All 4 datasets are further divided into 3 separate datasets I, II,III )
6 contexts
Nine data sets Each containing 10000 face images under 100 kinds of illumination
1.ROC 2.FAR 3.FRR
By creating similar offline system(without Result Evaluator)trained and tested on the same dataset
UCI DataSets 1.TicTacToe EndGame 2.Chess EndGame
1.Condition Diagnosis of Electric Power ApparatusGIS dataset 2. 15 UCI Benchmark Datasets
Velammal College of Engineering and Technology, Madurai
958 Instances 9 features 2 classes 3196 Instances 36 features 2 classes
Each dataset has different number of objects, classes and features
Cross Validation
In comparison with Other ensembles’ performance on the same data.
algorithm is proposed 4.Higher Classificati on Accuracy 1.Reduced Error Rates 2.Good Recognition Rate 1.Highest Recognition Rate than the individual classifiers 2.Most Stable Performanc e 1.Better Classificati on Accuracy than the individual Classifier. 2. DWMV has Higher classificatio n Accuracy. Better performanc e than popular ensemble methods like Bagging, boosting, Random forest and Rotation
Page 38
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Forest.
[3]
[24]
[25]
*
‘n’ number of SVMs
*
*
Any classifier capable of classifying CSP
1.Rough Set Theory 2.Decision Tree 3.SVM
Additional SVM is used to fusion the outputs of all SVMs
HyperSpectral AVIRIS Data
1.Majority Voting 2.Rejecting the Outliers
BCIEEG signals
Integrating the advantages of RST, DT and SVM
UCITeaching Assistant Evaluation (TAE) Dataset
220 data channels, 145 x 145 pixels, 6 land cover classes
*
151 Instances 6 Features 2 classes
1.Overall classification Results compared to an individual classifier. 2.By simple voting.
20 – Cross Validations
1. 102(68%) – Tr.Data 49(32%) -Tst. Data 2. 6-Folds Cross Validation
Good Accuracy than a single constituent classifier in the Ensemble 1.A new ensemble CSPE is designed 2.Better performanc e than LDA , RLDA and SVM 3.Average accuracy of 83.02% in BCI (Comparitiv ely a good Accuracy) 1.Improved Class Prediction with Acceptable Accuracy. 2. Enhanced Rule Generation
Note: * - Undefined in the literature REFERENCES:
[2]. Zhan Yu, Mi Young Nam and Phill Kyu Rhee, “Online Evolutionary Context-Aware Classifier Ensemble Framework For Object Recognition”, Proceedings of the 2009 IEEE International Conference on Systems, Man, and
Cybernetics,San Antonio, TX, USA - October 2009, Page(s): 3428 – 3433, 2009 [3]. Jón Atli Benediktsson, Xavier Ceamanos Garcia, Björn Waske, Jocelyn Chanussot, Johannes R. Sveinsson, and Mathieu Fauvel, “Ensemble Methods For Classification Of Hyperspectral Data”, IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2008, Volume: 1, Page(s): I-62 - I-65, 2008 [4]. Gao Daqi, Zhu Shangming, Chen Wei and Li Yongli “A Classifier Ensemble Model and Its Applns” IEEE
Velammal College of Engineering and Technology, Madurai
Page 39
[1]. Ludmila I. Kuncheva,” Combining Pattern Classifiers Methods and Algorithms” a john wiley & sons, inc., publication, tk7882.p3k83 2004
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
International Joint Conference Proceedings on Neural Networks, Volume: 2, Page(s): 1172 – 1177, 2005 [5]. Cheng Xian-yi and Guo Hong-ling, “The Technology of Selective Multiple Classifiers Ensemble Based on Kernel Clustering”, Second International Symposium on Intelligent Information Technology Application, Second International Symposium on Intelligent Information Technology Application, Volume: 2, Page(s): 146 – 150, 2008 [6]. Faisal M. Zaman Hideo Hirose, “Double SVMBagging: A New Double Bagging with Support Vector Machine”, Engineering Letter, 17:2, EL_17_2_09, (Advance online publication: 22 May 2009) [7]. In: L.R.B. Schomaker and L.G. Vuurpijl (Eds.), “Classifier combination: the role of a-priori knowledge”, Proceedings of the Seventh International Workshop on Frontiers in Handwriting Recognition, September 11-13, 2000, Amsterdam, ISBN 90-76942-01-3, Nijmegen: International Unipen Foundation, pp 143-152. [8]. Nima Hatami and Reza Ebrahimpour , “Combining Multiple Classifiers: Diversify with Boosting and Combining by Stacking”, IJCSNS International Journal of Computer Science and Network Security, VOL.7, No.1, January 2007 [9]. Suman Sedai and Phill Kyu Rhee “Evolutionary Classifier Fusion for Optimizing Face Recognition”, Frontiers in the Convergence of Bioscience and Information Technologies, Page(s): 728 – 733, 2007 [10]. Kun-Hong Liu, De Shuang Huang “Microarray data prediction by evolutionary classifier ensemble system” Kun-Hong Liu; De-Shuang Huang; Jun Zhang; IEEE Congress on Evolutionary Computation, Page(s): 634 – 637, 2007 [11]. Man Sing WONG, and Wai Yeung YAN “ Investigation of Diversity and Accuracy in Ensemble of Classifiers using Bayesian Decision Rules” , International Workshop on Earth Observation and Remote Sensing Applications, Page(s): 1 – 6, 2008 [12]. Abbas Golestani , Kushan Ahmadian Ali Amiri,and MohammadReza JahedMotlagh “ A Novel Adaptive-Boost-Based Strategy for Combining Classifiers Using Diversity Concept” 6th IEEE/ACIS International Conference on Computer and Information Science (ICIS 2007), Page(s): 128 – 134, 2007 [13]. Gou Shuiping ,Mao Shasha and Licheng Jiao,” Isomerous Multiple Classifier Ensemble Method with SVM and KMP “ International Conference on Audio, Language and Image Processing, Page(s): 1234 – 1239, 2008 [14]. D.M.J. Tax and R.P.W. Duin “Growing a Multi Class Classifier with a Reject Option” Pattern Recognition Letters 29 (2008 ) 1565-1570 [15]. Bing Chen, Hua-Xiang Zhang “An Approach of Multiple Classifiers Ensemble Based on Feature Selection”
Fifth International Conference on Fuzzy Systems and Knowledge Discovery, Volume: 2, Page(s): 390 – 394, 2008 [16]. Dr.S.Kanmani, Ch P C Rao, M.Venkateswaralu and M.Mirshad “Ensemble Methods for Data Mining” IE (I) Journal-CP, Vol 90, May 2009, 15-19 [17]. Barry Y. Chen, Tracy D. Lemmond, and William G. Hanley, “Building Ultra-Low False Alarm Rate Support Vector Classifier Ensembles Using Random Subspaces” IEEE Symposium on Computational Intelligence and Data Mining, Page(s): 1 – 8,2009. [18]. Hung-Yi Lo, Kai-Wei Chang, Shang-Tse Chen, Tsung-Hsien Chiang, Chun-Sung Ferng, Cho-Jui Hsieh, YiKuang Ko, Tsung-Ting Kuo, Hung-Che Lai, Ken-Yi Lin, Chia-Hsuan Wang, Hsiang-Fu Yu, Chih-Jen Lin, Hsuan-Tien Lin, Shou-de Lin” An Ensemble of Three Classifiers for KDD Cup 2009: Expanded Linear Model, Heterogeneous Boosting, and Selective Naıve Bayes” JMLR: Workshop and Conference Proceedings 1: 1-16 KDD cup, 57-64, 2009 [19]. L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp.123-140, August 1996. [20]. Y. Freund and R. Schapire, “Experiments with a new boosting algorithm,” Proceedings of 13th International Conference on Machine Learning, Bari, Italy, pp. 148-156, July 1996. [21]. T. K. Ho, “The random subspace method for constructing decision forests,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 8, pp. 832844, August 1998.
Velammal College of Engineering and Technology, Madurai
Page 40
[22]. L. I. Kuncheva, and C. J. Whitaker, “Ten measures of diversity in classifier ensembles: limit for two classifiers,” DERA/IEE Workshop on Intelligent Sensor Processing, Birmingham, U.K., pp.10/1-10/6, February 2001. [23]. Xu Lei, Ping Yang, Peng Xu, Tie-Jun Liu, and DeZhong Yao, “Common Spatial Pattern Ensemble Classifier and Its Application in Brain-Computer Interface”, Journal of Electronic And Technology Since Of China, Vol. 7, No. 1, March 2009 [24]. Li-Fei Chen, “Improve class prediction performance using a hybrid data mining approach “, International Conference on Machine Learning and Cybernetics, Vol 1, pp 210-214, 2009
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Efficient Apriori Hybrid Algorithm For Pattern Extraction Process 1,2
J.Kavitha1,D.Magdalene Delighta Angeline2,P.Ramasubramanian3 Lecturers,Department of Computer Science and Engineering, Dr.G.U.Pope College of Engineering, Sawyerpuram-628251, Tamilnadu, India.,3 Professor, Vellammal College of Engineering and Technology, Madurai 1
[email protected] [email protected] 3 [email protected] 2
Abstract— Student’s placement in industry for the practicum training is difficult due to the large number of students and organizations involved. Further the matching process is complex due to the various criteria set by the organization and students. This paper will discuss the results of a pattern extraction process using association rules of data mining technique where Apriori Hybrid algorithm was chosen. The data use consists of Bachelor of engineering students in the teaching organization from the year 2009 till 2010. Keywords— Data mining, KDD, association rules, apriori Algorithm, Knowledge Discovery
student’s majoring. Usually, organization will request student with a specific majoring details. Other criterion is student’s Percentage. Also, due to certain work prospect, some organization request student based on the gender and race. These criteria have been considered by the program coordinator in the placement process to ensure the right student being sent to the right organization. This study aim to identify the patterns in matching organization and student and to extract hidden information from previously matched practicum placement datasets. This paper discusses the application of data mining technique particularly association rules to extract the historical placement pattern. This pattern will be a useful guideline for future organization and student matching. The data consist of all engineering undergraduate students from the year 2009 till 2010. We then discuss how the best features of Apriori and AprioriTid can be combined into a hybrid algorithm, called AprioriHybrid. The problem of finding association rules falls within the purview of database mining, also called knowledge discovery in databases.
VI. INTRODUCTION The teaching organization is responsible with the placement of students in the industry for the internship program. It is experiencing difficulty in matching organization‘s requirement with students profile for several reasons. This situation could lead to a mismatched between organization’s requirement and students’ background. Hence, students will face problems in giving good service to the company. On the other hand, companies too could be facing difficulties in training the students and assigning them with a project. The placement must be based on certain criteria in order to best serve the organization and student. For example, student who lives in Chennai should not be sent to an organization located in Bangalore. This is to avoid problems in terms of accommodation, financial, and social. It has been decided that practicum students’ should match the organization’s requirement. However, due to the large number of students registered every semester, matching the organization with the students is a very tedious process. The current procedures in matching organization and students involve several steps. First, the registered city1 (is the first choice for students) and city2 (is the second choice for students) will be examined. A match between organizations location and student’s hometown will be determined. The next criterion is the
VII. LITERATURE REVIEW Data mining have been applied in various research works. One of the popular techniques used for mining data in KDD for pattern discovery is the association rule [1]. According to [2] an association rule implies certain association relationships among a set of objects. It attracted a lot of attention in current data mining research due to its capability of discovering useful patterns for decision support, selective marketing, financial forecast, medical diagnosis and many other applications. The association rules technique works by finding all rules in a database that satisfies the determined minimum support and minimum confidence [3]. An algorithm for association rule induction is the Apriori algorithm, proven to be one of the popular data mining
Velammal College of Engineering and Technology, Madurai
Page 41
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
techniques used to extract association rules [4], implemented the Apriori algorithm to mine single-dimensional Boolean association rules from transactional databases. The rules produced by Apriori algorithm makes it easier for the user to understand and further apply the result. [5] Employed the association rule method specifically Apriori algorithm for automatically identifying new, unexpected, and potentially interesting patterns in hospital infection control. Another study by employed Apriori algorithm to generate the frequent item sets and designed the model for economic forecasting, presented their methods on modelling and inferring user’s intention via data. VIII. PATTERN EXTRACTION All paragraphs must be indented. All paragraphs must be justified, i.e. both left-justified and right-justified. A. Selection The data have been generated by different reports among others Registered Students Report, Students’ Mark Report, Students’ List Based on City Report. This data include all 2009 and 2010 Bachelor in Engineering students. The initial data contains the performance profile gathered from a number of 125 students with 20 listed attributes which include Register Number, Programme, Duration, Program Code, City1, City2, Address, Address State, Percentage, Gender, Race Code, Race, Organization, Address1, Address2, Postcode, City3 and State. The data contains various types of values either string or numeric value. The target is represented as Organization’s name. The Organization’s name was grouped according to two categories (Government and Private). Based on the discussion with the program coordinator, all 125 data are used in this study. The selected attributes are Majoring, Percentage, Gender, City1, Race, Organization and City3 chosen based on the suitability of the condition of the problems being discussed. The data were then processed for generating rules. B. Pre-processing No Upon initial examination on the data, missing values of the attributes City1, Percentage, Race, Gender, Organization and City3 were found and removed according to the numbers of missing values in one instance as part of the data cleansing process. C. Transformation According to [9], after the cleansing process, data is converted into a common format to make sure that the data mining process can be easily performed besides ensuring a
Velammal College of Engineering and Technology, Madurai
meaningful result produced. The following rules are used to transform the Percentage to string data. 1. If the Percentage = 81 Till 90 THEN Replace Percentage by S1 2. If the Percentage = 75 Till 80 THEN Replace Percentage by S2 3. If the Percentage = 70 Till 74 THEN Replace Percentage by S3 4. If the Percentage = 65 Till 69 THEN Replace Percentage by S4 Transformation has also been applied to attributes city1 and city3 by grouping several cities together according to their location or region, decoded into new region using code of each state. For example, KODAMPAKAM and GUINDY have the same code 02 then they were converted into one Region (N_Region). Organization’s name was also transformed by into two categories (Government and Private). After all pre-processing and transformation have been implemented, the data was than ready to be mined using association rules.
D. Pattern Extraction using Apriori Algorithm In this study, the association rules using Apriori and Apriori TID Algorithm was applied to the data for generating rules. E. Apriori Algorithm Fig.1 gives the Apriori algorithm. The first pass of the algorithm simply counts item occurrences to determine the large 1-itemsets. A subsequent pass, say pass k, consists of two phases. First, the large itemsets Lk-1 found in the (k-1)th pass are used to generate the candidate itemsets Ck, using the apriori-gen function. Next, the database is scanned and the support of candidates in Ck is counted. For fast counting, we need to efficiently determine the candidates in Ck that are contained in a given transaction t. 1) L1 = {large 1-itemsets}; 2) for ( k = 2; Lk-1≠0; k++ ) do begin 3) Ck = apriori-gen(Lk-1 ); // New candidates 4) for all transactions t ε D do begin 5) Ct = subset(Ck , t); // Candidates contained in t 6) for all candidates c ε Ct do 7) c:count++; 8) end 9) Lk = {c ε Ck | c:count _ minsup} 10) end 11) Answer = n Lk; Fig. 1 Algorithm Apriori
Page 42
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
F. Algorithm AprioriTid The AprioriTid algorithm, shown in Fig. 2, also uses the apriori-gen function to determine the candidate item sets before the pass begins. The interesting feature of this algorithm is that the database D is not used for counting support after the first pass. Rather, the set Ck is used for this purpose. Each member of the set Ck is of the form < TID; fXkg >, where each Xk is a potentially large k-item set present in the transaction with identifier TID. For k = 1, C1 corresponds to the database D, although conceptually each item i is replaced by the item set fig. For k > 1, Ck is generated by the algorithm (step 10). The member of Ck corresponding to transaction t is <t:T ID, fc 2 Ckjc contained in tg>. If a transaction does not contain any candidate k-item set, then Ck will not have an entry for this transaction. Thus, the number of entries in Ck may be smaller than the number of transactions in the database, especially for large values of k. In addition, for large values of k, each entry may be smaller than the corresponding transaction because very few candidates may be contained in the transaction. However, for small values for k, each entry may be larger than the corresponding transaction because an entry in Ck includes all candidate k-item sets contained in the transaction.
earlier in the study while the degree of rules coverage was shown through the value of support parameter. IV EXPERIMENTS In this experiment, the data has been grouped into three groups based on the Organization category. Again, the experiment was conducted using Apriori algorithm with the same specifications. Table 1 shows the results generated by Apriori for all two categories of organizations.
TABLE 1 EXTRACTED PATTERN BASED ON ORGANIZATION CATEGORY Organization
Region
Criteria (Apriori) Major=Computer Science and Engineering Percentage=75-80
1) L1 = flarge 1-itemsetsg; 2) C1 = database D; 3) for ( k = 2; Lk-1 ≠0; k++ ) do begin 4) Ck = apriori-gen(Lk-1 ); // New candidates 5) Ck =0; 6) for all entries t ε Ck-1 do begin 7) // determine candidate itemsets in Ck contained // in the transaction with identi_er t.TID Ct = {c ε Ck | (c - c[k]) ε t:set-of-itemsets ^ (c - c[k-1]) ε t.set-of-itemsets}; 8) for all candidates c ε Ct do 9) c:count++; 10) if (Ct ≠0 ;) then Ck += < t:TID;Ct >; 11) end 12) Lk = {c ε Ck | c:count _ minsup} 13) end 14) Answer = Ụk Lk;
Gender=Male Race = Guindy
N_Region1
Major=Electronics Communication and Engineering Percentage=75-80
Government
Gender=Male Race = Guindy
Major=Electrical and Electronics W_Region2
Engineering Percentage=75-80
Fig. 2 Algorithm AprioriTid
Gender=Male Race = Guindy
G Interpretation/ Evaluation
During the process of pattern extraction, the acceptance of the output produced was evaluated in terms of accuracy and converge. This is to make sure that the generated rules are reliable and accurate. The accuracy of rules was obtained according to the value of confidence parameter determined
Velammal College of Engineering and Technology, Madurai
N_Region1
Major=Computer Science and Engineering Percentage=70-74 or 75-80 Gender=Female or Male
Page 43
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Race = Guindy or Kodambakam W_Region2
Major=Electronics Communication and Engineering
Private
Percentage=70-74 Gender=Male Race = Guindy
Major=Electrical and Electronics Engineering Percentage=70-74 Gender=Male Race = Guindy
A Discussion on the Apriori result From the pattern extracted, it was found that Apriori algorithm could generate patterns that are believed to be the factors that affect the matching process. From the experiment, extraction of the hidden information reveals that organization requirement can be fulfilled based on only three or four criteria. The best rules were selected where the Organization was set as the target of the students. The rules were evaluated based on the confidence and support. Upon examining Table 1, example of pattern extracted is IF students are from the Computer Science and Engineering AND Their Percentage is between 75-80 AND They are Guindy THEN The students were placed in the Northern Region and In an Government Organization IF students are from the Electrical and Electronics Engineering and Electronics Communication and Engineering Majoring AND Their Percentage is between 70-74 AND They are Guindy THEN The students were placed in the Western Region and In a Private Organization
Velammal College of Engineering and Technology, Madurai
V CONCLUSIONS This study has been implemented and conducted on existing data from the teaching organization. In this study data mining techniques namely association rule was used to achieve the goal and extract the patterns from the large set of data. Using organization category as the target, the patterns extracted can provide information of the practicum placement and how the matching of the organization’s requirement and student’s criteria was done previously. Further analysis can be done by changing the target attributes. ACKNOWLEDGMENT
Magdalene Delighta Angeline is a Lecturer in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. She obtained her Bachelor degree in Information Technology from Anna University, Chennai in the year 2007 and she is doing Master degree in Computer and Information Technology in Manonmaniam
Page 44
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Sundaranar University, Tirunelveli. She has over 3 years of Teaching Experience and published two paper in national conferences. Her current area of research includes Image Processing, Neural Networks, and Data Mining. Email: [email protected]. J.Kavitha is a Lecturer in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. She obtained her MCA from IGNOU and Master degree in Engineering from Anna University, Chennai in the year 2004 and 2007 respectively. She has over 3 years of Teaching Experience and published 2 papers in National conferences. Her current area of research includes Image Processing, Image Restoration and Data Mining. Email: [email protected]. P.Ramasubramanian, is Professor and Head in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. He obtained his Bachelor and Master degree in Computer Science and Engineering from M.K.University, Madurai in the year 1989 and 1996 respectively. He has submitted his Ph.D thesis to Madurai Kamaraj University, Madurai. He has over 22 years of Teaching Experience and authored 15 books and 22 research papers in International, National Journals and Conferences. His current area of research includes Data Mining, Data Ware housing, Neural Networks and Fuzzy logic. He is a member of various societies like ISTE, International Association of Engineers, Computer Science Teachers Association, International association of Computer Science and Information Technology and Fellow in Institution of Engineers (India). Email: [email protected], [email protected].
[4] Agrawal, R., C. Faloutsos, and A. N. Swami (1994). Efficient similarity search in sequence databases. In D. Lomet (Ed.), Proceedings of the 4th International Conference of Foundations of Data Organization and Algorithms (FODO), Chicago, Illinois, pp. 69-84. Springer Verlag [5] Ma, Y., Liu, B., Wong, C. K., Yu, .S., & Lee, S. M. (2000). Targeting the Right Student Using Data Mining , ACM, PP. 457-463. [6] R. Agrawal, T. Imielinski, and A.Swami.Database mining: A performance perspective.IEEE Transactions on Knowledge and Data Engineering, 5(6):914{925, December 1993. Special Issue on Learning and Discovery in Knowledge-Based Databases. [7] Almahdi Mohammed Ahmed , Norita Md Norwawi , Wan Hussain Wan Ishak(2009), Identifying Student and Organization Matching Pattern Using Apriori Algorithm for Practicum Placement, International Conference on Electrical Engineering and Informatics ,Selangor, Malaysia. [8] Jiawei Han, Micheline Kamber. "Data Mining : Concepts and Techniques " book: Data mining (2001). [9] Zhigang Li, Margaret H. Dunham, Yongqiao Xiao: STIFF: A Forecasting Framework for SpatioTemporal Data. Revised Papers from MDM/KDD and PAKDD/KDMCD 2002: 183-198. [10] Rakesh Agrawal Ramakrishnan Srikant, Fast Algorithms for Mining Association Rules, Proceedings of the 20th VLDB Conference Santiago, Chile (1994).
REFERENCES [1] Hipp, J., Guntzer, U., Gholamreza, N. (2000). Algorithm for Association Rule Mining: A General Survey and Comparison, ACM SIGKDD, volume 2 (Issue 1), p. 58. [2] Fayyad, U. M., Shapiro, G. P., Smyth, P., and Uthurusamy, R. (1996). Advances in Knowledge Discovery and Data Mining, Cambridge, AAAI/MIT press [3] Liu, B., Hsu, W., Ma, Y. (1998). Integrating Classification and Association Rule Mining, American Association for Artificial Intelligence
Velammal College of Engineering and Technology, Madurai
Page 45
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
CLD for Improving Overall Throughput in Wireless Networks Dr. P. Seethalakshmi1, Ms. A. Subasri2
1
Professor,Department of Computer Science, Anna University, Trichy, TamilNadu, India. 2 Student, Anna University, Trichy, TamilNadu, India 2
[email protected]
Abstract— In this project a cross-layer design approach to enhance throughput in multi-hop wireless network is proposed. The lossy nature of wireless link decreases the data rate of a given flow and the data flow becomes smaller and smaller along its routing path. As a consequence, the data rate received successfully at the destination node is typically lower than the transmission rate at the source node. The losses are due to interference and fading. These are to be minimized by adjusting the rate of flow and maximum throughput is to be achieved. The different types of interference at each layer are minimized applying various techniques like modulation and coding techniques at each layer. Maximum throughput is obtained by devising throughput maximization algorithms, and Interference aware algorithm.. The Received signal strength, quality of the link, signal to interference noise ratio play an important role in maximize overall throughput in the network. Keywords— Cross-layer design, Throughput , Interference, multi-hop
IX. INTRODUCTION In the past couple of decades, wireless communications have gained dramatic development and have been recently considered as an alternative to wire line networks in providing the last-mile broadband services. Such development further stimulates the emergence of multimedia applications, which require wireless networks to support broader bandwidth, higher transmission rate, and lower endto-end delay. For wireless communications, the challenge to provide multimedia services stems from the hostile wireless channel conditions. Besides channel noise, the time-variant channel fluctuation(i.e., channel fading) severely affects the transmission accuracy and the of quality of service(QoS).In order to combat interference and channel fading various diversity techniques, modulation and coding techniques are used. In the paper[1] the lossy feature of wireless links is studied and a leaky pipe flow model is designed where the flow rate changes per hop, which naturally points to hop-byhop rate control. The effective network utility is determined by considering two constraints, With link outage constraints and with path outage constraints. Hop-by-hop rate
Velammal College of Engineering and Technology, Madurai
control algorithm that are jointly optimized are used. CSIchannel state information is used. In the above work the interference considerations are not analysed. For achieving throughput maximization only estimations are done and analysed. Rayleigh fading model is used. II. RELATED WORKS This scheme [8] is based on belief propagation, which is capable of fully exploiting the statistics of interference. Consider, the detection of a sequence of symbols of the desired user with one strong interferer of the same signalling format, where the fading processes of both the desired user and the interference are Gauss-Markov in nature. Belief propagation is an iterative message passing algorithm for performing statistical inference on graphical models by propagating locally computed beliefs. Belief propagation algorithm has significant performance gain over the traditional interference suppression schemes. A. Local Balancing And Interference Aware Channel Allocation: This algorithm is used for reducing the overall interference in the network. Several approaches have been proposed for minimizing the Adjacent channel effects[5] ranging from coordinating the multiple radios in the wireless node and adjusting antenna parameters and the filter characteristics to using the channel overlaps for routing data across devices operating on the non-overlapping channels. One of the popular approaches for mitigating the interference effects[4] is to choose the transmission channels carefully by making sure that nearby links are on channels that do not interfere sufficiently. However, due to the dynamic nature of the links in a wireless network, the interference characteristics may vary, and therefore the channel allocation should be adaptable to these variations.
Page 46
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
B. Analog Network Coding Wireless networks strive to avoid scheduling multiple transmissions at the same time in order to prevent interference. This paper[6] encourages strategically picked senders to interfere instead of forwarding packets, routers forward the interfering signals. The destination leverages network level information to cancel the interference and recover the signal destined to it, The result is analog network coding because , signals are sent and not bits. C. Interference Aware Load Balancing. [4]Path weight function and routing schemes to provide interference aware and multichannel aware load balancing for mesh networks is done. The objective of load balancing is essentially to distribute traffic among different paths to avoid creating congested areas and improve network performance. D. Co- Operative Diversity To determine the optimal trade-off between the amount of throughput gain obtained via co-operation and the amount of interference introduced to the network. [2]Co-operative regionfor each active node is maintained, The nodes lying in such a region are allowed to co-operate with the source. They adopt decode and forward scheme at the relays and use the physical interference model to determine the probability that a relay node correctly decodes its corresponding source. Limitations Of Existing System All the techniques have their own limitations according to the design they have arrived. Tradeoffs between certain parameters is made. Link quality varies drastically. III. DESIGN AND IMPLEMENTATION In this project overall network throughput is to be enhanced by considering all the performance metrics and analyzing the parameters at each layer. The cross layer design is proposed for interactions between different layers. Using cross layer design the layers to be considered are Physical layer, MAC/ Data link layer.
S1
D1
Fig. 1 One source and one destination node
Velammal College of Engineering and Technology, Madurai
. In the above scenario there is no interference as there is only one source and one destination node.
S
S
D
D
S
D
Fig. 2 Nodes with two interfering sources In this fig. 2 there are two interfering sources which causes an adverse effect by introducing noise . As distance between source and destination node increases the received signal strength at the destination node decreases and interference increases which leads to the degradation in the quality of the signal. Normally by setting a threshold value for the received signal strength the packets are dropped or received. Here in this project an additional information is included in terms of interference which is to decide which packets are to be dropped and which are to be passed based on the signal to interference noise ratio calculation. A. Architectural Design: For each packet, a packet error rate(PER) can be computed as a function of received power, interference from concurrent transmissions. Here interference is of high importance. The underlying modulation is BPSK(Binary Phase Shift Keying) followed by a variable rate channel code. For each packet received a signal to interference and noise ratio is calculated. The bit error rate corresponding to this SINR is obtained by lookup tables. Gaussian approximation is used for the multiuser interference and for computing the SINR. From the computed BER the PER is obtained by standard approximation. The PER is used as the parameter of a binary random variable used to decide whether the packet is properly received or not.
Page 47
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The various interferences[3] are to be analysed at each layer and their effects in increasing or reducing the throughput is taken into account and the various qos metrics like throughput, delay and jitter have to be improved by changing the parameters that are in each layer which account for various degradation. The layers to be analysed are Physical layer – Received signal strength, bit error rate, Interference or signal to Interference noise ratio, modulation techniques, Data link layer- addressing or frame checking, reliability of link, Network layer- path reliability, routing, hop count, Transport –packet delivery ratio. A rate adaptation scheme is to be proposed. B. Implementation The impact of interference on the link from traffic on neighbouring links is determined. Quality of link is measured in terms of packet delivery ratio/loss rate. In the absence of interference the link capacity would be the product of the maximum sending rate of the sender and delivery ratio of the link. QoS - metrics in networks to be dealt with includes Available bandwidth, Packet loss rate/bit error rate, delay. Current routing protocols are AODV and DSR. They select the shortest path between source and Destination. The issue in shortest path selection is that intermediate nodes having less battery power, more delay, less bandwidth, high congestion, more noise ratio in such cases the path is not optimal for long and bandwidth oriented transmissions such as multimedia or real time applications, as such applications need more bandwidth, less delay and more link life. An adaptive and optimal route can only be designed using cross layer approach in which the source and destination select route on the basis of many parameters from different layers.
Fig. 3 End-to-end- Latency
The fig. 4 shows the computed packet delivery ratio with respect to the number of nodes active in the network . As packet delivery ratio increases the overall network throughput increases.
Fig. 4 Packet Delivery Ratio Vs Time
IV RESULTS As distance between the source and destination increases the delay increases due to interference. As delay increases throughput decreases. As packet delivery ratio increases delay decreases or end to end latency decreases. The fig. 3 shows the end to end latency with respect to time .
The fig. 5 represents the reporting node which are active and listening to the transmission and the number of packetsReceived reducing the interference and latency.
Velammal College of Engineering and Technology, Madurai
Page 48
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[8] Yan Zhu, Dongning Guo and Michael L. Honig , “Joint Channel Estimation and Co-Channel Interference Mitigation in Wireless Networks Using Belief Propagation”, Department of Electrical Engineering and Computer Science Northwestern University, Evanston, Illinois 60208. [9] http://www.wirelesscommunication.nl/reference/chaptr04 /outage/compouta.htm [10] http://www.wirelesscommunication.nl/reference/chaptr05/ spreadsp /ber.htm [11] htttp://www.isi.edu/nsnam/ns/tutorial/index.html
Fig. 5 No. Of Packets Received Vs Reporting Node
CONCLUSION This project aims at reducing the different types of interferences at each layer and improving the overall throughput by introducing throughput maximization algorithm and interference aware algorithm along with rate adaptation. NS2 simulation environment is implemented to test the results. By introducing the interference aware algorithm the throughput has been improved to an increase of 3% when compared to the standard existing system.
REFERENCES [1] Qinghai Gao, Junshan Zhang and Stephen V. Hanly, “Cross Layer rate control in wireless networks with lossy links: leaky pipeflow, effective network utility maximization and hop by hop algorithms”, IEEE Transactions of wireless communications vol8,no. 6,june 2000. [2] Kaveh Pahlavan,Prashant Krishnamurthy, “Principles of Wireless Networks”, Prentice Hall of India Private Limited, 2006. [3] Jochen Schiller, “Mobile Communication” 2nd Edition, Pearson Education 2003. [4] Yaling Yang, Jun Wang and Robin Kravets, “ Interference-aware Load Balancing for Multi-hopWireless Networks”, University of Illinois at Urbana- Champaign. [5] Nitin H. Vaidya , Vijay Raman , “Adjacent Channel Interference Reduction in Multichannel Wireless NetworksUsing Intelligent Channel Allocation”, Technical Report (August 2009),University of Illinois at Urbana Champaign. [6] Sachin Katti ,Shyamnath Gollakota, Dina Katabi, ”EmbracingWireless Interference: Analog Network Coding”, mit.edu. [7] David Tse, Pramod Viswanath , “Fundamentals of Wireless Communications “, University of California, Berkeley, August13, 2004.
Velammal College of Engineering and Technology, Madurai
Page 49
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Particle Swarm Optimization Algorithm In Grid Computing Mrs.R.Aghila1, Assistant Professor, [email protected]
Mr.R.Arun Kumar2 Lecturer [email protected]
M.Harine3, G.Priyadharshini3 Final year student, [email protected]
Department of Computer Science and Engineering, K.L.N College of Information Technology, Pottapalayam, Sivagangai district Abstract—Computational grid is a hardware and software infrastructure that provides dependable, consistent, pervasive and inexpensive access to high end computational capabilities. Grid computing environment has coordinated resources, Open Standard Protocols & frameworks, QOS (Quality of Service). Grid is mainly to extract enormous power of network, heterogeneous system. Grid can be applied in biomedicine, earth science, a high energy physics, astronomy, weather forecasting etc... The main need of grid computing application is discovering and scheduling of tasks and workflows. Scheduling plays an important role in load balancing and thus avoids processing delays and over commitment of resources. An effective task scheduling algorithm is necessary to satisfy the above need. In this paper, we’ve implemented Particle Swarm Optimization (PSO) algorithm for effective task scheduling using gridsim tool in Java. We’ve also compared its performance with Genetic Algorithm (GA). We aim to generate minimum completion time while executing the tasks using PSO than that obtained using GA. Keywords— Genetic algorithm, Simulated annealing, Particle Optimization algorithm
I. A.
INTRODUCTION
GRID COMPUTING:
Grid [1] is defined as an infrastructure that couples computers, software, databases, special instruments and people across the internet and presents them as an unified integrated resource. It is mainly used to extract enormous power of network, heterogeneous system. Grids have emerged as a global cyber-infrastructure for the nextgeneration of e-Science and e-business applications, by integrating large-scale, distributed and heterogeneous resources. Scientific communities in areas such as highenergy physics, gravitational-wave physics, geophysics, astronomy and bioinformatics, are utilizing Grids to share, manage and process large data sets.
Velammal College of Engineering and Technology, Madurai
B.
TASK SCHEDULING:
The need for task scheduling [1-2] arises from the requirement for most modern systems to perform multitasking (executing more than one process at a time) and multiplexing (transmit multiple flows simultaneously). Task scheduling is one of the NP-complete problems [5]. Heuristic optimization algorithm is widely used to solve a variety of NP-Complete problems. With the development of the network technology, grid computing used to solve larger scale complex problems becomes a focus technology. Task scheduling is a challenging problem in grid computing environment. If large numbers of tasks are computed on the geographically distributed resources, a reasonable scheduling algorithm must be adopted in order to get the minimum completion time. So task scheduling which is one of NPComplete problems become a focus by many of scholars in grid computing area. Abraham et al and Braun et al [6] presented three basic heuristics implied by Nature for Grid scheduling, namely Genetic Algorithm (GA) [3-5], Simulated Annealing (SA) [5] and Tabu Search (TS) [11], and heuristics derived by a combination of these three algorithms. GA and SA are powerful stochastic optimization methods, which are inspired form the nature.PSO [7-10] shares many similarities with evolutionary computational techniques such as Genetic Algorithms. The system is initialized with a population of random solutions and searches for optima by updating generations. However unlike GA, PSO has no evolution operators such as cross over and mutation. In PSO, the potential solutions, called particles, fly through the problem space by following the current optimum particles. This paper is organized as follows. In section 2, the mostly used task scheduling algorithms in grid computing. In section 3, our proposed system Particle Swarm Optimization algorithm is explained. Its implementation is also discussed in this section. The section 4 and 5 deals with concluding discussions and references respectively.
Page 50
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
II.
MOSTLY USED TASK SCHEDULING ALGORITHMS IN GRID COMPUTING:
In this paper, we will discuss the problem that ‘numt’ tasks work on ‘res’ computing resources with an objective of minimizing the completion time and utilizing the resources effectively. If the number of tasks is less than the number of resources in grid environment, the tasks can be allocated on the resources according to the first- come-first-serve rule. If the number of task is more than the number of resources, the allocation of tasks is to be made by some scheduling schemes. A.
SIMULATED ANNEALING (SA):
Simulated Annealing (SA) derives from the Monte Carlo method for statistically searching the global. The concept is originally from the way in which crystalline structures can be formed into a more ordered state by use of the annealing process, which repeats the heating and slowly cooling a structure [5]. SA has been used to select a suitable size of a set of machines for scheduling in a Grid environment.
Genetic algorithms (GAs) provide robust search techniques that allow a high-quality solution to be derived from a large search space in polynomial time by applying the principle of evolution [5]. A genetic algorithm combines exploitation of best solutions from past searches with the exploration of new regions of the solution space. Any solution in the search space of the problem is represented by an individual. A genetic algorithm maintains a population of individuals that evolves over generations. The quality of an individual in the population is determined by a fitness function. The fitness value indicates how good the individual is compared to others in the population. It first creates an initial population consisting of randomly generated solutions. After applying genetic operators, namely selection, crossover and mutation, one after the other, new offspring are generated. Then the evaluation of the fitness of each individual in the population is conducted. The fittest individuals are selected to be carried over next generation. The above steps are repeated until the termination condition is satisfied. Typically, a GA is terminated after a certain number of iterations, or if a certain level of fitness value has been reached.
Figure 2. Flowchart – Genetic Algorithm
III. A. PARTICLE ALGORITHM: Figure 1. Flowchart – Simulated Annealing algorithm
B.
GENETIC ALGORITHMS (GA):
Velammal College of Engineering and Technology, Madurai
OUR PROPOSED WORK: SWARM
OPTIMIZATION
PSO [7] is a robust stochastic optimization technique based on the movement and intelligence of swarms. PSO applies
Page 51
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the concept of social interaction to problem solving. It was developed in 1995 by James Kennedy (social-psychologist) and Russell Eberhart (electrical engineer). In PSO each single solution is a ‘bird’ in the search space. We call it “particle”. All of particles have fitness values which are evaluated by the fitness function to be optimized and have velocities which direct the flying of particles. Each particle keeps track of its coordinates in the solution space which are associated with the best solution (fitness) that has achieved so far by that particle. This value is called personal best, pbest. Another best value that is tracked by the PSO is the best value obtained so far by any particle in the neighborhood of that particle. This value is called gbest. The basic concept of PSO lies in accelerating each particle toward its pbest and the gbest locations, with a random weighted acceleration at each time
w: weighting function, cj : weighting factor, rand : uniformly distributed random number between 0 and 1, sik : current position of agent i at iteration k, pbesti : pbest of agent i, gbest: gbest of the group. The following weighting function is usually utilized in (1) w=wMax-[(wMax-wMin)xiter]/maxIter (2) Where wMax= initial weight, wMin = final weight, maxIter = maximum iteration number, iter = current iteration number. sik+1 = sik + Vik+1
(3)
3.2 FLOW CHART DEPICTING THE GENERAL PSO ALGORITHM:
Figure 3.Basic concept of PSO
sk : current searching point. sk+1: modified searching point. vk: current velocity. vk+1: modified velocity. vpbest : velocity based on pbest, vgbest : velocity based on gbest Each particle tries to modify its position using the following information: ) the current positions, ) the current velocities, ) the distance between the current position and pbest, ) The distance between the current position and the gbest.
Figure 4. PSO-Flowchart
The modification of the particle’s position can be mathematically modeled according the following equation:
B.
Vik+1 = wVik +c1 rand1 (…) x (pbesti-sik) + c2 rand2 (…) x (gbest-sik) ….. (1)
Unlike in genetic algorithms, evolutionary programming and evolutionary strategies, in PSO, there is no selection operation. All particles in PSO are kept as members of the population through the course of the run PSO [8] are the only
where, vik : velocity of agent i at iteration k,
Velammal College of Engineering and Technology, Madurai
ADVANTAGES OF PSO OVER GA AND SA:
Page 52
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
algorithm that does not implement the survival of the fittest. No crossover operation in PSO. C.
IMPLEMENTATION:
Module 1: PSO Implementation The number of processors i.e., the number of resources and the number of jobs are obtained as input. From the obtained input, gridlets are generated. The PSO algorithm is used to allocate the jobs to correct resources so that the completion time of the process will get minimized. Module 2: Comparison with GA The implementation is made to run for certain trials and the output is obtained. It is then compared with that of the results obtained using Genetic Algorithm and the graph is drawn.
Figure 6. Example implementation – Generation of completion time
Consider the number of resources is 3 and the number of tasks is 10. The speed of three resources are 4,3 and 2 respectively. The results of GA algorithm running 10 times were {26, 25.4, 25.8, 25.8, 25, 25, 25.8, 26, 25.4, 25}, with an average value of 25.52.the best result based on PSO algorithm for (3, 10) is shown in table 1. Table 1. Example of the best result
Figure 5.Example implementation-Getting the user inputs
Task(T)/Resources(R) T1 T2 T3 T4 T5 T6 T7 T8 T9 T10
R1 0 1 1 0 0 1 0 1 0 0
R2 1 0 0 0 1 0 0 0 1 0
R3 1 0 0 1 0 0 1 0 0 1
The results of PSO algorithm running 10 times were {25, 25, 25.4, 25.4, 25, 25, 25, 25, 25, 25.2}, with an average value of 25.1. PSO algorithm provided the best result 7 times, while GA algorithm provided the best result 3 times.
Velammal College of Engineering and Technology, Madurai
Page 53
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
proposed approach and compared it with genetic algorithm under the same condition. From the simulated experiment, the result of PSO algorithm is better than GA. Simulation results demonstrate that PSO algorithm can get better effect for a large scale optimization problem. Task scheduling algorithm based on PSO algorithm can be applied in the computational grid environment. REFERENCES:
Figure 7: Performance of PSO and GA scheduling algorithm about 5 processors and 100 tasks
Figure 8: The performance of curves of different numbers of processors running different Number of tasks
It shows PSO usually spent the shorter time to complete the scheduling than GA algorithm. It is to be noted that PSO usually spent the shorter time to accomplish the various task scheduling tasks and had the better result compared with GA algorithm. IV.
[9] Foster and C. Kesselman (editors), The Grid: Blueprint for a Future Computing Infrastructure,Morgan Kaufman Publishers, USA, 1999. [10] Y. Gao, H.Q Rong and J.Z. Huang, Adaptive grid job scheduling with genetic algorithms, Future Generation Computer Systems, pp.1510-161 Elsevier,21(2005). [11] M. Aggarwal, R.D. Kent and A. Ngom, Genetic Algorithm Based Scheduler for Computational Grids, in Proc. of the 19th Annual International Symposium on High Performance Computing Systems and Application (HPCS’05), ,pp.209-215 Guelph, Ontario Canada, May 2005. [12] S. Song, Y. Kwok, and K. Hwang, Security-Driven Heuristics and A Fast Genetic Algorithm for Trusted Grid Job Scheduling, in Proc. of 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05), pp.65-74, Denver, Colorado USA, April 2005. [13] Workflow scheduling algorithm for grid computing Jia Yu and Rajkumar Buyya Grid computing and distributed system(GRIDS) laboratory Department of Computer Science and Software Engineering The University of Melbourne,Australia. [14] Abraham, R. Buyya and B. Nath, Nature's Heuristics for Scheduling Jobs on Computational Grids, The 8th IEEE International Conference on Advanced Computing and Communications (ADCOM 2000), pp. 4552,Cochin, India, December 2000,. [15] Kennedy J. and Eberhart R. Swarm interllignece,Morgan Kaufmann, 2001. [16] J. Kennedy and R. C. Eberhard, “Particle swarm optimization”, Proc. of IEEE Int’l Conf. on Neural Networks, pp.1942-1948, Piscataway, NJ, USA, ,1995. [17] J.F. Schute and A.A. Groenwold, A study of global optimization using particle swarms, Journal of Global Optimization, pp.93-108, Kluwer Academic Publisher,31(2005). [18] M. Fatih Tasgetiren, Yun-Chia Liang, MehmetSevkli, and Gunes Gencyilmaz, “Particle Swarm Optimization and Differential Evolution for Single Machine Total Weighted Tardiness Problem,” International Journal of Production Research, pp. 4737-4754 , vol. 44, no. 22, 2006. [19] R. Braun, H. Siegel, N. Beck, L. Boloni, M. Maheswaran, A. Reuther, J. Robertson, M. Theys, B. Yao, D. Hensgen and R. Freund, A Comparison of Eleven Static Heuristics for Mapping a Class of Independent Tasks onto Heterogeneous Distributed Computing Systems, pp. 810-837, J. of Parallel andDistributed Computing, vol.61, No. 6, 2001.
CONCLUSION:
In this paper, scheduling algorithm based on PSO is proposed for task scheduling problem on computational grids. Each particle represents a feasible solution. The position vector is transformed from the continuous values to the discrete values based on SPV rules, accordingly, a permutation formed. Our approach is to generate an optimal schedule so as to complete the tasks in a minimum time as well as utilizing the resources in an efficient way. We evaluate the performance of our
Velammal College of Engineering and Technology, Madurai
Page 54
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
NTRU - Public Key Cryptosystem For Constrained Memory Devices V.Pushparani#1 and Kannan Balasubramaniam*2
#1
Faculty of Kamaraj College of Engg and Technology, Virudhunagar, Tamilnadu. 1
[email protected]
*2
Professor of Mepco Schlenk Engg College, Sivakasi. Tamilnadu. 2
[email protected]
ABSTRACT -- In many business sectors secure and efficient data transfer is essential. To ensure the security to the applications of business, the business sectors use Public Key cryptographic Systems (PKCS).. Secure public key authentication and digital signatures are increasingly important for electronic communications, commerce and security. They are required not only on high powered desktop computers, but also on smart cards and wireless devices with severely constrained memory and processing capabilities. An RSA and NTRU system generally belongs to the category of PKCS. The efficiency of a public key cryptographic system is mainly measured in computational overheads, key size. Although the security of RSA is beyond doubt, the evolution in computing power has caused a growth in the necessary key length. The fact that most chips on smart cards cannot process keys extending 1024 bit shows that there is a need for alternative. NTRU is such an alternative and it is a collection of mathematical algorithms based on manipulating lists of very small integers and polynomials. NTRU is the secure public key cryptosystem not based on factorization or discrete logarithm problems. The security of the NTRU cryptosystem came from the interaction of the polynomial mixing system with the independence of reduction modulo two relatively prime integers’ p and q. KeyWords -- NTRU, patterns, encryption, decryption, digital signature
I. INTRODUCTION A. NTRU NTRU (Number Theory Research Unit) is relatively new and was conceived by Jeffrey Hoff stein, Jill Pipher and Joseph. H. Silverman. NTRU uses polynomial algebra combined with clustering principle based on elementary mathematical theory. The security of NTRU comes from the interaction of polynomial mixing system with the independence of reduction modulo two relatively prime numbers. The basic collection of objects used by the NTRU Public Key Cryptosystem in the ring R that consists of all truncated polynomials of degree N-1 having integer coefficients a =a 2 + a 3X 3 +… + a N-2X N-2 + a N-1X N-1. 0+ a 1X + a 2X Polynomials are added in the usual way. They are also
Velammal College of Engineering and Technology, Madurai
multiplied more or less as usual, except that XN is replaced by 1, XN+1 is replaced by X, X N+2 is replaced by X2 and so on. 1)
Truncated Polynomial Rings:
The NTRU public-key algorithm, explained in chapter II, uses random polynomials which are generated from a [X]/(XN-1). The polynomial ring of the form R[X] = polynomials that form the ring R[X] have a degree smaller than N. The polynomials in the truncated ring R[X] are added in a regular way by adding their coefficients. The polynomial XN ≡ 1. Said differently, the maximum degree of the resultant polynomial of a multiplication between two polynomials of the ring cannot be greater than N - 1. The product operation of two polynomials in R[X], is defined as c(X) = a(X) * b(X) where ck is the kth coefficient of c(X) and is computed as ck=a0bk + aibk-1 + akb0 +.... + aN-1bk+1 The product of polynomials in R[X] is also called the star multiplication. II. RELATED WORKS A. NTRU: A Ring Based Public Key Cryptosystem - J. Hoffstein, J. Pipher, and J. H. Silverman N −1
The polynomial is represented as a(X)=
∑
aiXi € R. When
i =0
NTRU was formally introduced in 1998, Silverman presented the polynomial multiplication as the cyclic convolution of two polynomials as ck =
k
∑
i= 0
a i .b k − i +
N −1
∑
a i .b N
i= k +1
+ k −i
=
∑
i + j = k (mod
a i .b
j
N )
Ultimately, this straightforward method requires N2 multiplications to perform a polynomial multiplication for NTRU. B. NTRU in Constrained Devices – D. V. Bailey, D. Coffin, A. Elbirt, J. H. Silverman, and A. D. Woodbury
Page 55
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
In 2001, Bailey et al. introduced a fast convolution algorithm, to perform a polynomial multiplication for NTRU. This algorithm makes the realization that almost every polynomial multiplication involved with NTRU has one polynomial that is random. The random polynomials are assumed to have binary coefficients. In addition, the random polynomial is assumed to consist of three smaller polynomials of low Hamming weight(d), The product can be computed using only additions over and rotations of the coefficient vector of a. For each non-zero coefficient bi of b, polynomials of the form Xia are added in order to compute the product ab. The multiplication of a polynomial a with a monomial Xi in R corresponds to i right rotations of the coefficient vector of a, where the right rotation is defined as the mapping
( a 0 ,..., a N −1 ) a ( a N −1 , a 0 ,..., a N − 2 )
Fig.1 Multiplications of a, b using additions and rotations
The fast convolution algorithm reduces the complexity of NTRU's polynomial multiplication to dN additions and no multiplications. C. Sliding Window Method for NTRU - M.-K. Lee1, J. W. Kim, J. E. Song, and K. Park. In 2007, Lee et al observed that it is possible to reduce the number of additions needed to compute the product of polynomial a, b by using bit patterns of the binary polynomial b. By a bit pattern, we understand two 1s separated by a (possibly empty) sequence of 0s. We say that such a bit pattern has length l if the two 1s are separated by l - 1 0s. More generally, it is possible to reduce the number of additions needed to compute the product ab whenever a bit (a0 ,..., aN−1more ) a (than aN−1 ,once a0 ,..., pattern occurs inab. N −2It) is thus desirable to choose bit patterns in a way that maximizes the number of pattern occurrences and to efficiently identify the patterns in b. The algorithm of Lee et al. only considers bit patterns of length less than or equal to a parameter w. For each pattern l length l = 1…..w, the polynomial a + X a is precomputed and stored in a lookup table. The non-zero coefficients not belonging to any such bit pattern are treated as in the algorithm of Bailey et al. Binary polynomials are represented as bit strings.
Velammal College of Engineering and Technology, Madurai
Fig.2 Multiplications of a, b using bit patterns
Considering bit strings containing more than two 1s does not achieve any notable speedup because the probability that these strings occur more than once in b is very low. D. An Algorithm of Dynamic Patterns for NTRU - Bu Shan Yue Zhang Han Yan Wang RuChuan Bu Shan Yue et al proposed algorithm uses bit patterns, but the patterns can be of arbitrary length, and only the patterns actually occurring in b are considered. Thus, all non-zero coefficients of b belong to a pattern, except for a single coefficient in case that the hamming weight of b is odd. They omit the precomputation step of the algorithm of Lee et al. and instead compute the polynomials a+Xla when needed. They also represent binary polynomials as the sequence of the degrees of their monomials. It shows that pattern finding can be performed much easier and faster in this representation. E. NSS: The NTRU Signature Scheme - Jeffrey Hoffstein, Jill Pipher, Joseph H. Silverman Secure public key authentication and digital signatures are increasingly important for electronic communications and commerce, and they are required not only on high powered desktop computers, but also on Smart Cards and wireless devices with severely constrained memory and processing capabilities. Silverman et al introduced a NTRU signature scheme in which the public and private keys are formed as follows. Choosing two polynomials f and g having the form f = f0 + pf1 and g = g0 + pg1. Here f0 and g0 are fixed universal polynomials (e.g., f0 = 1 and g0 = 1-X) and f1 and g1 are polynomials with small coefficients chosen from the sets Ff and Fg, respectively. Next computes the inverse f-1 of f modulo q, that is, f-1 satisfies f-1* f ≡ 1 (mod q). Public verification key is the polynomial h * f-1 ≡g (mod q). Private signing key is the pair (f; g). III. NTRU PUBLIC KEY CRYPTOSYSTEM A. Parameters The basic collection of objects used by the NTRU Public Key Cryptosystem in the ring R that consists of all truncated polynomials of degree N-1 having integer coefficients
Page 56
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
a =a 0+ a 1X + a 2X 2 + a 3X 3 +…...+a N-2X N-2 +a N-1X N-1. All computations are performed in the ring of convolution modular polynomials R = Z [ X ] / ( X N − 1 ) where polynomials of degree less than N are used as representatives for the residue classes. Polynomials are added in the usual way. They are also multiplied more or less as usual, except that XN is replaced by 1, XN+1 is replaced by X, X N+2 is replaced by X2 and so on. The NTRU public-key algorithm is defined by the following parameters: • N: The degree parameter. Defines the degree N - 1 of the polynomials in R. • q: Large modulo. Polynomial coefficients are reduced modulo q. • p: Small modulo. The coefficients of the message are reduced modulo p in decryption. • dF: Private key space. Fixes the polynomial form defining the number of positive ones for the private key f, the negative ones are fixed by df - 1. • dg: Public key space. Fixes the polynomial form defining the number of positive and negative ones for the random polynomial g used to calculate the public key. • dr: Blinding value space. Fixes the polynomial form defining the number of positive and negative ones of the random polynomial r used in the encryption process. • D(d): The set of binary polynomials of degree less than N with hamming weight d. B. Key generation 1) Pattern Finding: A binary polynomial b of hamming weight d is represented by the sequence D0....Dd-1 of the degrees of its monomials in ascending order. The polynomial is traversed once in reverse order, starting at Dd-1. For each possible pattern length l € 1...N - d+1 a list Ll of pattern locations is created. Every pair of degrees (Di....Di-1) represents a bit pattern of length Di – Di1. The degree Di is stored in the list LDi....Di-1 and i is decreased by 2. In case that d is odd, the remaining single degree D0 is stored separately in a list L0. 2) Pattern multiplication: Each non-empty list Ll, , l > 0 represents a bit pattern of b with length l. For each such Ll, the corresponding pattern polynomial p = a + X l a is computed. For each element D of the list Ll, this pattern polynomial is right rotated D times and added to the resulting polynomial. A possibly remaining single degree stored in L0 is treated separately without computing a pattern polynomial.
During Key Generation the process Random Polynomial is invoked to generate the polynomial f and g in order to calculate the public key h. The polynomials are generated with random coefficients from a truncated ring of polynomials R. Random Polynomial receives the number of positive and negative ones and generates the random polynomial of N coefficients. 4) Inversion modulo q: During Key Generation the polynomial fq, the inverse of f modulo q, is also computed. The polynomial fq is necessary, together with g, to calculate the public key h. Inversion modulo q computes the inverse of a certain polynomial f modulo q in fq, computing fq = f-1 (mod q) which satisfies f* fq ≡ 1 (mod q). 5) Inversion modulo p: Inversion modulo p computes the inverse of a polynomial f in modulo p such that f * fp ≡1(mod p). 6) Key generation: Choose uniformly
at
random
and polynomials f . f −1 ≡ 1(mod q ) pF. If the congruence
the
binary
Compute f = 1 + has a solution,
calculate such a solution Otherwise, start over. = f − 1 pg mod q Compute the hpolynomial . The notation a = b mod q stands for reducing the coefficients of b modulo q and assigning the result to a. The private key is f, the public key is h.
Fig. 3 Key Generation
3) Random Polynomial generation:
Velammal College of Engineering and Technology, Madurai
Page 57
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
C. Encryption Encryption is the simplest part in the NTRU PKC. Encryption only requires generating a random polynomial r from the ring R that obscures the message. Then the polynomial r is multiplied by the public key h. And finally the product of r and h is added to the desired message to encrypt. This means Encryption just needs to receive a message in the polynomial form m and the public key h.
Dev(a; b) = #{i: Ai ≠ Bi}: Intuitively, Dev(a; b) is the number of coefficients of a mod q and b mod q that differ modulo p. 1) Key Generation: Chooses two polynomials f and g. Compute the inverse f-1 of f modulo q. Public verification key is the polynomial h * f-1 ≡g (mod q) and his private signing key is the pair (f; g). 2) Signing: User’s document is a polynomial m modulo p. Chooses a polynomial w € Fw of the form w = m + w1 + pw2; where w1 and w2 are small polynomials. Then computes s≡ f* w (mod q): Signed message is the pair (m; s).
Fig.4 Encryption specification D. Decryption
The Decryption process requires the encrypted message e and the private key set (f.. fp) to decrypt the encrypted message e into the clear message c.
3) Verification: In order to verify signature s on the message m, First checks that s≠0 and then verifies the following two conditions: 1. Compares s to f0*m by checking if their deviation satisfies Dmin≤ Dev(s. f0 * m)≤Dmax 2. Use public verification key h to compute the polynomial t≡h*s (mod q), putting the coefficients of t into the range [-q=2; q=2] as usual. Then checks if the deviation of t from g0*m satisfies Dmin≤ Dev(t. g0 * m)≤Dmax If signature passes tests (A) and (B), then accepts it as valid. IV. PERFORMANCE ANALYSIS In order to grasp how well NTRU performs for different applications, a timing analysis was conducted for the Key generation, Encryption, and Decryption functions. The test values for the parameters of NTRU used for this performance analysis are listed in Table ITABLE I TEST VALUES USED for PERFORMANCE ANALYSIS
Parameters
Fig.5 Decryption specification
E. The NTRU Signature Scheme The key computation involves the deviation between two polynomials. Let a(X) and b(X) be two polynomials in R. We first reduce their coefficients modulo q to lie in the range between -q=2 to q=2, then we reduce their coefficients modulo p to lie in the range between -p=2 and p=2. Let A(X) = A0+A1X+.....+AN-1XN-1 and B(X) = B0+.......+BN-1X N-1 be these reduced polynomials. Then the deviation of a and b is
Velammal College of Engineering and Technology, Madurai
N q p NumOnes f NumNegOnes f NumOnes g NumNegOnes g NumOnes r NumNegOnes r NumOnes m NumNegOnes m
107 NTRU 107 64 3 15 14 12 12 5 5 25 25
503 NTRU 503 256 3 216 215 72 72 55 55 165 165
Page 58
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The results of the timing analysis are shown in Table II. TABLE II TIMINGS for NTRU
Functions in NTRU Key Generation (ms) Encryption (ms) Decryption (ms)
107 NTRU
503 NTRU
16.2
699.5
0.6
15.0
1.4
29.4
Systems (CHES ’01), volume 2162 of Lecture Notes in Computer Science, pages 262–272. Springer Verlag. [3]. J. Hoffstein, J. Pipher, and J. H. Silverman. NTRU: A Ring-Based Public Key Cryptosystem. 1998, In Proceedings of the Third International Symposium on Algorithmic Number Theory, volume 1423 of Lecture Notes in Computer Science, pages 267–288.Springer Verlag. [4]. Jeffrey Hoffstein, Jill Pipher, Joseph H. Silverman,2005,NSS: The NTRU Signature Scheme. [5]. R. Lindner, J. Buchmann, M. Doering, 2008, Efficiency Improvements for NTRU, Sicherheit 2008, LNI vol. 128. [6]. M.-K. Lee1, J. W. Kim, J. E. Song, and K. Park. 2007, Sliding Window Method for NTRU. In Proceedings of ACNS 2007, volume 4521 of Lecture Notes in Computer Science, pages 432–442. Springer Verlag. [7]. Narasimham Challa and Jayaram Pradhan, 2008, Performance Analysis and Public key cryptographic systems RSA and NTRU, IJCSNS International Journal of computer science and Network security.
The Key Generation function takes the longest time because it requires two polynomial inversions and a polynomial multiplication. In addition, since the Decryption function requires two polynomial multiplications, it takes over two times as long as the Encryption functions. Since Encryption requires only one polynomial multiplication, it is fair to use Encryption's timing to estimate the time to perform a single polynomial multiplication. Altogether, the timing analysis in Table II shows that NTRU has potential in offering high performance. V. CONCLUSION The NTRU is suitable for applications where it requires security based on the environment. Because of the RSA’s time complexity O (n 3) and that of NTRU’s O (n log (n)), NTRU confirms its cryptography and delivers encryption, decryption and authentication at speeds of multiple times faster than RSA. NTRU is ideally suited for applications where high performance, high security and low power consumption are required. ACKNOWLEDGMENT The authors are grateful to the management of Kamaraj College of Engineering and Technology, Virudhunagar, India and Mepco Schlenk Engineering College, Sivakasi, India for granting permission to undertake this work. Our thanks are due to the Head of the Department of Computer Science and Engineering of Kamaraj College of Engineering and Technology for allowing us the use of the laboratories and computing facilities. REFERENCES [1]. Bu Shan Yue Zhang Han Yan Wang RuChuan Dept. of Comput. Eng., Huaiyin Inst. of Technol., Huaiyin; 2008, An Algorithm of Dynamic Patterns for NTRU Wireless communications, Networking and Mobile computing,. WiCOM ’08. [2]. D. V. Bailey, D. Coffin, A. Elbirt, J. H. Silverman, and A. D. Woodbury. 2001, NTRU in Constrained Devices. In Proceedings of the Third International Workshop on Cryptographic Hardware and Embedded
Velammal College of Engineering and Technology, Madurai
Page 59
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Novel Randomized Key Multimedia Encryption Algorithm Security Against Several Attacks S. Arul Jothi Department of Computer Science, NMS Sermathai Vasan College for Women Madurai, Tamilnadu, India [email protected] Abstract— Encryption is the science of changing data so that it is unrecognisable and useless to an unauthorised person. Decryption is changing it back to its original form. The wide availability of digital multimedia contents as well as the accelerated growth of wired and wireless communication technologies have brought the multimedia content security issue to the forefront. In particular, the problem of efficient multimedia data encryption has recently gained more attention in both academia and industry. Unlike ordinary computer applications, multimedia applications generate large amounts of data that has to be processed in real time. So, a number of encryption schemes for multimedia applications have been proposed in recent years. The most secure techniques use a mathematical algorithm and a variable value known as a 'key'. As the entire operation is dependent upon the security of the keys, it is sometimes appropriate to devise a fairly complex mechanism to manage them. With the rapid progress of information technology, security becomes one of the key factors in information storage, communication and processing. For the reason of speed and security I develop a novel encryption algorithm secure to several attacks. As opposed to traditional key algorithms here the key is randomized. Keywords— Cryptanalysis, Algorithm, Cryptography, Block cipher, Security, Encryption and Decryption.
I. INTRODUCTION Information security deals with several different "trust" aspects of information. Another common term is information assurance. Information security is not confined to computer systems, nor to information in an electronic or machine-readable form. It applies to all aspects of safeguarding or protecting information or data, in whatever form. Cryptography [3] [10]can also be defined as the science and art of manipulating message to make them secure. In this the original message to be transformed is called the plaintext and resulting message after transformation is called the cipher text. There are several ways of classifying cryptographic algorithms. They will be categorized based on the number of keys that are employed for encryption and decryption, and further defined by their application and use. The three types of algorithms that will be discussed are
Velammal College of Engineering and Technology, Madurai
secret key cryptography, public key cryptography and hash functions. With secret key cryptography[1] [7], a single key is used for both encryption and decryption. As shown in figure below, the sender uses the key to encrypt the plaintext and sends the cipher text to the receiver. The receiver applies the same key to decrypt the message and recover the plaintext. Because a single key is used for both functions, secret key cryptography is also called symmetric encryption Public-key cryptography is based on the notion that encryption keys are related pairs, private and public. The private key remains concealed by the key owner; the public key is freely disseminated to various partners. Data encrypted using the public key can be decrypted only by using the associated private key and vice versa. Because the key used to encrypt plaintext is different from the key used to decrypt the corresponding cipher text, public-key cryptography is also known as asymmetric cryptography. Cryptanalysis [21] is the study of methods for obtaining the meaning of encrypted information, without access to the secret information which is normally required to do so. Typically, this involves finding the secret key. In non-technical language, this is the practice of code breaking or cracking the code, although these phrases also have a specialised technical meaning. Block ciphers [13] [14] encrypt blocks of data (typically 64 or 128 bits) in a fixed key-dependent way. The design of block ciphers is a well-studied area of research. The best-known block ciphers are the Data Encryption Standard (DES) and the Advanced Encryption Standard (AES). In the past decade, many new attacks on block ciphers have emerged, the most important ones being differential and linear cryptanalysis. Differential cryptanalysis is an example of a chosen-plaintext attack, while linear cryptanalysis is a known-plaintext attack. A good design should at least be resistant to these attacks. This algorithm can overcome this two attacks. A randomized algorithm or probabilistic algorithm is an algorithm which employs a degree of randomness as part of its logic. The algorithm typically uses uniformly random bits as an auxiliary input to guide its behaviour, in the hope of
Page 60
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
achieving good performance in the "average case" over all possible choices of random bits [4] [5]. Formally, the algorithm's performance will be a random variable determined by the random bits; thus either the running time, or the output (or both) are random variables. Here in this algorithm the key is nothing but the random positions of the given n × n matrix.. II. BRIEF OVERVIEW OF CRYPTOGRAPHY The history of cryptography [3]begins thousands of years ago. Until recent decades, it has been the story of what might be called classic cryptography that is, of methods of encryption that use pen and paper, or perhaps simple mechanical aids. In the early 20th century, the invention of complex mechanical and electromechanical machines, such as the Enigma rotor machine, provided more sophisticated and efficient means of encryption; and the subsequent introduction of electronics and computing has allowed elaborate schemes of still greater complexity, most of which are entirely unsuited to pen and paper. The development of cryptography has been paralleled by the development of cryptanalysis of the "breaking" of codes and ciphers. The discovery and application, early on, of frequency analysis to the reading of encrypted communications has on occasion altered the course of history. Thus the Zimmermann Telegram triggered the United States' entry into World War I; and Allied reading of Nazi Germany's ciphers shortened World War II, in some evaluations by as much as two years. The Germans made heavy use (in several variants) of an electromechanical rotor based cipher system known as Enigma. The German military also deployed several mechanical attempts at a one time pad. Bletchley Park called them the Fish cipher's, and Max Newman and colleagues designed and deployed the world's first programmable digital electronic computer, the Colossus, to help with their cryptanalysis. The German Foreign Office began to use the one-time pad in 1919; some of this traffic was read in WWII partly as the result of recovery of some key material in South America that was insufficiently carefully discarded by a German courier. The era of modern cryptography [7] really begins with Claude Shannon, arguably the father of mathematical cryptography. In 1949 he published the paper Communication Theory of Secrecy Systems in the Bell System Technical Journal, and a little later the book Mathematical Theory of Communication with Warren Weaver. These, in addition to his other works on information and communication theory established a solid theoretical basis for cryptography and for cryptanalysis. And with that, cryptography more or less disappeared into secret government communications organizations such as
Velammal College of Engineering and Technology, Madurai
the NSA. Very little work was again made public until the mid '70s, when everything changed. In 1969[6] two major public (ie, non-secret) advances. First was the DES (Data Encryption Standard) submitted Cryptography/History by IBM , at the invitation of the National Bureau of Standards (now NIST), in an effort to develop secure electronic communication facilities for businesses such as banks and other large financial organizations. After 'advice' and modification by the NSA, it was adopted and published as a FIPS Publication (Federal Information Processing Standard) in 1977. It has been made effectively obsolete by the adoption in 2001 of the Advanced Encryption Standard, also a NIST competition, as FIPS 197. DES was the first publicly accessible cipher algorithm to be 'blessed' by a national crypto agency such as NSA. The release of its design details by NBS stimulated an explosion of public and academic interest in cryptography. DES [19], and more secure variants of it, are still used today, although DES was officially supplanted by AES (Advanced Encryption Standard)[18] in 2001 when NIST announced the selection of Rijndael, by two Belgian cryptographers. DES remains in wide use nonetheless, having been incorporated into many national and organizational standards. However, its 56-bit key-size has been shown to be insufficient to guard against brute-force attacks (one such attack, undertaken by cyber civil-rights group The Electronic Frontier Foundation, succeeded in 56 hours the story is in Cracking DES, published by O'Reilly and Associates). As a result, use of straight DES encryption is now without doubt insecure for use in new crypto system designs, and messages protected by older crypto systems using DES[19] should also be regarded as insecure. The DES key size (56-bits) was thought to be too small by some even in 1976, perhaps most publicly Whitfield Diffie. There was suspicion that government organizations even then had sufficient computing power to break DES messages and that there may be a back door due to the lack of randomness in the 'S' boxes. Second was the publication of the paper New Directions in Cryptography by Whitfield Diffie and Martin Hellman. This paper introduced a radically new method of distributing cryptographic keys, which went far toward solving one of the fundamental problems of cryptography [8], key distribution. It has become known as Diffie-Hellman key exchange. The article also stimulated the almost immediate public development of a new class of enciphering algorithms, the asymmetric key algorithms. In contrast, with asymmetric key encryption, there is a pair of mathematically related keys for the algorithm, one of which is used for encryption and the other for decryption. Some, but not all, of these algorithms have the additional property that one of the keys may be made public since the other cannot be (by any currently known method) deduced from the 'public'
Page 61
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
key. The other key in these systems is kept secret and is usually called, somewhat confusingly, the 'private' key[12]. An algorithm of this kind is known as a public key / private key algorithm, although the term asymmetric key cryptography is preferred by those who wish to avoid the ambiguity of using that term for all such algorithms, and to stress that there are two distinct keys with different secrecy requirements. III. SECURITY ANALYSIS In this section we discuss the security strength of the randomized encryption paradigm under five most common cryptographic attack types: Brute force attack, cipher textonly attack, known plaintext attack, chosen plaintext attack and chosen cipher text attack. G. Brute Force attack In cryptography, a brute force attack[15] is a strategy used to break the encryption of data. It involves traversing the search space of possible keys until the correct key is found. The selection of an appropriate key length depends on the practical feasibility of performing a brute force attack. By obfuscating the data to be encoded, brute force attacks are made less effective as it is more difficult to determine when one has succeeded in breaking the code. In my algorithm, If n is large, then the key space is of n2 dimension and it is impossible to search the key space. H. Cipher text only A known cipher text attack[16] is an attack where the cryptanalyst only has access to encrypted cipher text. A known cipher text attack is the easiest of the common cryptanalysis attacks to mount, because is requires the least amount of control over the encryption device. Conversely, the known cipher text is the most difficult of the common methods of cryptanalysis to execute successfully, because so little knowledge is known to begin with. Let the cipher text c be available. Since the key stores the bit positions for both encryption and decryption, it is not possible to obtain the key from the cipher text and the encryption /decryption process. I. Known plain text attack The known-plaintext attack (KPA)[20] is an attack model for cryptanalysis where the attacker has samples of both the plaintext and its encrypted version (cipher text) and is at liberty to make use of them to reveal further secret information such as secret keys and code books. Suppose a plain text_ cipher text pair (p, c) is available. The ijth element of c, cij is obtained by the form
Velammal College of Engineering and Technology, Madurai
⎛n l ⎞ cij= ⎜ ∑ cil ∑cmj-pij⎟ %2 --- > 1 ⎝i=1 j=1 ⎠ where cil, cmj values are defined based on the of (i, l) and (m, j) in the key. Since cil and cmj can take either 0 or 1, eqn1 involves 22n-2 possibilities. Hence, it is impossible to solve. Therefore we cannot find the key k. J. Chosen-plaintext attack A chosen-plaintext attack (CPA) [11] is an attack model for cryptanalysis which presumes that the attacker has the capability to choose arbitrary plaintexts to be encrypted and obtain the corresponding cipher texts. The goal of the attack is to gain some further information which reduces the security of the encryption scheme. In the worst case, a chosen-plaintext attack could reveal the scheme's secret key. Suppose we take a plaintext p as a n × n block of zeros then the encryption process yields the cipher text c as (n × n) block of zeros. If we take a plain text block p with 0 entry except one, say(i, j) only equal to 1. Then the encryption process spreads the value randomly (as per the key order) in the cipher block c. from the pair (p-c) it is impossible to obtain the key k. K. Chosen-cipher text attack A chosen-cipher text attack (CCA)[2] is an attack model for cryptanalysis in which the cryptanalyst gathers information, at least in part, by choosing a cipher text and obtaining its decryption under an unknown key. Suppose a cipher text c is choosen. The decryption process uses the key in the reverse order and yields the plain text. The key spreads zeros & ones randomly in the plain text according to the position order. Infact, there would be unbalanced number of zeros and ones, as compared to the cipher text. Therefore impossible to find the key. IV ALGORITHM Security is a major concern in an increasingly multimediadefined universe where the internet serves as an indispensable resource for information and entertainment. This algorithm protect and provide access to critical and time-sensitive copyrighted material or personal information. L. Step 1: Randomized key generation Generate a pair of values (i,j) randomly, 1<=i<=m, 1<=j<=m, such that no pair (i,j) is repeated more than once.
Let K denote the collection of pairs (i,j) stored as per the order of generation. M. Step 2: Encryption Let p be a plaintext of size n. Let a = n(mod m²).
Page 62
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
If a = 0 set k = n/m² else set k = n/m² + 1. The plaintext p is divided in to k blocks, say p1,p2,---,pk, such that the first m entries of p form the first row of p1, second m elements the second row of p1 and so on. Similarly we define p2,p3,---,pk. If a ≠ 0 then we add the require number of zeros at the end of the plaintext, so that the last block pk is of order m × m. For l = 1,…,k. Let aij be the (i,j)th element of pl. Change aij as per the generation order of the pair(i,j) as follows, n ⎛ n ⎞ a^ij = ⎜ ∑arj+ ∑ais -aij ⎟ ⎝ r=1 s=1 ⎠ If a^ij (mod 2) = 0, set aij= 1 else set aij = 0. Let cl be the m × m matrix with the modified aij. c1 is the ciphertext corresponding to p1. N. Step 3: Decryption For l = 1,---,k. Let aij be the (ij)th element of cl. Change aij as per the reverse order of generation of the pair (i,j) as follows, n ⎛ n ⎞ a^ij = ⎜ ∑arj + ∑ais -aij ⎟ ⎝ r=1 s=1 ⎠ If a^ij (mod 2) = 0, set aij = 1 else set aij = 0. The (aij) = pl, the original plaintext block. The plaintext p = p1,p2,---,pk.
IV. EXPERIMENTS AND PERFORMANCE EVALUATION With the rapid development of communication techniques, it becomes more and more practical to talk with anyone anywhere. As an important aspect, multimedia (image, video, audio, etc.) enriches humans’ daily life. Nowadays, multimedia communication is in close relation with the activities in entertainment, politics, economics, militaries, industries, etc., which makes it urgent for protecting multimedia security, e.g., confidentiality, integrity, ownership or identity. Generally, multimedia security is different from text/binary data security since multimedia content is often of large volumes, with interactive operations, and requires real-time responses. As a result, the last decade has witnessed an explosive advance in multimedia applications that are dramatically changing every aspect of our modern life, from business, manufacturing, transportation, medicine and finance to education, research, entertainment and government. Although encrypting the entire multimedia content by a
Velammal College of Engineering and Technology, Madurai
traditional cryptographic cipher (e.g., the block or stream cipher) yields a satisfactory level of security, such an approach does have several shortcomings. First, the computational cost associated with encrypting the entire multimedia content is often high due to the large data size. Second, the encryption and decryption operations add another level of complexity to the system. In most cases, additional hardware or software functions are needed in order to implement it. Hence, I have developed an efficient yet secure multimedia encryption technique. The algorithm uses the experimental environment, CPU: Intel(R) core(TM)2 DVOE7200 @ 2.53GHz, 0.99 GB of RAM; Operating System: Windows XP Professional. O. Results: Block size : 8 bits COMPARISON OF DIFFERENT FILE FORMATS S. NO.
FILE TYPE & SIZE
TIME FOR ENCRYPTION
TIME FOR DECRYPTION
1.
Txt,1kb
78ms
63ms
2.
Doc,551kb
1secs 360 ms
1secs 47 ms
3.
Bmp,14.3kb
141 ms
143 ms
4.
Jpeg,1.69mb
3 secs 469 ms
3 secs 266ms
5.
Pdf,245kb
562 ms
531 ms
6.
Xls,29kb
141 ms
125 ms
7.
Mp3,50.3kb
171 ms
128 ms
8.
Wav,45.3mb
36 secs
35 secs
9.
Vob,62.8mb
47 secs
46 secs
Recent advances in the field of video communications have emphasized the importance of delivering video contents to a wide variety of devices, through different networks, and using a growing number of codecs and standards. In addition to the performance challenges behind video communications, the security issues have received increasing attention from the research community. Among the most relevant security aspects of contemporary video communication systems are: encryption, watermarking and authentication. File format: video file File size:62.8mb
Page 63
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
COMPARISON OF TIME FOR DIFFERENT BLOCK SIZES OF A VIDEO FILE BLOCK SIZE
TIME FOR
TIME FOR
(BITS)
ENCRYPTION
DECRYPTION
8
47 secs
46 secs
16
1 min 10 secs
1 min 9 secs
32
1 min 31 secs
1 min 30 secs
64
2 min 35 secs
2 min 34 secs
V. CONCLUSION Since the development of cryptology in the industrial and academic worlds in the seventies, public knowledge and expertise have grown in a tremendous way, notably because of the increasing presence of electronic communication means in our lives. Block ciphers are inevitable building blocks of the security of various electronic systems. Recently, many advances have been published in the field of public-key cryptography, being in the understanding of involved security models or in the mathematical security proofs applied to precise cryptosystems. Unfortunately, this is still not the case in the world of symmetric-key cryptography and the current state of knowledge is far from reaching such a goal. In this paper we developed a novel dynamic symmetric key generation scheme for multimedia data encryption. VI. REFERENCES [1] M. Bellare and P. Rogaway, “Robust computational secret sharing and a unified account of classical secret-sharing goals”, Proceedings of the 14th ACM Conference on Computer and Communications Security (CCS), ACM, 2007. [2] J. Zhou, Z. Liang, Y. Chen, and O. C. Au, “Security analysis of multimedia encryption schemes based on multiple Huffman table”, IEEE Signal Processing Letters, vol. 14, no. 3, pp. 201– 204, 2007. [3] M. Bellare, A. Boldyreva and A. O'Neill, “Deterministic and efficiently searchable encryption”, Advances in Cryptology - Crypto 2007 Proceedings, Lecture Notes in Computer Science Vol. 4622, A. Menezes ed, Springer-Verlag, 2007. [4] Elaine Barker and John Kelsey, “Recommendation for Random Number Generation Using Deterministic Random Bit Generators”, NIST Special Publication 800-90. Revised March 2007. [5] M. Grangetto, E. Magli and G. Olmo, “Multimedia selective encryption by means of randomized arithmetic
Velammal College of Engineering and Technology, Madurai
coding”, IEEE Transactions on Multimedia, vol. 8, no. 5, 2006. [6] M. Abdalla, M. Bellare, D. Catalano, E. Kiltz, T. Kohno, T. Lange, J. Malone-Lee, G. Neven, P. Paillier and H. Shi, “Searchable Encryption Revisited: Consistency Properties”, Relation to Anonymous IBE, and Extensions. Advances in Cryptology - Crypto 2005 Proceedings, Lecture Notes in Computer Science Vol. 3621, V. Shoup ed, SpringerVerlag, 2005. [7] K. Pietrzak, M. Bellare and P. Rogaway, “Improved Security Analyses for CBC MACs”, Advances in Cryptology - Crypto 2005 Proceedings, Lecture Notes in Computer Science Vol. 3621, Springer-Verlag, 2005. [8] Y. Lu and S. Vaudenay, “Faster correlation attack on Bluetooth keystream generator E0. In M. Franklin, editor, Advances in Cryptology - Crypto 2004, 24th Annual International Cryptology Conference, Santa Barbara, California, USA, August 15-19, 2004. Proceedings, volume 3152 of Lecture Notes in Computer Science, pages 407 - 425. Springer-Verlag, 2004. [9] L. S. Choon, A. Samsudin, and R. Budiarto, “Lightweight and cost-effective MPEG video encryption, in Proc. of Information and Communication Technologies”, From Theory to Applications, 2004, pp. 525–526. [10] B. Canvel, A. Hiltgen, S. Vaudenay, and M. Vuagnoux,”Password interception in a SSL/TLS channel”, In D. Boneh, editor, Advances in Cryptology - Crypto 2003, 23rd Annual International Cryptology Conference, Santa Barbara, California, USA, August 17-21, 2003, Proceedings, volume 2729 of Lecture Notes in Computer Science, pages 583-599. Springer-Verlag, 2003. [11] ] M. Hellman, R.Merkle, R. Schroeppel, L.Washington, W. Diffie, S. Pohlig, and P. Schweitzer, “Results of an initial attempt to cryptanalyze the NBS Data Encryption Standard”. Technical Report SEL 76-042, Department of Electrical Engineering, Stanford University, 1976. [12] M. Bellare and B. Yee, “Forward-Security in PrivateKey Cryptography”, Topics in Cryptology - CT-RSA 03, Lecture Notes in Computer Science Vol. 2612, M. Joye ed, Springer-Verlag, 2003. [13] H. Gilbert and M. Minier, “New results on the pseudorandomness of some block cipher constructions”, In M. Matsui (Ed.) Fast Software Encryption - FSE 2001, Lecture Notes in Computer Science 2355, Springer-Verlag, 2002, 248– 266. [14] P. Rogaway, M. Bellare, J. Black and T. Krovetz, “OCB: A block-cipher mode of operation for efficient authenticated encryption”, Proceedings of the 8th ACM Conference on Computer and Communications Security (CCS), ACM, 2001.
Page 64
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[15] J. Kelsey, Bruce Schneier, David Wagner, and C. Hall, “Cryptanalytic Attacks on Pseudorandom Number Generators, Fast Software Encryption”, Fifth International Workshop Proceedings (March 1998), Springer-Verlag, 1998, pp. 168-188. [16] Mitsuru Matsui, “The First Experimental Cryptanalysis of the Data Encryption Standard”. In Advances in Cryptology Proceedings of CRYPTO ’94, Lecture Notes in Computer Science 839, Springer-Verlag, 1994. [17] Alfred J.Menezes, Paul C. van Oorschot, and Scott A. Vanstone, Handbook of Applied Cryptography, CRC Press, 1997. [18] Christof Paar and Jan Pelzl, "The Advanced Encryption Standard", Chapter 4 of "Understanding Cryptography, A Textbook for Students and Practitioners". Springer, 2009. [19] Diffie, Whitfield and Martin Hellman, "Exhaustive Cryptanalysis of the NBS Data Encryption Standard" IEEE Computer 10(6), June 1977, pp74–84. [20] John Kelsey, Stefan Lucks, Bruce Schneier, Mike Stay, David Wagner, and Doug Whiting, Improved Cryptanalysis of Rijndael, Fast Software Encryption, 2000 pp213–230.
Velammal College of Engineering and Technology, Madurai
Page 65
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Secure Multiparty Computation Based Privacy Preserving Collaborative Data Mining J.Bhuvana#1,T.Devi#2 #
School of Computer Science and Engineering, Bharathiar University Coimbatore-46, India 1
[email protected] 2 [email protected]
Abstract— In today’s competitive world, businesses are competing to stay ahead of their competitors by predicting the future based on intelligent analysis of available data in decision making. Towards this, various techniques such as reporting, online analytical processing, analytics, data mining, business performance management, text mining and predictive analytics are used. Data Mining is the process of processing large volumes of data, searching for patterns and relationships within that data. In case of voluminous business data being analysed, collaborative data mining techniques are practised. In collaborative data mining, preserving data privacy is a biggest challenge. This paper gives a brief description of Data Mining, Privacy Preserving in Data Mining (PPDM), introduces Secure Multiparty Computation (SMC) based Privacy Preserving Collaborative Data Mining (PPCDM). Keywords— Privacy, Security, Privacy Preserving Data Mining, Privacy Preserving Collaborative Data Mining, Secure Multiparty Computation.
I. INTRODUCTION Recent advances in data collection, data dissemination and related technologies have initiated a new era of research where existing data mining algorithms should be reconsidered from the point of view of privacy preservation. Privacy refers to the right of users to conceal their personal information and have some degree of control over the use of any personal information disclosed to others. To conduct data mining, one needs to collect data from various parties. Privacy concerns may prevent the parties from directly sharing the data and some types of information about the data. The way in which multiple parties conduct data mining collaboratively without breaching data privacy is a challenge. In recent years, the research community has developed numerous technical solutions for the privacy preserving data mining. However, Clifton has pointed out “the notion of privacy that satisfies both technical and societal concerns is unknown” [3]. Security is a necessary tool to build privacy, but a communication or transaction environment can be
Velammal College of Engineering and Technology, Madurai
very secure, yet totally unprivate. Security and privacy are closely related technologies, Privacy is about informational self-determination, the ability to decide what information about one goes where. Security offers the ability to be confident that those decisions are respected. Data Mining is an information extraction activity whose goal is to discover hidden facts contained in databases. Using a combination of machine learning, statistical analysis, modeling techniques and database technology, data mining finds patterns and relationships in data and infers rules that allow the prediction of future results. Privacy preserving data mining allows multiple data holders to collaborate to compute important information while protecting the privacy of other information using the technological tools such as cryptography, data perturbation and sanitization, access control, inference control and trusted platforms. To use the data mining algorithms, all parties need to send their data to a trusted central place to conduct the mining. However, in situations with privacy concerns, the parties may not trust anyone. This type of problem is called Privacy Preserving Collaborative Data Mining (PPCDM) [10]. For each data mining problem, there is a corresponding PPCDM problem. Collaborative data mining is categorised into homogeneous collaboration and heterogeneous collaboration. In homogeneous collaboration, each party has the same sets of attributes and in heterogeneous collaboration, each party has different sets of attributes. Fig.1 shows how a traditional data mining problem could be transformed to PPCDM problem. Security and privacy are related but different [11]. In the privacy preserving data mining, the system needs secured channels to protect network attackers. The system categorizes the protection into two layers as shown in Fig.2. One is the protection against the collaborative parties; the other is protection against network attackers. Without loss of generality, attacks from collaborative parties are considered inside attacks, these parties are called inside attackers; attacks outside the collaborative parties are considered outside attacks, the attackers who conduct the attacks are called outside attackers. To protect against outside attackers, the system need
Page 66
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
to rely on secure channels. The research work by Dolve et al [5] assumes that the communications between collaborative parties are encrypted by a non-malleable encryption scheme.
The focus of this research is to prevent inside attackers from knowing private data in collaborative data mining. Prevention of inside attackers is different from prevention of outside attackers in that the inside attackers usually have more knowledge about private data than outside attackers. Furthermore, the goal of collaborative data mining is to obtain valid data mining result. However, the result itself may disclose the private data to inside attackers. Therefore, the system cannot hope to achieve the same level of protection in privacy preserving collaborative data mining as in general secure communications which protect against the outside attackers. However, the system would like to prevent the private data from being disclosed during the mining stage [5]. Particularly, in this work, the privacy preservation means that multiple parties collaboratively get valid data mining results while disclosing no private data to each other or any party who is not involved in the collaborative computations. The authors are currently carrying out research on Secure Multiparty Computation based Privacy Preserving Collaborative Data Mining with the specific objectives of designing and developing an algorithm for Secure Multiparty Computation based Privacy Preserving Collaborative Data Mining and to analyse efficiency of the algorithm by scaling up with various factors such as number of parties involved in the computation, encryption key size and size of data set.
Fig.1 Privacy Preserving Non-Collaborative and Collaborative Data Mining .
Fig.2 Inside Attackers vs Outside Attackers
Velammal College of Engineering and Technology, Madurai
II. LITERATURE REVIEW In privacy-preserving data mining, to protect actual data from being disclosed, one approach is to alter the data in a way that actual individual data values cannot be recovered, while certain computations can still be applied to the data. This is the core idea of randomization-based techniques. The random perturbation technique is usually realised by adding noise or uncertainty to actual data such that the actual values are prevented from being discovered. Since the data no longer contains the actual values, it cannot be misused to violate individual privacy. Randomization approaches were first used by Agrawal and Srikant [1] to solve the privacy-preserving data mining problem. Following the idea of secure multiparty computation [6], Lindell and Pinkas [7] introduced a secure multi-party computation technique for classification using the ID3 algorithm, over horizontally partitioned data. Encryption is a well-known technique for preserving the confidentiality of sensitive information. Comparing with other techniques described, a strong encryption scheme can be more effective in protecting the data privacy. One of the encryption schemes is the Homo-morphic encryption which allows certain computations performed on encrypted data without preliminary decryption operations. Homo-morphic encryption is a very powerful cryptographic tool and has been applied in several research areas such as
Page 67
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
electronic voting and on-line auction. Privacy preserving data mining has generated many research successes. However, as Clifton [3] pointed out. The proposed system present a definition of privacy and use Homo-morphic encryption and digital envelope technique to achieve collaborative data mining without sharing the private data among the collaborative parties. In particular, the research provides a solution for k-nearest neighbor classification with vertical collaboration. III. SECURED MULTIPARTY COMPUTATION MODEL FOR COLLABORATIVE DATA MINING
i=1
Support A =
Tot. Nodes
∑ Database Size (i) i=1
Tot. Nodes
∑ Support Count AB (i) i=1
Support AÆB = Tot. Nodes
∑ Database Size (i)
P. Secured Multiparty Computation (SMC) The basic idea of SMC is that a computation is secure if at the end of the computation, no party knows anything except its own input and the results. One way to view this is to imagine a trusted third party; everyone gives their input to the trusted third party, who performs the computation and sends the results to the participants. The concept of SMC was first introduced by Yao [13] and it has been proved that for any function, there is a SMC solution [12]. The approach used is: the function F to be computed is first represented as a combinatorial circuit, and then the parties run their private function in their node on the network. Every participant gets corresponding input values and out values for every node. SMC not only preserves individual privacy, it also preserves leakage of any information other than the final result. However, traditional SMC method requires a high communication overhead; they do not scale well with the network size. Q. Association Rule Mining Computing Association rules without disclosing individual data items is straightforward. Distributed association rule mining techniques can discover rules among multiple nodes [14]. Mining association rules from vertically partitioned data, where the items are partitioned and each itemset is split between nodes, can be done by extending the apriori algorithm. Most of the apriori algorithm can be done locally at each of the nodes. The crucial step involves finding the support count of an itemset. The support determines how often a rule can be applied to a given data set and the confidence determines how frequently items in B appear in transactions that contain A. One can compute the global support and confidence of an association rule AÆB knowing only the local support of A, B and the size of each node’s data using the following formulae [1]:
i=1
Support AÆB Confidence
Velammal College of Engineering and Technology, Madurai
= Support A
R. Privacy Formalization A privacy-oriented scheme S preserves data privacy if for any private data T; the following holds [11]: |Pr (T |PPDMS) − Pr(T )| ≤ € Where, PPDMS: Privacy-Preserving data mining scheme. Pr(T | PPDMS): The probability that the private data T is disclosed after a Privacy-Preserving data mining scheme has been applied. Pr(T): The probability that the private data T is disclosed without any Privacy-Preserving data mining scheme being applied. Pr(T | PPDMS) - Pr(T): The probability that private data T is disclosed with and without Privacy-Preserving data mining schemes being applied. To achieve privacy-preserving data mining, reduce the whole algorithm to a set of component privacy oriented protocols. The privacy preserving data mining algorithm preserves privacy if each component protocol preserves privacy and the combination of the component protocols does not disclose private data. In the secure multiparty computation literature, a composition theorem describes a similar idea. A privacyoriented component protocol CP preserves data privacy if for any private data T, the following is held:
Tot. Nodes
∑ Support Count A (i)
AÆB
|Pr(T |CP) − Pr(T )| ≤ € Where,
Page 68
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
CP: Component protocol. Pr (T | CP): The probability that the private data T is disclosed after a Privacy-Preserving component protocol has been applied. Pr(T | CP) - Pr(T): The probability that private data T is disclosed with and without a Privacy-Preserving component protocol. S. Digital Envelope A digital envelope is a random number (or a set of random numbers) only known by the owner of private data[2] . To hide the private data in a digital envelope, to conduct a set of mathematical operations between a random number (or a set of random numbers) and the private data. The mathematical operations could be addition, subtraction, multiplication, etc. For example, assume the private data value is A. There is a random number R which is only known by the owner of A. The owner can hide A by adding this random number, e.g., A + R. T. Privacy-Preserving Collaborative Data Mining Problems Multiple parties, each having a private data set which is denoted by D1, D2, · · · and Dn respectively, want to collaboratively conduct a data mining task on the concatenation of their data sets. Because they are concerned about their data privacy or due to the legal privacy rules, neither party is willing to disclose its actual data set to others. P1 has a private data set D1, P2 has a private data set D2, · · · and Pn has a private data set Dn. The data set [D1 ∪ D2 ∪· · · ∪ Dn] forms a dataset, which is actually the concatenation of D1, D2, · · · and Dn. The n parties want to conduct k-nearest clustering over [D1 ∪ D2 ∪ · · · ∪ Dn] to obtain the mining results satisfying the given constraints. There are two types of collaborative models. In the vertical collaboration, diverse features of the same set of data are collected by different parties. In the horizontal collaboration, diverse sets of data, all sharing the same features, are gathered by different parties. U. Privacy-Preserving (K-NN)Classification
K-Nearest
Neighbor
Classification introduces a privacy constraint that is not easily addressed in the SMC model. If the target class in the training data is private information, the classifier could be used to scratch the privacy. One solution is not to build a global classifier. Instead, the parties collaborate to classify each instance. Therefore, the problem becomes how to share a classifier among the various distributed nodes while still supporting the classification function.
Velammal College of Engineering and Technology, Madurai
Currently many researchers developing methods for secure distributed K-NN classification; two different solutions for different level of security. One is security based on cryptographic assumptions and another one is finding roots of the large degree polynomials. The later ones is computationally hard. The K-NN classification is an instance based learning algorithm that has been shown to be very effective for a variety of problem domains. The objective of K-NN classification is to discover k nearest neighbors for a given instance, then assign a class label to the given instance according to the majority class of the k nearest neighbors. The nearest neighbors of an instance are defined in terms of a distance function such as the standard Euclidean distance. IV. PERFORMANCE EVALUATION The proposed system discloses the support values for each tested item set to all the participants. This discloses may cause for full disclosure in some cases. The research presents a new approach that does not disclose the support values during the execution and discloses less amount of information in general. In any site there is no external information about any other database. The collusion between parties should not be beyond a given threshold value. The various parties follow the protocol honestly. However, any attempt to deviate from the protocol will not result in disclosed information. They may try to use a correct protocol to infer information, but shall show that this will not be helpful for them. The parties may store intermediate or final results. The analysis of privacy is in terms of what can be inferred from those stored results only done with the help of a third untrusted party (this party is not trusted with the database, but it is trusted with computations). The parties change roles at the end of the phase. When information between the parties is shared, then only information in which some attributes are real (in one of the databases) is of use. That is, a fake transaction whose corresponding Transaction Identification (TID) in the other database is empty, is not considered at all. Its important to note that when each site computes large itemsets it doesn’t know whether the attributes corresponding to his real transaction, are real or not. V. CONCLUSIONS Worldwide markets are competitive and in order to capture markets, businesses are practising business intelligence techniques that involve team based data mining. Data mining helps in decision making and future prediction of market. In modern business world, collaboration is highly essential for carrying out business successfully. In this paper, data mining, privacy preserving data mining, secure multiparty computation techniques and the privacy problem of privacy preserving collaborative data mining are discussed. The authors are currently working on design of privacy preserving
Page 69
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
collaborative data mining algorithms using the concepts of secure multiparty computation. ACKNOWLEDGMENT I record my sincere thanks to Bharathiar University for providing necessary financial assistance to carry out my research. REFERENCES [1] R. Agrawal and R. Srikant, “Privacy-Preserving data mining”, in Proc. ACM SIGMOD, May 2000, pages 439–450. [2] D. Chaum, “Security without identification Communication”, ACM, 28(10), pp. 1030–1044, Oct. 1985. [3] C. Clifton, “What is privacy? critical steps for Privacy-Preserving data mining”, in IEEE ICDM Workshop on Security and Privacy Ascepts of Data Mining, Houston, Texas, USA, November, 2005, pp. 27-30. [4] T. Cover and P. Hart “Nearest neighbor pattern classification”, iIn IEEE Transaction of Information Theory, Vol. 13, pp. 21- 27, January, 1968. [5] Dolev, D. Dwork, and M. Naor, “Non-malleable cryptography”, in Proc. twenty-third annual ACM symposium on Theory of computing, New Orleans, Louisiana, United States, 1991,pp: 542 – 552. [6] O. Goldreich, “The foundations of cryptography”, Vol. 2, Cambridge University Press, 2004. [7] Y. Lindell and B. Pinkas, “Privacy preserving data mining”, Advances in Cryptology - Crypto2000, Lecture Notes in Computer Science, volume 1880, 2000. [8] P. Paillier, “Public-key cryptosystems based on composite degree residuosity classes”, Advances in Cryptography - EUROCRYPT ’99, Prague, Czech Republic, 1999, pp 223-238. [9] R. Rivest, L. Adleman, and M. Dertouzos, “ On data banks and privacy homomorphisms”, Foundations of Secure Computation, eds. R. A. DeMillo et al., Academic Press, pp. 169-179., 1978. [10] J. Zhan, “Privacy Preserving Collaborative Data Mining”, PhD thesis, Department of Computer Science, University of Ottawa, 2006. [11] J. Zhan, “Using Cryptography for Privacy Protection in Data Mining Systems”, Springer-Verlag Berlin Heidelberg, 2007. [12] O. Golreich (1998) “Secure Multiparty Computation (working draft)” [online]. Available: http://www.wisdom.weizmann.ac.il/home/oded/public html/foc.html [13] A.C. Yao, “Protocols for secure computations”, Proc. 23rd Annual IEEE Symposium on Foundations of Computer Science, 1982. [14] Agrawal, R. and Shafer, J.C,“Parallel mining of association rules,” IEEE Transactions on Knowledge and Data Engineering, Vol. 8, No. 6, 1996, pp.929-969.
Velammal College of Engineering and Technology, Madurai
Page 70
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Towards Customer Churning Prevention through Class Imbalance M.Rajeswari#1, Dr.T.Devi#2 #
School of Computer Science and Engineering, Bharathiar University, Coimbatore, India 1
[email protected] 2 [email protected]
Abstract— Technologies such as data warehousing, data mining, and campaign management software have made Customer Relationship Management (CRM) a new area where firms can gain a competitive advantage. Particularly through data mining a process of extracting hidden predictive information from large databases, organisations can identify their valuable customers, predict future behaviors, and enable firms to make proactive, knowledge-driven decisions. Data Mining along with Customer Relationship Management plays a vital role in today’s business environment. Customer churn, a process of retaining customer is a major issue. Prevention of customer churn is a major problem because acquiring new customer is more expensive than holding existing customers. In order to prevent churn several data mining techniques have been proposed. One among such method is solving class imbalance which has not received much attention in the context of data mining. This paper describes Customer Relationship Management (CRM), customer churn and class imbalance and proposes a methodology for preventing customer churn through class imbalance. Keywords— Data Mining, Customer Relationship Management, Churn, Class Imbalance.
I. INTRODUCTION World-wide businesses are competing to capture and retain customers and towards this Customer Relationship Management (CRM) are practiced. Organisations turn to CRM to enable them to be more effective in acquiring, growing and retaining their profitable customers. Customer relationship management has emerged as one of the demanding ways for the firms and it provides an effective way for customer satisfaction and retention [12]. In order to carry out CRM, the historical data about customers and their behavior need to be analysed. Towards such analysis, data mining techniques are used nowadays. Due to increase in competition between the companies they realized that customers are their valuable assets and retaining existing customers is the best way to survive. But it is analysed that it is more profitable to satisfy existing customers than to look for a new one [3]. Customer attrition,
Velammal College of Engineering and Technology, Madurai
also known as customer churn, customer turnover, or customer defection, is a business term used to describe loss of clients or customers. Recent studies analysed the importance of predicting as well as preventing customer churn in customer relationship management. Customer churn is often a rare event in service industries, but of great interest and great value, until recently, however, class imbalance has not received much attention in the context of data mining. A class imbalance is a term where one class occurs much more often than the other. Now, as increasingly complex real-world problems are addressed, the problem of imbalanced data is taking centre stage and class imbalance concept comes under rarity. Many researchers are concentrating on rarity that relates to rare classes, or, more generally, class imbalance. This type of rarity requires labeled examples and is associated with classification problems. The data set used to detect oil spills from satellite images provides a good example of a rare class. Because only 41 of the 937 satellite images contain oil slicks, it can be said that oil slicks are rare (i.e., a rare class). Several problems arise when mining rare classes and rare cases. A case corresponds to a region in the instance space that is meaningful with respect to the domain under study and the rare case is a case that covers a small region of the instance space and covers relatively few training examples, for example if bird is taken as a class, non-flying bird is rare case because only very few birds come under this category [22]. In order to prevent customer churn, researchers have proposed number of problems that arise when mining rare classes and rare cases. The problems include: improper evaluation metrics, lack of data: absolute rarity, relative lack of data: relative rarity, data fragmentation, inappropriate inductive bias and noise. To overcome these problems, several methods such as more appropriate evaluation metrics, non-greedy search techniques, using a more appropriate inductive bias, knowledge/human interaction, learn only the rare class, segmenting the data, accounting for rare items, cost-sensitive learning, sampling and other methods such as boosting, two
Page 71
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
phrase rule induction, place rare cases into separate classes are followed [22]. Among these methods non-greedy search techniques, using a more appropriate inductive bias, learn only the rare classes are at conceptual stage only [5]. The rest of this paper is organized as follows: In Section 2, an overview of Data Mining is presented. Section 3 discusses about Customer Relationship Management and particularly about Customer Churn. An introduction to Class imbalance usage in predicting customer churn is described in Section 4. Methods followed in order to solve class imbalance are detailed in Section 5 and finally Conclusions are given in Section 6. II. DATA MINING (DM) Organizations are implementing data warehousing technology, which facilitates enormous enterprise-wide databases and the amount of data that organizations possess is growing at a phenomenal rate and the next challenge for these organizations is how to transform it into useful information and knowledge. Data Mining is one of the technologies used for meeting this challenge; data mining is the exploration and analysis, by automatic or semiautomatic means, of large quantities of data in order to discover meaningful patterns and rules. Data mining is defined as the process of extracting interesting and previously unknown information from data, and it is widely accepted to be a single phase in a complex process known as Knowledge Discovery in Databases (KDD). This process consists of a sequence of the following steps: data cleaning, data integration, data selection, data transformation, data mining, pattern evaluation and knowledge presentation [23]. Thus data mining is the search of valuable information in large volumes of data. This combines the efforts of human, who design the databases, describe the problems and set the goals, and computers who process the data looking for patterns that match this goal. Data mining sits at the common frontiers of several fields including database systems, artificial intelligence, statistics, machine learning, and pattern recognition or data visualization. In order to ensure that the extracted information generated by the data mining algorithms is useful, additional activities are required, like incorporating appropriate prior knowledge and proper interpretation of the data mining results. The basic objective is to construct a model for one situation in which the answer or output is known and then apply that model to another situation in which the answer or output is desired. The best applications of the above techniques are integrated with data warehouses and other interactive, flexible business analysis tools. There are many reasons by which all business environment started to analyze the performance of data mining in order to recover from all problems [9]. Databases today are huge containing more than 1,000,000 entities/records/rows, from 10 to 10,000 fields/attributes/
Velammal College of Engineering and Technology, Madurai
variables, gigabytes and terabytes. Databases are growing at an unprecedented rate and decisions must be made rapidly with maximum knowledge. Currently business to business marketers and service retailers are the major users of data mining [1]. Data mining involves many different algorithms to accomplish different tasks. All these algorithms attempt to fit a model to the data. The algorithms examine the data and determine a model that is closest to the characteristics of the data being examined. The created data mining model can be either predictive or descriptive in nature (e.g. Fig. 1). A Predictive Model – makes a prediction about values of data using known results found from different data. Predictive modeling may be made based on the use of other historical data. A Descriptive Model – identifies patterns or relationships in data. Unlike the predictive model, a descriptive model serves as a way to explore the properties of the data examined, not to predict new properties.
Fig. 1 Data Mining Models and Tasks [10]
Clustering, Summarization, Association rules, and Sequence discovery are usually viewed as descriptive in nature. The individual data mining tasks are combined to obtain more sophisticated data mining applications. There are several data mining techniques available and used. Mostly used techniques are associations, classifications, sequential patterns and clustering. The basic aim of an association is to find all associations, such that the presence of one set of items in a transaction implies the other items. Classification develops profiles of different groups. Sequential patterns subject to a user–specified minimum constraint. Clustering segments a database into subsets or clusters. The classifications of the data mining techniques are: user–guided or verification–driven data mining and discovery–driven or automatic discovery of rules. Most of the techniques of data mining have elements of both the models. Data mining is used in several applications such as retail, banking, telecommunication and other relevant applications. For example, when telecommunication companies around the world face competition which is forcing them to aggressively market special pricing programs aimed at retaining existing
Page 72
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
customers and attracting new ones. Knowledge discovery in telecommunication includes Call detail record analysis and Customer loyalty, Call detail record analysis: Telecommunication companies accumulate detailed call records. By identifying customer segments with similar use patterns, the companies can develop attractive pricing and feature promotions and Customer loyalty: Some customers repeatedly switch providers, or “churn”, to take advantage of attractive incentives by competing companies. The companies can use data mining to identify the characteristics of customers who are likely to remain loyal once they switch, thus enabling the companies to target their spending on customers who will produce the most profit. Similarly knowledge discovery applications are emerging in variety of industries such as customer segmentation and manufacturing [8]. III. CUSTOMER RELATIONSHIP MANAGEMENT (CRM) The way in which companies interact with their customers has changed dramatically over the past few years. A customer’s continuing business is no longer guaranteed. As a result, companies have found that they need to understand their customer better, and to quickly respond to their needs. Customer Relationship Management is the term used for the business practice allowing businesses that have more than a few customers to better serve and manage the interactions with those customers. CRM is defined as “a comprehensive strategy and process of acquiring, retaining and partnering with selective customers to create superior value for the company and the customer. It involves the integration of marketing, sales, customer service, and the supply chain functions of the organization to achieve greater efficiencies and effectiveness in delivering customer value” [18]. CRM is defined by a framework consisting of four elements: Know, Target, Sell, and Service. CRM requires the firm to know and understand its markets and customers. This involves detail customer intelligence in order to select the most profitable customers and identify those longer worth targeting. CRM also entails development of the offer: which product to sell to which customers and through which channel. In selling, firms use campaign management to increase the marketing department’s effectiveness. Finally, CRM seeks to retain its customers through services such as call centers and help desks. CRM is essentially a two-stage concept: the task of the first stage is to master the basics of building customer focus. This means moving from a product orientation to a customer orientation and defining market strategy from outside-in and not from inside-out. The focus should be on customer needs rather than product features. Companies in the second stage are moving beyond the basics; they do not rest on their laurels but push their development of customer orientation by integrating CRM across the entire customer experience chain, by leveraging technology to achieve real-time customer
Velammal College of Engineering and Technology, Madurai
management, and by constantly innovating their value proposition to customers [8]. CRM consists of four dimensions: Customer Identification, Customer Attraction, Customer Retention/Churn, and Customer Development. Customer Management System contains these four dimensions in closed cycle and used to create a better understanding of their customers and create a long term relationship with the organization [2]. The three stages in customer life cycle are acquiring customers, increasing the value of the customer and retaining good customers. Data mining can improve profitability in each stage through integration with operational CRM systems or as independent applications (e.g. Fig. 2). Data mining helps to accomplish such a goal by extracting or detecting hidden customer characteristics and behaviors from large databases. The generative aspect of data mining consists of the building of a model from data [7]. CRM is not only applicable for managing relationships between businesses and consumers, but even more crucial for business customers. In B2B environments, transactions are more numerous, custom contracts are more diverse and pricing schemes are more complicated. CRM strategies, such as customised catalogues, personalised business portals, and targeted product offers, can help smooth this process and improve efficiencies for both companies [8].
Fig. 2 Three simple business goals for CRM [14]
From the perspective of customer behaviour, Reference [6] suggests that B2B buyers choose a supplier with whom they can develop a relationship; one they can go back to as required and one on which they feel they can depend. Once they have chosen a supplier, having invested this time and effort, they are more likely to stay with that supplier for longer. This invokes the equal importance of deploying CRM in both recruiting new customers and maintaining existing customers. V. Customer Churn Attrition or churn is a growing problem in many industries and is characterised by the art of customer switching companies,
Page 73
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
usually to take advantage of “a better deal”. One of the most important issues for companies is customer retention. A persistent change in both the cost and revenue sides is customer churn, the movement of customers from provider to provider in search of better and cheaper products and services. Specifically churn is the gross rate of customer loss during a given period. Churn is a term used to represent the loss of a customer, in order to prevent churn one should acquire more loyal customers initially and then identify customers who are most likely to churn. This type of predictive churn modeling is applied in the field of banking, mobile telecommunication and life insurances. Churn can be shown as follows: Monthly churn = (C0 + A1 - C1) / C0 C0 is the number of customers at the start of the month, A1 is the number of customers at the end of the month, C1 is the gross new customers during the month. There are several reasons for a particular customer to churn such as Price, Service Quality, Fraud, Lack of carrier responsiveness, Brand disloyalty, Privacy concerns, Lack of features, new technology or product introduced by competitors. Minimizing customer churn provides a number of benefits, such as: Minor investment in acquiring a new customer, higher efficiency in network usage, Increase of added-value sales to long term customers, Decrease of expenditure on help desk, Decrease of exposure to frauds and bad debts and higher confidence of investors. Churn can be defined and measured in different ways: “Absolute” Churn is number of subscribers disconnected, as a percentage of the subscriber base over a given period, “Line” or “Service” Churn is number of lines or services disconnected, as a percentage of the total amount of lines or services subscribed by the customers, “Primary Churn” is number of defections and “Secondary Churn” is drop in traffic volume, with respect to different typology of calls. Measuring churn is getting more and more difficult, growing tendency for Business users to split their business between several competing fixed network operators. Carrier selection enables Residential customers to make different kind of calls with different operators, Carrier pre-selection and Unbundling of the Local Loop makes it very difficult to profile customers according to their “telecommunication needs”. IV. CHURN PREDICTION USING CLASS IMBALANCE A dataset is imbalanced if the classification categories are not approximately equally represented. Recent years brought increased interest in applying machine learning techniques to difficult "real-world" problems, many of which are characterized by imbalanced data. Additionally the distribution of the testing data may differ from that of the training data, and the true misclassification costs may be unknown at learning time. Predictive accuracy, a popular choice for evaluating
Velammal College of Engineering and Technology, Madurai
performance of a classifier, might not be appropriate when the data is imbalanced and/or the costs of different errors vary markedly [17]. The class imbalance problem corresponds to domains for which one class is represented by a large number of examples while the other is represented by only few. Class imbalance falls under rarity which deals with both rare classes (i.e. class imbalance) and rare cases. Fig. 3 shows an artificial domain with two classes, A and B, where A is the rare (minority) class and B is the common (majority) class. Holding with established conventions, the rare class is designated the positive class and the common class is designated the negative class. The true decision boundaries are displayed with solid lines while the learned boundaries are displayed with dashed lines. The labeled examples are represented in the figure using the “+” and “-” symbols. The five subconcepts associated with class A are labeled A1A5. Subconcepts A2-A5 correspond to rare cases, whereas A1 corresponds to a fairly common case, covering a substantial portion of the instance space. The majority class is comprised of two subconcepts, B1 and B2. Subconcept B1 is a very general case that covers most of the instance space. Subconcept B2, on the other hand, corresponds to a rare case, demonstrating that common classes may contain rare cases. However, it is expected that rare classes, by their very nature, to contain a greater proportion of rare cases (a very rare class cannot contain any common cases). Fig. 3 also shows that rare cases A3, A4, and A5 cause small disjuncts to be formed in the learned classifier (rare case A2 is “missed” completely by the classifier). Rare cases, like rare classes, can be considered the result of a form of data imbalance and have in fact been referred to as within-class imbalances [13].
Fig. 3 Graphical representation of rare classes and rare cases [22]
To predict customer churn many approaches are carried out and one such approach is to handle imbalance classes, in order to solve the problem of imbalance classes researchers have introduced various methods such as: Improper evaluation metrics, Lack of data: absolute rarity, Relative lack of data: relative rarity, Data fragmentation, Inappropriate inductive bias,
Page 74
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
and Noise. Improper evaluation metrics: often, not the best metrics are used to guide the data mining algorithms and to evaluate the results of data mining, Lack of data: absolute rarity: the number of examples associated with the rare class is small in an absolute sense, which makes it difficult to detect regularities within the rare class, Relative lack of data: relative rarity : objects are not rare in absolute sense, but are rare relative to other objects, which makes it hard for greedy search heuristics, and more global methods are, in general, not tractable, Data fragmentation : Many data mining algorithms, like decision trees, employ a divide-and-conquer approach, where the original problem is decomposed into smaller and smaller problems, which results in the instance space being partitioned into smaller and smaller pieces. This is a problem because regularities can then only be found within each individual partition, which will contain less data, Inappropriate inductive bias: Generalizing from specific examples, or induction, requires an extra-evidentiary bias. Without such a bias ‘‘inductive leaps” are not possible and learning cannot occur. The bias of a data mining system is therefore critical to its performance. Many learners utilize a general bias in order to foster generalization and avoid overfitting. This bias can adversely impact the ability to learn rare cases and rare classes, Noise: Noisy data will affect the way any data mining system behaves, but interesting is that noise has a greater impact on rare cases than on common cases [22]. V. PROPOSED WORK Customer churn prevention as a part of customer relationship management approach is high on the agenda, and it is found that almost each and every company is forced to implement churn prediction model to detect possible churners before they effectively leave the company. In order to predict churn, more and more data mining techniques are applied and fortunately for the companies involved, churn is often a rare object, but of great interest and great value, there is a need for understanding how to model rare events. ‘‘The models developed in marketing are typically applied to situations where the events of interest occur with some frequency (e.g., customer churn, customer purchases). These models can break down when applied to setting where the behaviour of interest is rare. For example, when modelling the correlates of customer acquisition in a lowacquisition rate setting, the performance of the familiar logit model is often unacceptable. There may be opportunity to gain valuable insights from the statistics literature on the modelling of rare events” [11]. Due to the increase in complex real-world problems, the problem of imbalance data is taking center stage and many researchers are found to solve such problems using various techniques and methods. There are six data mining problems related to rarity, and lists ten methods to address them, they are improper evaluation metrics, lack of data: absolute rarity, relative lack of data:
Velammal College of Engineering and Technology, Madurai
relative rarity, data fragmentation, inappropriate inductive bias and noise. To overcome these problems several methods are followed to deal with rarity such as more appropriate evaluation metrics, non-greedy search techniques, using a more appropriate inductive bias, knowledge/human interaction, learn only the rare class, segmenting the data, accounting for rare items, cost-sensitive learning, sampling and other methods such as boosting, two phrase rule induction, place rare cases into separate classes [22]. Among those methods non-greedy search techniques, using a more appropriate inductive bias, learn only the rare classes are at conceptual stage. Non-Greedy Search Techniques: Due to several disadvantages in greedy method, Reference [22] has suggested carrying out non-greedy techniques to handle imbalance classes, one among such non-greedy techniques is genetic algorithm. To solve the problem genetic algorithm makes use of candidate solution instead of single solution and then employs stochastic operators. It gives a conclusion that genetic algorithms work better with attribute interaction and thus avoids getting stuck into local maxima, which all together makes genetic algorithms suitable finding rarity. This would give a clear idea that for what reason genetic algorithms are being increasingly used for data mining. Several systems have relied to the power of genetic algorithms to handle rarity. Some of them used a genetic algorithm to predict very rare events while and other researches tried to use a genetic algorithm to discover “small disjunct rules” [21]. Using a More Appropriate Inductive Bias: Several attempts have been taken to improve the performance of data mining systems with respect to rarity by choosing a more appropriate bias. The simplest approach involves modifying existing systems to eliminate some small disjuncts based on tests of statistical significance or using error estimation techniques. The hope is that these will remove only improperly learned disjuncts [20]. Unfortunately, this approach was shown not only to degrade performance with respect to rarity, but also to degrade overall classification performance. Many researches worked on this concept with some techniques and obtained better results, like the usage of C4.5 algorithm. But fortunately the result showed only mixed success, because they tried out using only overall classification accuracy instead of small disjuncts. Learn Only The Rare Classes: When learning a set of classification rule for all classes, it is found that several rare classes are ignored and hence, the only solution is to predict rare classes. This technique is proved using Support vector machines and neural networks, and one data mining technique that utilizes this recognition-based approach is Hippo. Ripper is a rule induction system that utilizes a separate-and-conquer approach to iteratively build rules to cover previously uncovered training examples. Each rule is grown by adding conditions until no negative examples are covered. It normally
Page 75
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
generates rules for each class from the most rare class to the most common class. Given this architecture, it is quite straightforward to learn rules only for the minority class—a capability that Ripper provides. VI. CONCLUSION Churn is often considered as a rare object and in order to handle these rare objects, class imbalance (i.e. rare classes) have been utilised by many researchers. This paper gives ideas about various data mining techniques that have been applicable for customer relationship management concept particularly in the prevention of customer churn by using class imbalance. The authors are currently designing a methodology for preventing customer churn through class imbalance.
ACKNOWLEDGMENT I record my sincere thanks to Bharathiar University for providing necessary financial assistance to carry out my research. REFERENCES [1] Armstrong, G., and P. Kolter, Principles of Marketing, Prentice Hall New Jersey, 2001. [2] Au, W. H., Chan, K. C. C., & Yao, X, “A novel evolutionary data mining algorithm with applications to churn prediction,” IEEE Transactions on Evolutionary Computation, vol. 7, pp. 532–545, 2003. [3] Berry, M. J. A., & Linoff, G. S, Data mining techniques second edition - for marketing, sales, and customer relationship management, Wiley, 2004. [4] Berson, A., Smith, S., & Thearling, K, Building data mining applications for CRM, McGraw-Hill, 2000. [5] Burez, J., Van den Poel. D, “Handling class imbalance in customer churn prediction,” Expert Systems with Applications, vol. 36, pp. 4626–4636, 2009. [6] Bush, R., The Interactive and Direct Marketing Guide, The Institute of Direct Marketing, Middlesex, 2002, Chapter 3.6. [7] Carrier, C. G., & Povel, O, “Characterising data mining software. Intelligent Data Analysis,” vol. 7, pp. 181–192, 2003. [8] Chris Rygielski, Jyan-Cheng Wang, David C.Yen, “Data Mining and Customer RElatonship Management,” Technology in Society, vol. 24, pp. 483-502, 2004. [9] Fayyad, U.M. (2003). Editorial. SIGKDD Explorations, 5(2). [10] Garver. M. S, “Using Data Mining for Customer Satisfaction Research. Marketing Research” vol. 14, no. 1, pp. 8–12, 2002. [11] Gupta, S., Hanssens, D., Hardie, B., Kahn, W., Kumar, V., Lin, N., et al, “Modeling customer lifetime value,” Journal of Service Research, 2006, 9 (2), 139–155. [12] Jayanthi Ranja., Vishal Bhatnagar, “Critical Success Factors For Implementing CRM Using Data Mining,” Journal of Knowledge Management Practice, vol. 9, 2008. [13] Japkowicz, N, “Concept learning in the presence of between class and within-class imbalances,” In Proceedings of the Fourteenth Conference of the Canadian Society for Computational Studies of Intelligence, 2001, pages 67-77, SpringerVerlag. [14] Jorg-Uwe Kietz, “Data Mining for CRM and Risk Management,” Knowledge discovery services and knowledge discovery applications, 2003.
Velammal College of Engineering and Technology, Madurai
[15] Kracklauer, A. H., Mills, D. Q., & Seifert, D, “Customer management as the origin of collaborative customer relationship management. Collaborative Customer Relationship Management - taking CRM to the next level,” 2004, 3–6. [16] Neslin, S., Gupta, S., Kamakura, W., Lu, J., & Mason, C, “Detection defection: Measuring and understanding the predictive accuracy of customer churn models. Journal of Marketing Research,” vol. 43(2), pp. 204–211, 2006. [17] Nitesh V. Chawla, “Data Mining For Imbalanced Datasets: An Overview,” pp. 853-867, 2005. [18] Parvatiyar, A., & Sheth, J. N, “Customer relationship management: Emerging practice, process, and discipline,” Journal of Economic & Social Research, vol. 3, pp. 1–34, 2001. [19] Regielski, C., Wang, J.C. & Yen, D.C, “Data Mining Techniques for Customer Relationship Management,” Technology in Society, vol. 24, pp. 483502, 2002. [20] Weiss, G. M, “Learning with rare cases and small disjuncts,” In Proceedings of the Twelfth International Conference on Machine Learning, pages 558-565, Morgan Kaufmann, 1995. [21] Weiss, G. M, “Timeweaver: a genetic algorithm for identifying predictive patterns in sequences of events,” In Proceedings of the Genetic and Evolutionary Computation Conference, pages 718-725, Morgan Kaufmann, 1999. [22] Weiss, G. M. (2004), “Mining with rarity: A unifying framework. SIGKDD Explorations,” vol. 6 (1), pp. 7–19, 2004. [23] Wirth, R. and Hipp, J, “CRISP-DM: Towards a standard process model for data mining,” In Proceedings of the 4th International Conference on the Practical Applications of Knowledge Discovery and Data Mining, 2000, pages 29-39, Manchester, UK.
Page 76
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Designing health care forum using semantic Search engine & diagnostic ontology #1
#2
Prof.Mr.V.Shunmughavel #1, Dr.P.Jaganathan #2, Department of CSE, K.L.N. College of Information Tech. Pottapalayam, 630611. Professor& Head, Department of Computer Applications, P.S.N.A. College of Engg & Tech, Dindigul. [email protected], [email protected]
Abstract. Healthcare Forum is a coalition of doctors, hospitals, health plans and state associations, that have joined together to improve the health care system. It should follow the best practices and evidence-based guidelines, with a strong emphasis on patient safety. Chronic diseases are on the rise and the population is ageing. Added to this, advancing technologies, new medications, and new procedures and providers have a wide array of tools to manage in order to deliver high quality health care. With the advent of semantic web, the fast dissemination of ontologies to represent and solve information causes a deep impact on knowledge retrieval as a whole. Ontology is increasingly seen as a key factor for enabling interoperability across heterogeneous systems and semantic web applications. The main objective of this research paper is to develop an effective functioning forum which integrates the unique association thinking of humans with an intelligent diagnostic Ontology to devise an automated semantic search engine which provides recommended solutions for patients, doctors and research activities in the medicine worldwide.
Keywords: Ontology, Semantic Search 1.
INTRODUCTION
In general any Forum is a group creativity technique designed to generate a large number of ideas for the solution to a problem. Although forum has become a popular group technique, researchers have generally failed to find evidence of its effectiveness for enhancing either quantity or quality of ideas generated. Because of such problems as distraction, social loafing, evaluation apprehension, and production blocking, brainstorming groups are little more effective than other types of groups, and they are actually less effective than individuals working independently. For this reason, there have been numerous attempts to improve traditional forum design
Velammal College of Engineering and Technology, Madurai
techniques or replace it with more effective variations of the basic technique. A good example to quote for the current forum sessions is the simple forum available over the internet. In these forums questions will be posted by users and answers for those questions will be posted by various users. Until now, the best answer for a particular question will be judged by a human. There is no guarantee that the answer selected as the best answer to be the perfect solution always. The thinking and decision making capability of human mind differs from person to person. So on the whole, opinion differs. In order to overcome the demerits in the current traditional forum sessions, in this research a novel idea is proposed for choosing the best recommended solution to a question. In general in any forum, first a user posts a question. Now users who are interested in that question can view that question and cast their answers. It is not necessary that the answer given by a user should be the perfect solution. Instead any number of solutions can be posted by users for a particular question. These solutions will now be available for everyone to view. In addition to this, the research focuses on providing a recommended solution using Ontology. With this aim, this work is organized as follows: Section 2 discusses the related work. Section 3 discusses the diagnostic ontology building methodologies. Section 4 discusses the proposed architecture of Health Care Forum. Finally Section 5 presents conclusion and future scope of this work. 2.
RELATED WORK
The main objective of the website [9] is to deliver updated knowledge for Practicing Anesthesiologists and extend helping hand at the time of facing critical cases and also to
Page 77
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
create awareness amongst them. Semantically enhanced information extraction system providing automatic annotation with references to classes in the ontology and instances in the knowledge base is presented. Based on these annotations, we perform IR-like indexing and retrieval, further extended by using the ontology and knowledge about the specific entities [10]. Web Ontology Language (OWL) is a language for defining and instantiating web ontologies (a W3C Recommendation). OWL is designed for use by applications that need to process the content of information instead of just presenting information to humans. This is used for knowledge representation and also is useful to derive logical consequences from OWL formal semantics [11]. Matthew Horridge[12]introduces the Protege-OWL plug-in for creating OWL ontologies. It gives a brief overview of the OWL ontology language and focuses on building OWLDL ontology and by using a description logic reasoner to check the consistency of the ontology and automatically compute the ontology class hierarchy. The vision of a semantic Web is extremely ambitious and would require solving many long-standing research problems in knowledge representation and reasoning, databases, computational linguistics, computer vision, and agent systems [13]. In OWL Web Ontology Language Guide [14], The OWL Web Ontology Language is intended to provide a language that can be used to describe the classes and relations between them that are inherent in Web documents and applications. This document demonstrates the use of the OWL language to 1. formalize a domain by defining classes and properties of those classes, 2. define individuals and assert properties about them, and 3. reason about these classes and individuals to the degree permitted by the formal semantics of the OWL language. Oscar Corcho [7] presents how to build ontology in the legal domain. Graciela Brusa [8] presents a discussion on the process and product of an experience in developing ontology for the Public Sector whose organization requires a strong knowledge management. Particularly, this process has been applied to develop ontology for Budget Domain. Pablo Castells [15] assumes a knowledge base has been built and associated to the information sources (the document base), by using one or several domain ontologies that describe concepts appearing in the document text. Our system can work with any arbitrary domain ontology with essentially no restrictions, except for some minimal requirements, which basically consist of conforming to a set of root ontology classes. The concepts and instances in the KB are linked to the documents by means of explicit, non embedded annotations to the documents.
Velammal College of Engineering and Technology, Madurai
3.
Building Diagnostic Ontology
This section describes the process of an experience in developing a Diagnostic Ontology for Health Care Domain (Fig.1) [7] [8]. 3.1.
Specification: The Ontology Goal and Scope
Fig. 1. Specification The scope limits the ontology, specifying what must be included and what must not. It is an important step Fig. 1 for minimizing the amount of data and concepts to be analyzed, especially for the extent and complexity of the diagnostic semantics. In successive iterations for verification process, it will be adjusted if necessary. 3.2.
Specification: Domain Description
In this analysis, the application to formulating the diagnostic ontology and its related documentations were studied and revised. Furthermore, meetings with a group of experts were carried out. This group was conformed by Doctors, and software engineers who bring informatics supports for these tasks. As it can be seen, the group of experts was very heterogeneous. In addition, they do not have much time to assign the meetings. This group was the support for knowledge acquisition during the ontology development. Then, we have to define different intermediate representations to communicate the knowledge acquired to the experts considering the background of each one and the time of meetings. Then, a brief description of the domain is presented.
Page 78
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3.3.
Specification: Motivating Scenarios
The motivation scenarios show problems that arise when people need information that the system does not provide. Besides, the scenario description contains a set of solutions to these problems that includes the semantic aspects to solve them. In order to define motivation scenarios and communicate them to the involved people, templates have been used. These templates were based on those proposed to specify case uses in object oriented methodology. The template describes: the name of the scenario, people who participate in the scenario, a brief scenario description, and a list of possible terms related to the scenario. Since this template shows the most important information in a concise way, it is useful when the experts do not have a lot of time to analyze the scenarios. 3.4. Specification: Ontology Granularity and Type According to the level of conceptualization and granularity, the ontology proposed here is domain ontology. Domain ontology describes the vocabulary related to a specific domain and, the ontology objective is to facilitate the discussion among the Health Care Forum members. 3.5. Conceptualization: Domain Conceptual Model
defined concepts. The UML class diagram can be used to express concepts in terms of classes and relationships among them. Although UML in its standard form is not suitable for semantic representation, the information modeled in the UML class diagram was the base for building the ontology term glossary, trying to include other concepts by means of generalization and specialization techniques. 3.5. Conceptualization: Identification of Classes, Relations and Attributes We used Concepts classifier trees to analyze hierarchies and attributes, binary relations, axioms and instances tables. For determining classes, we identified those terms of independent existence from the key terms list and the glossary. Once the conceptual model of the ontology has been created, the next step is to define relevant instances 3.7. Implementing the Diagnostic Ontology with PROTEGE
Fig. 3. Implementation
Fig. 2. Conceptualization In this step Fig. 2, a list of the most important terms was elaborated. The core of basic terms is identified first and then they are specified and generalized if necessary. Then with these concepts as reference, the key term list was defined. To properly understand the conceptual aspects in the context, a Unified Modeling Language (UML) [5] diagram was elaborated with the main relations among
Velammal College of Engineering and Technology, Madurai
In order to implement Fig. 3 the diagnostic ontology, Protege tool is an opt tool because of the fact that it is extensible and provides a plug-and-play environment that makes it a flexible base for rapid prototyping and application development. Protege ontologies can be exported into different formats including Relational Description Framework Schema and Web Ontology Language (OWL). Particularly, it is proposed to implement the Diagnostic Ontology in OWL. To compare the ontology implementation with its conceptualization, graphics using the OWLViz and Ontoviz plug-ins will be generated and compared with UML diagrams. On the one hand, OWLViz enables the class hierarchies in OWL Ontology to be viewed, allowing comparison of the asserted class hierarchy and the inferred class hierarchy.
Page 79
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
4.HEALTH CARE FORUM ARCHITECTURE & ENVIRONMENT This system aims at constructing a new Health Care Forum model, integrating semantic based diagnostic ontology module using semantic search engine. As depicted in Fig. 4, the architecture and the environment have several components. It also provides an overall high-level description of the interactions between the main components of the system. These parts and their interactions are described as follows:
concepts appearing in the document text. The concepts and instances in the KB are linked to the documents by means of annotations to the documents. Documents are annotated with concept instances from the KB by creating instances. Weights are computed automatically, based on the frequency of occurrence of the instances in each document. The annotations are used by the retrieval process. The Knowledge Base records every client’s domain knowledge, which comprises idea instances and the relationships between these instances. That is, this study assumes the existence of a given set of participants and their relevant knowledge. 4.3.
Semantic Search Engine Module
Semantic Search Engine Module is the prime module which acts like interface between the Data Module and Diagnostic Ontology. In this module Jena framework has been used for the implementation of RDF Data Query Language (RDQL) [3]. Jena is a Java toolkit which provides an API for creating and manipulating RDF models. Jena sources can be retrieved at http://jena.sourceforge.net/. RDQL is an implementation of an SQL-like query language for RDF. The RDQL query is executed against the knowledge base, which returns a list of instance tuples that satisfy the query. The semantic search engine extracts recommended solutions from the diagnostic ontology module with the help of instance tuples. These recommended solutions will be sent to the forum blackboard. Fig. 4: Proposed Architecture of Health Care Forum 4.1.
User Interface Module
This module has two components: Open Problem and Participants Component [1]. An “open problem” is an initial topic or issue given to the system. “Participants” means the registered users attending the Health Care Forum session. Both the components represent the given open problem, the registered users and posted questions with related topic.
4.2.
Data Module
This comprises two parts: Document Base and Knowledge Base(KB). The document base stores fundamental information about registered users, posted questions and replies. A knowledge base has been built and associated to the information sources (the document base), by using one or several domain ontologies that describe
Velammal College of Engineering and Technology, Madurai
4.4.
Diagnostic Ontology Module
Diagnostic Ontology is already created with the help of Domain Experts and implemented by using Protege tool and kept for use. Very often the Diagnostic Ontology is updated with recent findings and developments and used by Health Care Forum. 5. CONCLUSION Building domain ontologies is not a simple task when domain experts have no background knowledge on engineering techniques and/or they don’t have much time to invest in domain conceptualization. Sharing the best practice on ontology building can be useful for the whole community. In our work, we have also shown how ontologies could develop merging different methodologies and software engineering. Particularly this approach has been used to define Domain ontology for Medical Diagnostic System which could be used by Doctors and Experts of various domains belongs to Medicine. The proposed architecture for Health Care forum integrates the unique association thinking of humans with an
Page 80
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
intelligent diagnostic Ontology to devise an automated and suitable solution for Patients, Doctors and Research activities in the medicine worldwide. As future work, we plan to complete the implementation of our approach with practical evaluation. 6.
REFERENCES
[1] Soe-Tsyr Yuan and Yen-Chuan Chen (2008) “Semantic Ideation Learning for Agent-Based E-Brainstorming”, IEEE Transactions on Knowledge And Data Engineering, Vol. 20, No. 2. [2] A.R. Dennis, A. Pinsonneault, K.M. Hilmer, H. Barki, R.B. Gallupe, M. Huber, and F.Bellavance (2005) “Patterns in Electronic Brainstorming: The Effects of Synergy and Social Loafing on Group Idea Generation” Int’l J. e-Collaboration, vol. 1, no. 4, pp. 38- 57. [3] W.-L.Chang and S.-T. Yuan (2004) “iCare Home Portal: A Quest for Quality Aging e-Service Delivery,” Proc. First Workshop Ubiquitous Seaborne A., RDQL - A Query Language for RDF, W3C Member Submission. http://www.w3.org/Submission/2004/SUBM-RDQL-20040109/ [4] Smith,M., Welty.C., McGuinness.D. (2004) OWL Web Ontology Language Guide, W3C Recommendation 10 http://www.w3.org/TR/owl-guide/ [5] UML (2006) Unified Modeling Language. http://www.uml.org/ Brickley, D., Guha, R.V. (2004) RDF Vocabulary Description Language 1.0: RDF Schema. W3C Recommendation. http://www.w3.org/TR/rdf-schema/ [6] Caliusco M. L. (2005), “A Semantic Definition Support of Electronic Business Documents in e–Colaboration”, PhD thesis. UTN - F.R.S.F. Santa Fe, Argentina. [7] Corcho O, Fernández-López M, Gómez-Pérez A, LópezCima A. (2005), “Building legal ontologies with METHONTOLOGY and WebODE. Law and the Semantic Web, Legal Ontologies, Methodologies, Legal Information Retrieval, and Applications”, Australian Computer Society, Inc. T Australasian Ontology Workshop (AOW 2006), Hobart, Australia. Conferences in Research and Practice in Information Technology, Vol. 72. [8] Graciela Brusa, “A Process for Building a Domain Ontology: an Experience in Developing a Government Budgetary Ontology” Dirección Provincial de Informática, San Martín 2466, Santa Fe (Santa Fe), Argentina, [email protected] [9] www.isakanyakumari.com [10] www.ontotext.com/kim/ semanticannotation.html [11] Goutam Kumar Saha, ACM Ubiquity, v.8, 2007 [12] Matthew Horridge, Holger Knublauch, Alan Rector, Robert Stevens, Chris Wroe, “A Practical Guide To Building OWL Ontologies Using The Prot´eg´e-OWL”, Plugin and CO-ODE Tools Edition 1.0, August 27, 2004. [13] Ian Horrocks, Ontologies and the Semantic Web, communications of the acm|december 2008 | vol. 51| no. 12. [14] http://www.w3.org/TR/owl-guide/ [15] Pablo Castells, Miriam Ferna´ndez, and David Vallet An Adaptation of the Vector-Space Model for Ontology-Based Information Retrieval, IEEE Transactions On Knowledge And Data Engineering, sVol. 19, No. 2, February 2007.
Velammal College of Engineering and Technology, Madurai
Page 81
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
An Enhancing the Life Time of Wireless Sensor Networks Using Mean Measure Mechanism P.Ponnu Rajan #1, D.Bommudurai*2 #
Department of Computer Applications, K.L.N College of Information Technology, Pottapalayam, Sivagangai District, Tamilnadu, India. 1
[email protected]
*
Department of Computer Science and Engineering, K.L.N College of Information Technology, Pottapalayam, Sivagangai District, Tamilnadu, India. 2
[email protected]
Abstract— Wireless sensor Network function with many routing protocols from the LEACH to High power short distance routing protocol. Still there are some unresolved issues in routing protocol in wireless sensor networks. In this paper the proposed algorithm is Distance based energy aware routing named as DER. The distance based Energy aware routing protocol that selects the route to the base station based on mean measure and the node that is nearest to the base station with in the radius. The energy of the node with in the range is calculated as mean measures or threshold. By comparing the threshold with the energy level of the node the transmission is take place to the base station. The distance-based energy aware routing will efficiently reduces the time delay, number of nodes and energy consumption in WSN. The energy consumption of the node in the network is well balanced and prolonging the survival time of the wireless sensor network. Keywords— Wireless Sensor Networks; Mean measure; Energy aware routing protocol
I. INTRODUCTION In wireless sensor network [1, 2] the sensors are deployed randomly or densely in deterministic or non-deterministic environment. The sensors node is small in nature with limited power. The energy is vital for wireless sensor network because it is not easy to recharge the nodes because it has been deployed in a hostile environment. So conservation of the energy efficiently is the primary motive in a wireless sensor networks. The sensor node consumes energy for sensing the activity, processing and transmitting. The energy required for transmitting the data is high compared to sensing and processing the data. The energy [3,4] spent for transmitting a single bit of data over 100 meters is equal to processing 3000 instructions. WSNs have been widely employed in several applications such as habitat monitoring, disaster supervising, defence application, and commercial application. An important challenge in designing of a wireless sensor network is very limited bandwidth and energy than in wired network environment. The innovative techniques are needed to eliminate energy inefficiency that shortens the lifetime of the network and efficient use of the limited bandwidth. So it is highly desirable to find new methods for energy efficient route
Velammal College of Engineering and Technology, Madurai
discovery and relaying of data from sensor node to the base station. DER is the significant energy aware routing mechanisms. In this scheme, the sensed data are transmitted to the base station based on assumed transmission range by the sensor nodes. In the transmission range, the energy level of the sensor nodes are aggregated together and mean measure is calculated. The target node is finding based on the criteria such that the node which one is nearest to the base station in the range. And the energy level of the node is greater than the threshold level. This scheme effectively reduces the delay and save network energy compared to some of the previous schemes. This protocol highly useful in military based application especially in cross border terrorisms for detecting an object .The detected events are it quickly to the base station for further processing or actions. The remainder of this paper is organized as follows. In section 2, some previous related works, which motivate the research, are briefly reviewed and discussed. The design of the proposed protocol DER is detailed in section 3. The algorithm specified in the section 4 and section5 shows the performance evaluation of the protocol. II. RELATED WORKS Low Energy Adaptive Clustering Hierarchy [5,6] (LEACH) is operated on two phases, the setup phase and steady phase. In the setup phase, the clusters are organized and cluster heads are selected. In the steady phase, the actual transfer to the base station takes place. LEACH randomly selects a few sensor nodes as cluster heads and rotates its role to evenly distribute the energy load among the sensor in the network life time. It holds some of the weaknesses are lots of cluster heads are selected and cluster heads are not selected in a distributed manner leads to unbalanced energy consumption of the nodes and prolonging lifetime of the network get shortly reduced. HPSD [7,8] (High power short distance protocol) is a nodebased protocol that selects route to the base station based on closest node that has the highest battery power relative to its surrounding neighbours. When an event occurs the source node sends the data to the base station using the shortest path. It consists of 10 randomly
Page 82
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
deployed sensor nodes. The base station is fixed and located in the left of middle corner. Initially, node N2 selects the next neighboured node (Node N7) that has the shortest distance to the base station, because all nodes have the same power level. Hence the path will be N7, N6, N9, N11, N13, and N15. Later on that node N7 changes its power level from high to medium due to participating in some routing task. And let node N3 has new data to send to the base station. Then this source node has three possible neighbours to use (Node5, Node7 and node4). That Node7 already participated in the previous routing and power level is medium. Therefore the source node will select either Node 5 or Node 4 because it posses high energy. The process is repeated until it reaches the base station. It posses some of the drawbacks are time delay and more number of nodes required for transmission and energy consumption for more node processing and additional overhead for sending control packets to its neighbours. N1
N4
N13
N11 N9 N6
N15
N7
the base station. Finally the base station calculates distance based on time and speed .The summarized data’s are broadcast it again to all nodes in the network by the base station. Equation for calculating the Distance: Distance= (Speed*Time) Where • Distance=Distance between the Node and the Base station. • Speed=Speed of wave in air. • Time=Time elapsed of a message to reach from the base station to the node. Constructing route table: After identification of node position, the base station broadcast the routing information. The routing table is maintained at each node for taking routing strategy. The routing table contains three fields such as Node-ID, energy level of the node, distance from the base station. The routing strategy is taken by analyzing the above fields in the routing table. The Table I shows the routing table maintained at each node in the sensor network.
N 2
TABLE I ROUTING TABLE
Node-ID N10 N1 4
N 12
N8
N5 N3
Base station
Sensor Field
Fig.1. HPSD Protocol operation
III. PROPOSED METHODOLOGY W. DER: Protocol Operation The Distance based Energy Aware Routing protocol operations are to increase the node lifetime by distributing energy load among sensor nodes. This protocol operated on three phases: 1) Initialization phase 2) Forwarding phase 3) Updating phase 1) Initialization Phase: Once the sensors are deployed in the environment, it is not so easy to re deploy the sensor in the same field. The energy is vital for sensor to survive in the environment. So the energy should be conserve efficiently in the sensing field. Identification of node position: The sensor deployed in the environment randomly. To know its position first the base station broadcast [7] the initialization packet to all the nodes in the sensor network. On receiving such packet, it calculates the time taken to reach every node in the network. Each node in the network stored its time. Then the node transmits its attributes (N_ID, Time and Energy) to
Velammal College of Engineering and Technology, Madurai
N1 N2 ,, ,, Nn BS
Distance
25Meters 37Meters ,, ,, NMeters -
Energy
1Joules 0.97Joules ,, ,, NJoules High Energy
2) Forwarding Phase: After the event triggered by the sensor nodes, the source node route the packet to the base station by considering the routing table. Successive Transmission: In successive transmission, the energy level of the sensor nodes gets changes in the sensor field. Here the source node takes the transmission range of 50m.The target node is predicted based on the criteria: a) The node that is nearest to the base station from the nodes with in the range. b) The Energy level of the node should be greater than the threshold level. Calculation of Threshold level: The energy level of the nodes with in the range is calculated on the basis of mean measures. n ⎛ EN (i ) + EN (i +1) + .......EN (n ) ⎞ ⎟⎟ Th = ∑ ⎜⎜ i =1 ⎝ No. of nodes in the range ⎠
Where, • •
Th is the Threshold level. EN(i) is the Energy level of node.
Page 83
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
If the energy level of the node greater than the threshold level, then the data to be transmitted to the node. Suppose if the energy level of the node lesser than the threshold level, then previous least distance node from the routing table should be taken and compared with its threshold level. If the condition is met, then make transmission to the node. During the successive transmission, the energy level of nodes gets changed. Suppose the Node N2 act as source node after sense the event. It takes the transmission range of 50 meters. The Node N4, N5 and N7 are located with in the range and mean threshold is calculated. Then Node N2 decided to take transmission to N7 that is least distance to the base station but the energy level is greater than the mean threshold level .so it act as target node. Next the N7 take transmission range of 50mts.The Node N6, N8, N11, and N10 are located with in the range and mean threshold is calculated. The Node N7 transmits the data to Node N10 that is least distance to the base station, but the energy level of the node lesser than the mean threshold. So it takes the previous least distance node N11 and compares energy level with the mean threshold. Suppose if the energy level is more than the mean threshold takes it as target node N11. The same procedure is repeated by N11 to transmit the data to Node 15 and finally to the base station. Sensor Field N14
N4 N11
N9
N15
N2
N 7
N10
N1
N 6 N8 N12 Base station
N5
N3
N13
Sensed Data
Node-ID
Target Node Energy
Fig.3. DER Updating Phase
IV. DER IN DETAIL The proposed algorithm is Distance based Energy aware routing Protocol. The procedure for forwarding phase is after the event triggered the source node takes the transmission range of 50 meters. From the source node, the transmission range of 50 meters should be taken. Within that range, find the node, which is nearest to the base station, and compare that node’s energy level with the threshold level. Suppose if energy of the node is lesser than the threshold level, then the previous least distance node from the routing table is taken and compared with the threshold level until the node energy is greater than the threshold level. If the energy of the node is greater than the threshold level, then take that node as target node. The process is repeated until the information reaches the base station. 1. Const Tr=50; 2. For each (Node n in Transmission range) 3. Loop beginning 4. If (n. Energy level is not same) //Node energy level is not uniform with in the range 4. Get_least_distance_node ( ); 5. End If 6. Loop End //If the node is least distance to the base station 7. Get_least_distance_node ( ) 8. Begin 9. For each (Node n in Transmission range) 10. Loop Beginning 11. Nn=Minimum_dist (); 12. Get_target_node (Nn); 13. Loop End 14. End
Fig.2. DER Forwarding phase
3) Updating Phase: After every transmission of information from source to base station, the energy level of the node in the sensor fields get changed. So it is essential for updating routing information at each node after every transmission of packets. Here the energy required for updating the routing information at each node lies on the hands of the base station. a) During transmission, the target nodes in every range not only send the sensed data in addition to that it sends energy level and Node-ID to the base station. b) The base station broadcast the messages to all nodes in the network. c) After receiving such message the nodes updates its routing information. The above process is repeated after every transmission to update the routing table at every node in the sensor network.
Velammal College of Engineering and Technology, Madurai
// If the Mean threshold greater than basic threshold and // Energy level of node greater than the mean threshold 15. Get_target_node (Nn) 16. Begin 17. For each (Node n in Transmission range) 18. Loop Beginning 19. EN= EN1+EN2… +ENn; 20. MTh=EN/n; 21. Const BTh=0.002j; 22. Loop End 23. If ((MTh >= BTh) && (ENn >= MTh)) 24. target node=Nn //Route to target node 25. Else 26. Get_next_minimum_dist (); 27. If (target. BS==true)
Page 84
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
28. //Route to base station 29. End If 30. End If 31. End //Finding the Minimum Distance Node 32. Minimum_dist () 33. Begin 34. For each (Node n in Transmission range) 35. For each (Node m in Transmission range) 36. Loop Beginning 37. If (N [n] < M [n]) 38. t=N[n] 38. N [n]=M [n] 39. M [n]=t 40. End If 41 Loop End 42. End 43. Get_next_minimum_dist () 44. Begin 45. For each (Node m in Transmission range) 46. Loop Beginning 47. Nn=M [n--] 48. Get_target_node (Nn) 49. Loop End 50. End
Y. B. Comparing HPSD with DER Table IV shows the comparison between the High Power Short Distance Protocol and Distance based Energy Aware Routing Protocol. TABLE III PARAMETERS USED IN THE SIMULATION
Parameters Number of nodes Network dimension Initial Energy Dead node Sense radius Deployment Base station location Data packet size Transmission Range Topology
Values 50 50*50 m2 1 Joule < 0.002 Joule 20 Meters Random Middle of left side 53 bytes (variable length) 50 Meters Static topology TABLE IV
S.No. 1)
HPSD Source node selects the target node from the nearest neighbor nodes.
2)
Energy level mention as High, Medium, and Low Energy level. More number of nodes required for transmitting a packet
Fig. 3. Pseudo code of DER
In the fig.3, the pseudo code describes the forwarding phase in DER. The Table II shows the variable used in the pseudo code. TABLE II VARIABLE MEANING
Name Tr Nn EN1... ENn EN BTh MTh BS
Usage Transmission range Minimum distance node Energy level of nodes in the range Total energy of the nodes in the range Base Threshold Mean Threshold Base Station
4)
5)
Time delay is more for transmitting a packet.
DER Select the target node based on the criteria: a) In the range, Node that nearest to base station. b) The node holds the energy level greater than the threshold level. Threshold Energy level is calculated based on mean measure. Less number of nodes required for transmission, due to assumed transmission range. Time delay is less for transmitting a packet.
HPSD VERSUS DER
X. A. Simulation environment and parameters Using NS2 [Network Simulator] to simulate DER Protocol from the following aspects: total energy consumption [9, 10, 11, and 12] of the node in the network and the Time delay in data transmission. The parameters used in the simulation are set as shown in Table III. It is assumed that the transmission frequency range is set dynamically in the applications
Z. C. Simulation Result Fig.4 shows the simulation results for total energy consumption for three compared schemes. The energy consumption of DER protocol is the least among three protocols (DER, LEACH and HPSD). As LEACH protocol adopts one- hop communication, the death of nodes will lead to the increasing of communication distance, which causes the energy consumption of LEACH protocol increase fast after 300s.The additional cost is required by transferring of control packets to the neighbourhood node leads to the energy cost of HPSD is more than the DER protocol. Figure.5 shows the simulation results of the time consumption for two schemes. In the HPSD Protocol, the time consumption for transmitting the packet is high compared to DER. Because the number of nodes required for transmitting a
Velammal College of Engineering and Technology, Madurai
Page 85
V. PERFORMANCE EVALUATION In this section, it describes the simulation environment and its parameter and simulation result.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
packet is more compared to DER. In DER protocol by taking the assumed transmission frequency the number of nodes required to reach the base station is shortly reduced lead to reduce the time delay in the sensor network.
Fig.4. Average Energy Consumption
Fig.5. Time Consumption
VI. CONCLUSION & FUTURE WORK The Distance based Energy Aware Routing protocol (DER) selects routes based on energy aware of the nodes in the network by mean measure approach. The distance based energy aware routing efficiently reduces the time and energy in WSN. The energy consumption of the node in the network is well balanced and prolonging the survival time of the wireless sensor network. The performance of DER is analyzed through simulation, the result show that DER can increase the lifetime of the network compared to other conventional protocol. For future work dynamic readjustment of the transmission frequency of the target node, and it leads to reduce the time and energy of the nodes in the sensor networks. In broadcasting of packets to the nodes, DER needs improvement in security aspects.
[22] Routing in wireless sensor networks. Mayank Saraogi. www.cs.utk.edu/~saraogi/594paper.pdf by M Saraogi – 2005. [23] A survey of wireless sensor network, EURASIP Journal on Wireless Communications and Networking 2005:5, pp 774–788. [24] Zheng, Yugui Qu, Baohua Zhao, Data Aware Clustering for Data Gathering in Wireless Sensor Networks, International Conference on Networks Security, Wireless Communications and Trusted Computing, 2009, vol. 1, pp.192-195, 2009. [25] Shemshaki, M.; Shahhoseini, H.S., Energy Efficient Clustering Algorithm with Direct Path Supports 2009 International Conference on Signal Processing Systems Volume, Issue, 15-17 May 2009 Page(s): 277 – 281. [26] Al-Karaki, J.N.; Al-Malkawi, I.T, On energy efficient routing for wireless sensor networks, International Conference on Volume, Issue, 16-18 Dec. 2008 Page(s): 465-469. [27] Hey, L.A.Power Aware Smart Routing in Wireless Sensor Networks Next Generation Internet Networks, 2008. NGI 2008 Date: 28-30, Apri2008, On page(s): 195-202. [28] Weizheng Ren, Lianming Xu, Zhongliang Deng, A LowPower Routing Protocol for Wireless Sensor Networks Based on Data Aggregation, International Symposium on Intelligent Information Technology Application Workshops, 2008 IEEE. [29] Fatma Bouabdallah, Nizar Bouabdallah, and Raouf Boutaba, On Balancing Energy Consumption in Wireless Sensor Networks, IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58,NO. 6, JULY 2009. [30] Hong Luo; Luo, J.; Liu, Y.; Das, S.K, Adaptive Data Fusion for Energy Efficient Routingin Wireless Sensor Networks, IEEE Transactionson Volume 55, Issue 10, Oct. 2006 Page(s): 1286 – 1299. [31] Asudeh, A, Movaghar, A., MEHR: Multi-Hop EnergyAware Hierarchical Routing for Wireless Sensor Network Onpage(s): 1-6. [32] Fariborzi, H.; Moghavvemi, M., EAMTR: energy aware multi-tree routing for wireless sensor networks Communications, IET, Volume 3, Issue 5, May 2009 Page(s): 733-739. [33] WeizhengRen, LianmingXu, ZhongliangDeng, A LowPower Routing Protocol for Wireless Sensor Networks Based on Data Aggregation Intelligent Information Technology Application Workshops, 2008. IITAW '08. Date: 21-22Dec.2008, Onpage(s): 661-664.
REFERENCES [20] Routing Techniques in Wireless Sensor Networks: A. Survey. Jamal N. Al-Karaki. Ahmed E. Kamal. – 2004. [21] Energy Conservation in Wireless Sensor Networks: a Survey. Giuseppe Anastasi, Marco Conti, Mario Di Francesco, Andrea Passarella, Volume 7, Issue 3 (May 2009) Year of Publication: 2009 Pages 537-568.
Velammal College of Engineering and Technology, Madurai
Page 86
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Study of Similarity Metrics for Genomic Data Using GO-Ontology Annalakshmi V#1, Priyadarshini R#2, Bhuvaneshwari V#3 #
School of Computer Science and Engineering, Bharathiar University Coimbatore-641 046, Tamil Nadu, India 1
[email protected] [email protected] 3 [email protected]
2
Abstract— Bioinformatics is the science of managing, mining and interpreting information from biological sequences and structures. The Gene Ontology (GO) is a major bioinformatics initiative to unify the representation of gene and gene product attributes across all species. Similarity is quantity that reflects the strength of relationship between two objects or two features. In this paper, a study on similarity metrics for genomic data is done. The modeling of data for measuring the measure is explained in detail. The comparison of traditional similarity and ontology based similarity measures are discussed and compared. Keywords— Bioinformatics, Gene Ontology, Similarity Measures, GO terms, genes.
I. INTRODUCTION Data mining, the extraction of hidden predictive information from large databases, is a powerful new technology with great potential to help companies focus on the most important information in their data warehouses [1]. Data mining techniques are the result of a long process of research and product development. Data mining is a component of a wider process called Knowledge discovery from databases. Bioinformatics is the science of organizing and analyzing biological data that involves collecting, manipulating, analyzing, and transmitting huge quantities of data. Bioinformatics and data mining provide exciting and challenging researches in several application areas especially in computer science. Bioinformatics is the science of managing, mining and interpreting information from biological sequences and structures [2]. Data are collected from genome analysis, protein analysis, microarray data and probes of gene function by genetic methods. The Gene Ontology (GO) is a major bioinformatics initiative to unify the representation of gene and gene product attributes across all species. The ontology covers three domains such as, cellular component, molecular function and biological process. GO is used to explore and analyze the data by using the GO term object. A gene is the basic unit of heredity in a living organism. All living things depend on genes. Genes hold the information to build and maintain an organism's cells and pass genetic traits to offspring. GO represents terms within a Directed Acyclic Graph (DAG) consisting of a number of terms, represented as nodes within the graph,
Velammal College of Engineering and Technology, Madurai
connected by relationships, represented as edges. Terms can have multiple parents, as well as multiple children along the \is-a' relationships, together with part-of relationships [6]. Similarity is defined as the ratio of the shared information between the objects to the total information about both objects [18]. Measure of semantic similarity for the knowledge component of bioinformatics resources should afford a biologist a new tool in their collection of analyses. The similarity has been computed by four general approaches: the set-based approach, the graph-based approach, the vectorbased approach and the term based approach. In the set-based approach an annotation is viewed as a ‘bag of words’. Two annotations are similar if there is a large overlap between their sets of terms. A graph-based approach views similarity as a graph-matching procedure. Vector-based methods embed annotations in a vector space where each possible term in the ontology forms a dimension. Term-based approaches compute similarity between individual terms and then combine these similarities to produce a measure of annotation similarity [19]. In this paper we have done a detailed study on the similarity metrics using GO ontology for genomic and proteomic databases using annotations. The paper is organized as follows. Section 2 provides the literature study of the various semantic similarity measures. Sections 3 describe similarity metrics using gene ontology. In section 4 provides modeling of dataset for similarity measure. The implemented results for the yeast dataset are analyzed and validated within the go ontology functionalities. The final section draws the conclusion of the paper. II. REVIEW OF LITERATURE The literature related to the various similarity measures based on the GO ontology is discussed. Semantic Similarity measures have been proposed for comparing concepts within ontology. Lord [6] has proposed a semantic similarity measure based on the information content of each term. This is defined as the number of times each term, or any child term, occurs in the corpus, which is expressed as a probability. Resnik [3], [4] developed a measure of semantic similarity for "is-a" ontology based on the information content of the lowest common ancestor (LCA) of two terms. The more frequently a term occurs, i.e., the higher its probability of occurring, the
Page 87
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
lower its information content [5]. One node based approach to determine the conceptual similarity is called the information content approach. If the LCA of two terms describes a generic concept, these terms are not very similar and this is reflected in a low information content of their LCA. This measure considers how specific the LCA of the two terms is but disregards how far away the two terms are from their LCA [5]. Lin [Lin, 1998] has presented a semantic similarities of two GO terms based on the annotation statistics of their common ancestor terms [9]. Jiang and Conrath, [1997] proposed a semantic similarity measure based on hybrid approach, i.e. it combines information content and conceptual distance [10]. Lord et al. Resnik, Lin, and Jiang's have provided similarity methods based on the information content of the GO terms. The two approaches target semantic similarity from quite different angles. Jiang and Conrath have proposed the edge-based distance method is more intuitive, while the node-based information content approach is more theoretically sound. Both have inherent strength and weakness [7]. Similarity measures that satisfy the only condition of reflexive, i.e., every object are most similar to it. Chabalier and Mosser have presented the vector Space model measures, they consider only m*n matrix representation, where m is the total number of genes in the corpus and n is the total number of GO terms. Each row in the matrix represents a gene vector of its annotations. Each vector is binary valued, with 1 represents the presence of the GO term in the gene’s annotation and 0 representing its absence [11]. The major contribution of this approach is the possibility of using a weighting scheme over the annotation.Chabalier and Mosser have proposed the Cosine similarity; the measure can be calculated by using the vector for each gene in the pair. This similarity first generates a weight, for each GO term based on the frequency of its occurrence in the corpus [20]. Huang et al developed the Kappa statistics are used to measure cooccurrence of annotation between gene pairs. Kappa statistics are used for continuous and non-categorical data. It can specifically detect the gene-gene functional relationships [12]. Lee et al has proposed the Term Overlap (TO) Measure score for two genes are calculated as the number of terms that occur in the intersection set of the two gene product annotation sets. The Normalized Term Overlap (NTO), in which the term overlap score is divided by the annotation set size for the gene with the lower number of GO annotations [13]. III.
SIMILARITY METRICS USING GENE ONTOLOGY
measures such as, Resnik, Lin’s, Jiang and Conrath measure. These measures are fully focused on the Information Content (IC). Group wise similarity is calculated among the group of genes. This method consists of the following measures such as, Cosine similarity, Kappa Statistics and Term Overlap (TO). These measures consider only matrix representation of gene products. AA. Node-based (Information Content) Approach One node based approach to determine the conceptual similarity is called the information content approach (Resnik 1992, 1995). This method contains the following measures such as, Resnik, Lin’s, Jiang and Conrath measure. 1) Lord Measure: Lord measure calculates the similarity between two terms by using only the IC of the t1: The information content of a GO term ti is:
Eq., (1) Where p(ti) is the probability of a term occurring in the corpus:
Eq., (2) Where the corpus is the set of annotations for all genes under consideration. "Root" represents one of the three root ontology terms and freq (root) is the number of times a gene is annotated with any term within that ontology. freq(ti) is given by:
Eq., (3) The Lord measure calculates the information content of term ti. 2) Resnik’s Measure: Resnik's measure calculates the similarity between two terms by using only the IC of the lowest common ancestor (LCA) shared between two terms t1 and t2:
Eq., (4)
Gene ontology is a collection of controlled vocabularies that describes the biology of a gene product. It consists of approximately 20,000 terms arranged in three independent ontology: Biological Process, Cellular Component, and Molecular Function. Similarity measures are used to find the similarity between genes. We can measure the similarity in two ways. There are Pair wise similarity and Group wise similarity. The pair wise similarity measures are calculated by using gene pairs. This method contains the following
Where LCA denotes the lowest common ancestor between ontological terms T1 and T2. This measure only accounts for the commonality between terms.
Velammal College of Engineering and Technology, Madurai
Page 88
3) Lin’s Measure: Lin's measure of similarity takes into consideration the IC values for each of terms t1 and t2 in addition to the LCA shared between the two terms and is defined as follows.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Eq., (5) Which has the advantage that it maps onto values on the interval [0, 1] unlike Resnik's measure which maps onto the interval [0, ∞]. Lin's measure also accounts for both the commonality and difference between terms. Resnik's measure does have the desirable property that terms close to the root of the ontology have a low similarity however. This is not the case for Lin's measure. 4) Jiang & Conrath Measure: Jiang and Conrath proposed an IC based semantic distance, which can be transformed into a similarity measure.
Eq., (6) This measure considered probability of term t1, t2 and LCA of the two terms. The highest score for Lin and Jiang is 1, and Resnik's measure has no upper bound. 5) Term Overlap: The term overlap score for two genes is then calculated as the number of terms that occur in the intersection set of the two gene product annotation sets.
Eq., (7) Normalized term overlap (NTO), in which the term overlap score is divided by the annotation set size for the gene with the lower number of GO annotations.
Eq., (8) Traditional cardinality-based similarity measures such as Jaccard and Dice [21] are computed similarly to NTO, but use the union or sum, respectively, of the two gene annotation set sizes as the normalizing factor. BB.
absence (resp. presence) of a term (along a particular dimension) in an annotation. This has the advantage that standard clustering techniques on vector spaces such as kmeans can be applied to group similar terms. What is required is a means of measuring the size of vectors. This can be achieved by embedding terms in a metric space (usually Euclidean). The most common method of measuring similarity between vectors of terms is the cosine similarity
Eq., (9) Where vi represents a vector of terms constructed from an annotation (group of terms) Gi. |·| corresponds to the size of the vector and • corresponds to the dot product between two vectors. The source of descriptiveness, commonality and difference is the same as the situation for set-based approaches. CC.
Edge-based (Distance) Approach
The edge based approach is a more natural and direct way of evaluating semantic similarity in a Go Ontology. It estimates the distance (e.g. edge length) between nodes which correspond to the terms/genes being compared. Given the multidimensional concept space, the conceptual distance can conveniently be measured by the geometric distance between the nodes representing the concepts. Obviously, the shorter the path from one node to the other, the more similar they are. In a more realistic scenario, the distances between any two adjacent nodes are not necessarily equal. It is therefore necessary to consider that the edge connecting the two nodes should be weighted. Most of these are typically related to the structural characteristics of a hierarchical network. Some conceivable features are: local network density (the number of child links that span out from a parent node), depth of a node in the hierarchy, type of link, and finally, perhaps the most important of all, the strength of an edge link. This method contains the following measures. Sussna (1993) considered the first three factors in the edge weight determination scheme. The weight between two nodes c1 and c2 is calculated as follows:
Eq., (10) Given
Vector-based Approach
Vector-based methods embed ontological terms in a vector space by associating each term with a dimension. Usually a vector is binary consisting of 0's and 1's where 0 denotes the
Velammal College of Engineering and Technology, Madurai
Eq., (11)
Page 89
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
where ®r is a relation of type r, ®r' is its reverse, d is the depth of the deeper one of the two, max and min are the maximum and minimum weights possible for a specific relation type r respectively, and nr(x) is the number of relations of type r leaving node x. In determining the overall edge based similarity, most methods just simply sum up all the edge weights along the shortest path. To convert the distance measure to a similarity measure, one may simply subtract the path length from the maximum possible path length (Resnik 1995):
Eq., (12) Where dmax is the maximum depth of the taxonomy, and the len function is the simple calculation of the shortest path length (i.e. weight = 1 for each edge). DD. Set based Approach Set based methods for measuring the similarity of annotations are based on the Tversky ratio model of similarity, which is a general model of distance between sets of terms. It is represented by the formula
EE. Graph based Approach Ontology is a directed, acyclic graph (DAG) whose edges correspond to relationships between terms. Thus it is natural to compare terms using methods for graph matching and graph similarity. We may consider the similarity between annotations in terms of the sub-graph that connects terms within each annotation. Annotation similarity is then measured in terms of similarity between two graphs. Graph matching has only a weak correlation with similarity between terms. It is also computationally expensive to compute, graph matching being an NP-complete problem on general graphs. 1) Improving Similarity Measures by Weighting Terms: Set, vector and graph-based methods for measuring similarity between annotations can be improved by introducing a weighting function into the similarity measure. For example, the weighted Jaccard distance can be formulated as:
Eq., (16) Where, as before, G1 and G2 are annotations or sets of terms describing data (e.g. a gene product), Tx is the xth term from a set of terms and m(Tx) denotes the weight of Tx.
Eq., (13) Where G1 and G2 are sets of terms or annotations from the same ontology and ƒ is an additive function on sets (usually set cardinality). For α=β=1 we get the Jaccard distance between sets:
Eq., (14) And for α=β=1/2 we get the Dice distance between sets
Instance-Based Weights: One approach to assigning weight to an ontological term is to measure how informative a term is in describing data. A method of measuring information is to analyze a term's use in a corpus against the general use of ontological terms in the same corpus. Information is measured using the surprisal function:
Eq., (17) Where p(Ti) corresponds to the probability of a term Ti or its taxonomic descendants occurring in a corpus. Other Weighting Approaches: Other measures of information can be used not necessarily relying on corpus data. One measure [14] relies on the assumption that how the ontology is constructed is semantically meaningful:
Eq., (15) In this situation the source of descriptiveness of an annotation is its set of terms. Each term and its set of associated instances are considered independent of other terms. The commonality and difference between annotations is modeled as set intersection and difference of sets of terms respectively. Setbased approaches return a similarity of zero if they do not share common terms ignoring the fact that terms may be closely related. Because of the atomic nature of terms in the set-based approach the monotonicity property does not apply .
Where desc(Ti) returns the number of descendants of term Ti and numTerms refers to the total number of terms in the ontology.
Velammal College of Engineering and Technology, Madurai
Page 90
Eq., (18)
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
FF. Term-based Approaches In term-based approaches similarity between pairs of terms from each annotation are computed. These weightings are then combined in order to characterize the similarity between annotations as a whole. There are several ways to combine similarities of pairs of terms such as the min, max or average operations. Term-based approaches depend on a function s(Ti, Tj) where Ti and Tj are terms from two annotations G1 and G2 respectively. s(Ti, Tj) provides a measure of distance/similarity between these two terms. Once distances has been measured between all possible pairs of terms they are then aggregated using an operation such as max or the average of all distances. For example:
Graphical Measures of Term Similarity: The simplest approach to measuring similarity between ontological terms using the graph structure is to measure the shortest path distance between terms in the graph [15, 16]. A more refined use of graph distance as a basis for a measure of term similarity is found in the Wu-Palmer measure of similarity [17]. It uses the idea that the distance from the root to the lowest common taxonomic ancestor (LCTA) measures the commonality between two terms while the sum of the distance between the LCTA and each term measures the difference between two terms. Combining these aspects results in the formula:
Eq., (20)
Eq., (19) More sophisticated term based approaches combine multiple measures of term similarity and aggregate similarity values using more complex functions.
Where T1 and T2 are the two terms being compared, Tlcta is the term that corresponds to the lowest common taxonomic ancestor between T1 and T2. Troot denotes to root node of the ontology (assuming that the ontology has only one root). dist(Ti, Tj) denotes the graph distance between terms Ti and Tj. The 2 * dist(Tlcta, Troot) component of the denominator serves to normalize the measure.
The various approaches with author reference are given in Table 1. TABLE I. COMPARISON OF VARIOUS APPROACHES AND MEASURES
Approach Node-based Approach
Measures The similarity value is defined as the information content value of specific gene pairs.
Author Resnik, Lin’s, Jiang and Conrath. Meeta Mistry and Paul Pavlidis
Edge-based Approach
This method estimates the distance between nodes which correspond to the terms/genes being compared. This method for measuring the similarity of annotations is based on the Tversky ratio model of similarity, which is a general model of distance between sets of terms.
Jiang JJ, Conrath DW
Set-Based Approach
Graph-Based Approach
Term-Based Approach
Vector-Based Approach
This method is measuring in terms of similarity between two graphs, and then measuring similarity between annotations can be improved by introducing a weighting function into the similarity measure This method is measuring similarity of between pairs of terms from each annotation are computed. This method is measuring similarity of matrix representation of genes.
Velammal College of Engineering and Technology, Madurai
Huang da W, Sherman BT, Tan Q, Collins JR, Alvord WG, Lempicki RA, Jeffrey Hau, William Lee, Brendan Sheehan Brendan Sheehan, Aaron Quigley, Benoit Gaudin and Simon Dobson, Lee JH, Kim MH, Lee YJ Rada R, Mili H, Bicknell E, Bletner, Wu-Palmer, Lee JH, Kim MH, Lee YJ Resnik P , Lord PW, Stevens RD ,Meeta Mistry and Paul Pavlidis, Brendan Sheehan
Ref No [6], [4], [3], [13],[19], [21],[22] [7],[8], [10] [12],[18], [19]
[13],[16], [19] [15],[16], [17] [4],[6], [13],[19]
Page 91
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IV.
MODELING DATASET FOR SIMILARITY MEASURE
GO annotation site contains various data sets. Yeast data set is one of the types of data set. We can download the yeast data set form the Go annotation website. Yeast microarray data contains about 6400 Genes (e.g. YALO51W, YALO54C) with their corresponding Yeast values (0.1650, 0.2720). The file “yeastgenes.sgd” was obtained from the GO annotation site. SGDann is a master structure of Yeast microarray data. It contains the parameters namely SGDaspect, SGDgenes, SGDgo. SGDaspect contains the corresponding functionality for the SGDgenes. SGDgenes contains the whole genes. Similarly SGDgo contains the corresponding Goid for SGDgenes. The entire microarray data are mapped with this structure (SGDgenes) and the microarray data is splitted based on the functionality of genes using GO ontology. For each gene the corresponding functionality of genes are found (Biological Process, Cellular Component, and Molecular Function). All the genes with same functionality are grouped for the corresponding Yeast values from the microarray dataset. Yeast data set contains 6400 genes. The sample genes for yeast data set can be shown in Fig. 1.
The similarity measure can be calculated based on the gene pair. So we are taking the gene from the yeast micro array data. Consider one gene name is ‘YKL050C’. This gene contains the following Goids (3674, 5575, 8150). This can be mapped from the entire SGDgenes and find out the goid values. The gene and goid can be shown in Fig. 3.
Fig. 3 Gene for ‘YKL050C’ & Goid values
Each descendant can have their own genes. The genes can be finding out by using the goid values. Descendants contain the corresponding goid for the genes. The respective genes can be shown in Fig. 4.
Fig. 1 Genes from the yeast data set Fig. 4 Genes for the descendant’s goid
SGDann structure is the entire yeast micro array data set. It contains the attributes such as, Database, DB_Object_ID, DB_Object_Symbol, Goid, Aspect, and DB_Object_Type. It can be shown in Fig. 2.
Descendant means children for the respective goid. We can easily get the children (descendants) from the GO ontology using the get descendant method. For example consider one goid ‘3674’. It contains 8675 descendants. The sample descendants can be shown in Fig. 5.
Fig. 2 SGDann Data Structure
Velammal College of Engineering and Technology, Madurai
Page 92
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
methods. Using this method we can’t get the accurate result for the semantic similarity for analyzing gene functionalities for genes. The semantic measures calculated using traditional approaches are shown in Table 2. TABLE II. RESULTS OF TRADITIONAL METHODS
Genes YKL050C
Fig. 5 Descendant for the goid ‘3674’
Similarly we can get the genes from the yeast dataset and can be arranged in the structure. For example we consider the six genes namely YKL050C, YGL114W, YBR241C, YIL166C, YBL095W, YOR390W. And the corresponding goid, descendants can be shown in Fig. 6.
Fig. 6 Multiple Genes & descendants structure
YGL114W
YBR241C
Euclidean Jaccard Cosine Euclidean Jaccard Cosine Euclidean Jaccard Cosine
YKL050C
YGL114W
YBR241C
0 0 0 21.1896 0.7143 0.0061 11.0905 0.7143 0.0020
21.1896 0.7143 0.0061 0 0 0 21.9089 0.8571 0.0065
11.0905 0.7143 0.0020 21.9089 0.8571 0.0065 0 0 0
For the above we consider the genes for ‘YKL050C’, ‘YGL114W’, and ‘YBR241C’ for traditional method. In this table we calculated the values for the Euclidean, Jaccard and Cosine similarity. The GO ontology based semantic similarity measures are calculated using the discussed approaches above. These measures identify the relation among genes based on their children and parents. The Table 3 provides with the result of the approaches based on ontology. In table3 we consider the genes for ‘YKL050C’, ‘YGL114W’, and ‘YBR241C’ for calculating the semantic similarity measures based on the genes functionality. In this table we are used Lin, Jiang and Conrath measures. TABLE III. RESULTS FOR GO ONTOLOGY FUNCTIONALITY
Genes YKL050C Lin Jiang YGL114W Lin Jiang YBR241C Lin Jiang
YKL050C 0 0 1.5728 0.9039 1.5749 0.9037
V.
Fig. 7 Graphical representation for ‘YKLO50C’ gene with 3 Go Terms
A. Experimental Results The traditional method for semantic similarity measure is calculated based on the content comparison. The basic types of traditional methods are Euclidean, Jaccard and Cosine
Velammal College of Engineering and Technology, Madurai
YGL114W 1.5728 0.9039 0 0 1.6721 0.8248
YBR241C 1.5749 0.9037 3.6721 0.8248 0 0
CONCLUSION
In this paper we have done a detailed study of the semantic similarity measures for finding the relationships among genes. The modeling of the dataset for the measure calculation is provided in detail. The results on implementation we have found that gene ontology based similarity measures finds the accurate association of genes based on Go Ontology. And also we compared the traditional measures and gene ontology based similarity measures. We conclude from our study that the gene ontology based semantic measures can be used for accessing genes for various factors than traditional based measures.
Page 93
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
REFERENCES [34] Arun. K. Pujari, Data Mining Techniques, Universities press (India) Limited 2001, ISBN-81-7371-380-4. [35] Daxin Jiang, Chun Tang, and Aidong Zhang, “Cluster Analysis for Gene Expression Data: A Survey, IEEE Transactions on knowledge AND Data Engineering”, Vol 16, No. 11, November 2004. [36] Resnik P: “Using Information Content to Evaluate Semantic Similarity in a Taxonomy” Proc 14th Int'l Joint Conf Artificial Intelligence, 1995, Vol 1, p.448-453. [37] Resnik P: “Semantic Similarity in a Taxonomy: An InformationBased Measure and its Application to Problems of Ambiguity in Natural Language” J Artif Intell Res 1999, 11:95-130. [38] Schlicker A, Domingues FS, Rahnenfuhrer J, Lengauer T: “A new measure for functional similarity of gene products based on Gene Ontology” BMC Bioinformatics 2006, 7:302. [39] Lord PW, Stevens RD, Brass A, Goble CA: “Semantic similarity measures as tools for exploring the gene ontology” Pac Symp Biocomputing 2003:601-612. [40] Jiang JJ, Conrath DW: “Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy” ROCLING X: 1997; Taiwan 1997. [41] Tatusova TA, Madden TL: “BLAST 2 Sequences, a new tool for comparing protein and nucleotide sequences” FEMS Microbiol Lett 1999, 174(2):247-250. [42] Guangchuang Yu: “GO-terms Semantic Similarity Measures” October 28, 2009. [43] Francisco M. Couto, Mário J. Silva, Pedro Coutinho: “Implementation of a Functional Semantic Similarity Measure between GeneProducts” DI–FCUL TR–03–29, 2003. [44] Chabalier J, Mosser J, Burgun A: “A transversal approach to predict gene product networks from ontology-based similarity” BMC Bioinformatics 2007, 8:235. [45] Huang da W, Sherman BT, Tan Q, Collins JR, Alvord WG, Roayaei J, Stephens R, Baseler MW, Lane HC, Lempicki RA: “The DAVID Gene Functional Classification Tool: a novel biological module-centric algorithm to functionally analyze large gene lists” Genome Biol 2007, 8(9):R183. [46] Meeta Mistry1 and Paul Pavlidis*2: “Gene Ontology term overlap as a measure of gene functional Similarity” Bioinformatics 2008, 9:327. [47] Veale N, Seco JHT: “An Intrinsic Information Content Metric for Semantic Similarity in WordNet” ECAI 2004 2004:1089-1090. [48] Rada R, Mili H, Bicknell E, Bletner M: “Development and Application of a Metric on Semantic Nets” IEEE Transactions on Systems, Man, and Cybernetics 1989, 19:17-30. [49] Lee JH, Kim MH, Lee YJ: “Information Retrieval Based on Conceptual Distance in IS-A Hierarchies” Journal of Documentation 1993, 49:188-207. [50] Wu Z, Palmer M: “Verb semantics and lexical selection” In 32nd. Annual Meeting of the Association for Computational Linguistics New Mexico State University, Las Cruces, New Mexico; 1994:133-138. [51] Jeffrey Hau, William Lee, John Darlington: “A Semantic Similarity Measure for Semantic Web Services” Imperial College London, 180 Queens Gate, London, UK. [52] Brendan Sheehan*, Aaron Quigley, Benoit Gaudin and Simon Dobson: “A relation based measure of semantic similarity for Gene Ontology annotations” BMC Bioinformatics 2008. Vol 9,9:468. [53] Chabalier J, Mosser J, Burgun A: “A transversal approach to predict gene product networks from ontology-based similarity” BMC Bioinformatics 2007, 8:235. [54] Popescu M, Keller JM, Mitchell JA: “Fuzzy measures on the Gene Ontology for gene product similarity” IEEE/ACM Trans Comput Biol Bioinform 2006, 3(3):263-274. [55] Lin D: “An information-theoretic definition of similarity” In 15th International Conf on Machine Learning Morgan Kaufmann, San Francisco, CA; 1998:296-304
Velammal College of Engineering and Technology, Madurai
Page 94
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hybrid PSO based neural network classifier and decision tree for brain MRI mining #
Dr.V.Saravanan#1 , T.R.Sivapriya*2 Dr.N.G.P Institute of Technology, Coimbatore, India *
[email protected]
Lady Doak College, Madurai, India [email protected]
Abstract Artificial neural networks have been applied in a variety of realworld scenarios with remarkable success. In this paper, a hybrid PSO based back propagation neural network for classifying brain MRI is proposed. The results show that there is a marked difference while training BPN with PSO. Also a customized PSO is used to train the BPN which again results in better performance compared to the conventional PSO used in training. The Decision tree extracts rules from the trained network that aids in medical diagnosis. The neural model based on Particle Swarm Optimisation was trained with 150 samples (including all patients with mild and severe dementia). Additional hundred samples have been used for validation. The proposed algorithm outperforms the result of conventional training algorithms and is found to have 95% sensitivity and 96% accuracy. The samples were tested with a radiologist and psychiatrist by means of blind folded study. When compared with the experts, the algorithm achieved good accuracy with higher rate of reliability for the assessment of mild and severe dementia. Keywords --- Image mining, Neural networks, PSO, Decision tree
1.INTRODUCTION There has been a growing number of research applying ANNs for classification in a variety of real world applications. In such applications, it is desirable to have a set of rules that explains the classification process of a trained network The classification concept represented as rules is certainly more comprehensible to a human user than a collection of ANNs weights . This paper proposes a fast hybrid PSO based BPN classifier, whose results are mined by decision tree for the diagnosis of dementia. A neural network, by definition, is a massively parallel distributed processor made up of simple processing units, each of which has a natural propensity for storing experiential knowledge and making the knowledge available for use [11]. Neural networks are fault tolerant and are good at pattern recognition and trend prediction. In the case of limited knowledge, artificial neutral network algorithms are frequently used to construct a model of the data.
Velammal College of Engineering and Technology, Madurai
A. Artificial Neural Networks In the last few years artificial neural networks have been applied successfully to a variety of real-world problems. For example, neural networks have been successfully applied in the area of speech generation and recognition, vision and robotics, handwritten character recognition , medical diagnostics [9], and game playing. Neural network method is used for classification, clustering, feature mining, prediction and pattern recognition. It imitates the neurons structure of animals, bases on the M-P model and Hebbian learning rule, so in essence it is a distributed matrix structure. Through training data mining, the neural network method gradually calculates (including repeated iteration or cumulative calculation) the weights the neural network connected. Knowledge can be more easily recognized when presented in the form of images. For example, geophysical and environmental data from satellite photos, Web pages containing images, medical imaging including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), is sources of useful information used in our daily life. B. The Artificial Neural Network Models for Data Mining The center piece of the analytic plan is modeling ANN architectures. Artificial neural network (ANN) modeling has been used for various tasks, including speech recognition, stock market prediction, mine detection, cancerous cell identification, and handwriting recognition. We use the ANN as a computational modeling tool, the center piece of the datamining approaches taken in this application. In practice, however, reports of ANN’s ability to surpass linear models for classification and prediction problems have been inconsistent (Doig et al., 1993; Baxt, 1995; Dybowski et al., 1996; Anderer et al., 1994; Duhet al., 1998), depending on data complexity, evaluation measures, and the user’s familiarity with the data as well as modeling.
Page 95
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
II. BACK PROPAGATION NEURAL NETWORKS The backpropagation algorithm for estimating parameters in neural networks has been the most popular in the medical literature [Reggia, 1993. One of the problems encountered by researchers utilizing the backpropagation algorithm is that low-frequency patterns (or rare patterns) may take a long training time to be recognized, because frequent patterns “dominate” the error.This problem is overcome by faster training using PSO. A hidden layer is used for ANN’s internal transformation: hidden neurons transform the “signals” from the input layer and generate signals to be sent to the next layer, so that MLP is often called the “feedforward” neural network. The input layer consists of input variables plus an input bias, which adjusts the weighted mean of the input variables. All input variables are regressed on the next layer, in the same fashion as multiple regression estimation.
Fig. 1 Backpropagation Network
A. BPN architecture design BPN architecture is selected based on testing and training of 7 different models.The steps are as follows: 1. Create a customized bpn model n1. 2. Train and test with new data 3. Check for the efficiency and ability to recognize new patterns 4. If the model n1 has comparatively higher efficiency and detects new patterns ,it is selected as the ideal model for diagnosis. III. PARTICLE SWARM OPTIMISATION Particle Swarm Optimisers (PSO) are a new trend in evolutionary algorithms, being inspired in group dynamics and its synergy. PSO had their origins in computer simulations of the coordinated motion in flocks of birds or schools of fish. As these animals wander through a three dimensional space, searching for food or evading predators, these algorithms make use of particles moving in an n-dimensional space to search for solutions for a variable function optimization problem. In PSO individuals are called particles and the population is called a swarm [1]. PSO are inspired in the intelligent behaviour of beings as part of an experience sharing community as opposed to an isolated individual reactive response to the environment
Velammal College of Engineering and Technology, Madurai
A. Training the BPN with PSO A Neural Network’s effectiveness is in its weights (assuming the topology is correct) Other methods exist to train NNs, but they can be slow and/or difficult to implement (GA,Simulated Annealing. Every edge (or weight) in the network is an element in our particle. The number of weights in the network equalizes the number of dimensions of the particles in PSO. The fitness value is how well the NN performs on the test data. The algorithms is run for a given number of iterations or until a minimum error is reached as follows: 1. Initialize a population size, positions and velocities of agents, and the number of weights and biases. 2. The current best fitness achieved by particle p is set as pbest. The pbest with best value is set as gbest and this value is stored. 3. Evaluate the desired optimization fitness function fp for each particle as the Mean Square Error (MSE) over a given data set. 4. Compare the evaluated fitness value fp of each particle withits pbest value. If fp < pbest then pbest = fp and bestxp = xp, xp is the current coordinates of particle p, and bestxp is thecoordinates corresponding to particle p’s best fitness so far. 5. The objective function value is calculated for new positions of each particle. If a better position is achieved by an agent, pbest value is replaced by the current value. if fp < gbest then gbest = p, where gbest is the particle having the overall best fitness over all particles in the swarm. 6. Change the velocity and location of the particle based on random number assigned. vid = vid + ϕ1*rnd()*(pid-xid) + ϕ2*rnd()*(pgd-xid); xid = xid + vid; • Where i is the particle,ϕ1,ϕ2 are learning rates governing the cognition and social components • g represents the index of the particle with the best pfitness, and d is the dth dimension. IV. DECISION TREE Decision tree induction is one of the simplest, and yet most successful forms of learning algorithm. It serves as a good introduction to the area of inductive learning, and is easy to implement. A decision tree takes as input an object or situation described by a set of attributes and returns a decision-the predicted output value for the input. The input attributes can be discrete or continuous. For now, we assume discrete inputs. The output value can also be discrete or continuous; learning a discrete-valued function is called classification learning; learning a continuous function is called regression. A. Evaluation of Rule set The constructed rule set is evaluated against the instance test set. Each instance is sequentially tested with rules in the rule set, until it matches one rule antecedent. The instance class is then compared to the matching rule predicted class. When all
Page 96
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
instances have been tested, the reliability of the rule set is computed regarding correctly classified instances and presented in a percentage value. CLASSIFY MRI
CLASS I
CLASS II
4. Test the network 5.Train the same BPN network using PSO 6. Test the network 7.Compare the performance 8. Apply data mining algorithm to produce object association rules from the weight vector of the trained Hybrid PSO based BPN. READ MRI IMAGE
PREPROCESSING
NORMAL
HIGLHLY DEMENTED
LESS DEMENTED
Fig 2. Decision Tree for Mining V. MRI DATASET OASIS provides brain imaging data that are freely available for distribution and data analysis. This data set consists of a cross-sectional collection of 416 subjects covering the adult life span aged 18 to 96 including individuals with early-stage Alzheimer’s Disease (AD). For each subject, 3 or 4 individual T1-weighted MRI scans obtained within a single imaging session are included. The subjects are all right-handed and include both men and women. 100 of the included subjects over the age of 60 have been diagnosed with very mild to mild AD. Additionally, for 20 of the non-demented subjects, images from a subsequent scan session after a short delay (less than 90 days) are also included as a means of assessing acquisition reliability in the proposed study.
FEATURE EXTRACTION
TRAINNG IN ANN
TESTING
CLASSIFY CC,VT,
EXTRACT RULES
Fig 2. System Overview
For each subject, a number of images are taken for analysis, including: 1) images corresponding to multiple repetitions of the same structural protocol within a single session to increase signalto-noise, 2) an average image that is a motion-corrected co registered average of all available data, 3) a gain-field corrected atlas-registered image to the 1988 atlas space of Talairach and Tournoux (Buckner et al., 2004), 4) a masked version of the atlas-registered image in which all non-brain voxels have been assigned an intensity value of 0, and 5) a gray/white/CSF segmented image (Zhang et al., 2001). VI. HYBRID PSO BASED BPN 1. Feature extraction. Segment images into regions identifiable by region descriptors This step is also called segmentation. a. extract cortex b. extract hippocampus c. ventricles d. basal ganglia 2. Feature reduction by PCA 3. Train conventional BPN with features to generate desired output
Fig 3. Brain MRI features Initially the conventional Particle Swarm Optimisation technique with 25 particles is used to train the backpropagation algorithm. Although the convergence was quick, performance improved further when the number of particles is reduced to 10. A. Training
Velammal College of Engineering and Technology, Madurai
Page 97
A multilayered back-propagation neural network was used (37 inputs from each of the 150 adolescents, comprising the input patterns, and the two binary outputs). The network was exposed to the data, and the parameters (weights and biases)
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
were adjusted to minimize error, using a back-propagation training algorithm. . The input layer has 10 neurons, where each neuron represents reduced pixel group. The number of neurons in the hidden layer is calculated based on the following formula: N3 =
((2/3)*(N1))+N2
Rules : IF CC=+++ AND HIP=+++ AND BG=+++ AND VT=++++ THEN DEMENTIA= HIGH IF CC= +++ AND HIP=+ AND BG=+ AND VT=+++ THEN DEMENTIA= MODERATE
N1 represents number of nodes in the input layer; N2 represents number of nodes in the output layer; N3 represents number of nodes in the hidden layer.
IF CC=+ AND HIP= -- AND BG = -- AND VT=+ THEN DEMENTIA= MILD
B. Validation
IF CC= -- AND HIP= -- AND BG = -- AND VT=-THEN DEMENTIA= NO TABLE III RESULTS OF CONVERGENCE
PSO-BPN
BPN
Iterati ons 175
Iterati ons 175
MRI
The ability of the neural network model (a 10 X 7 X 1 structure was used) to perform the classification was examined by setting aside 20% of the patterns (or observations) as validation (or testing) data. In this crossvalidation approach, the training involved repeatedly exposing the network to the remaining 80% of the patterns (training data) for several epochs, where an epoch is one complete cycle through the network for all cases. (Data were normalized before training.) A network trained in this manner is considered generalizable, in the sense that it can be used to make predictions. VII. RESULTS
200
TABLE I COMPARISON OF CONVERGENCE
225
575
BPN
Sensitivity %
85
HYBRID PSOBPN 92
Accuracy % 83 90 Specificity % 87 93 CC- Cerebral cortex HIP= Hippocampus VT – Ventricle BG – Basal Ganglia +++ Æ high increase ;+ Æ less increase -- Æ no increase
275 300 175 200
TABLE 2 COMPARISON OF EFFICIENCY
MEASURE
250 NORMAL
ITERATIONS >2000 800
CUSTOMISED PSOBPN 95 96 97
Velammal College of Engineering and Technology, Madurai
HIGHLY DEMENTED
TECHNIQUE BPN BPN-PSO WITH 25 PARTICLES BPN –PSO WITH 10 PARTICLES
225 250 275 300
GBest 0.247863299824 5176 0.247863290721 40684 0.247863290493 80179 0.247863290484 66602 0.247863290484 66071 0.247863290484 66071 0.247863299824 5176
200
0.247863290721 40684 0.247863290493 80179 0.247863290484 66602 0.247863290484 66071 0.247863290484 66071
200
225 250 275 300 175
225 250 275 300
MSE 0.003164 58/0 0.003122 28/0 0.003033 84/0 0.002969 5 0.002870 65 0.002763 27/0 0.003164 58/0 0.003122 28/0 0.003033 84/0 0.002969 5 0.002870 65 0.002763 27/0
Page 98
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the weights of the BPN correlated with the diagnosis of the experts.
All Data, Blue = known, Red = net 1 0.9 0.8
REFERENCES
0.7 0.6 out
0.5 0.4 0.3 0.2 0.1 0 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
in
Fig. 3. Trained PSO-BPN
Fig 4. Converging speed of conventional BPN
HYBRID BPN VS BPN 35 30 TIME
25 20 15 10 5 0 1
2
3
4
5
6
7
8
9
10
NO. OF MRI IMAGES
Fig 5. Comparison of converging speed of Hybrid PSO-BPN and conventional BPN
VII. CONCLUSION The hybrid PSO based BPN effectively classifies the brain MRI images of dementia patients. It appears to prove for highly reliable for use by practitioners in the medical field. Mining large database of images poses severe challenges. However, the hybrid classifier is being offered s to overcome difficulties in diagnosis and treatment. The rules mined from
Velammal College of Engineering and Technology, Madurai
[1] Ashish Darbari, “Rule Extraction from Trained ANN: A Survey,” Technical Report, Department of Computer Science, Dresden University of Technology, Dresden, Germany, 2000. [2] Heon Gyu Lee, Ki Yong Noh, Keun Ho Ryu, “Mining Biosignal Data: Coronary Artery Disease Diagnosis using Linear and Nonlinear Features of HRV,” LNAI 4819: Emerging Technologies in Knowledge Discovery and Data Mining, pp. 56-66, May 2007. [3] Cristianini, N., Shawe-Taylor, J. “An introduction to Support Vector Machines”, Cambridge University Press, Cambridge, 2000. [4] Li, W., Han, J., Pei, J., “CMAR: Accurate and Efficient Classification Based on Multiple Association Rules”, In: Proc. of 2001 Interna’l Conference on Data Mining. 2001. [5] Parpinelli, R., Lopes, H. and Freitas, A., “An Ant Colony Algorithm for Classification Rule Discovery”, Idea Group, 2002. [6] F. H. Saad, B. de la Iglesia, and G. D. Bell, “A Comparison of Two Document Clustering Approaches for Clustering Medical Documents”, Proceedings of the 2006 International Conference on Data Mining (DMIN06), 2006. [7] Gerhard Münz, Sa Li, and Georg Carle, “Traffic anomaly detection using k-means clustering”, In Proc. of Leistungs-, Zuverlässigkeits- und erlässlichkeitsbewertung von Kommunikationsnetzen und Verteilten Systemen, 4. GI/ITG-Workshop MMBnet 2007, Hamburg, Germany, September 2007. [8] R. Reed, “Pruning algorithms-A survey,” IEEE Transactions on Neural Networks, vol. 4, pp. 740-747, 1993. [9] Han Jiawei, Micheline Kamber, “Data Mining: Concepts and Techniques,” Morgan Kaufmann Publisher: CA, 2001 [9] Evangelou, I. E.: Automatic Segmentation of Magnetic Resonance Images (MRIs) of the Human Brain using Self-Organizing Maps (SOMs). MS Thesis, Clarkson University, Potsdam, NY, USA (1999). [10] Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., and R. Uthurusamy, R., (Eds.). Advances in Knowledge Discovery and Data Mining. AAAI Press, The MIT Press, CA, USA (1996). [11] J. S. D. Bonet. Image preprocessing for rapid selection in “Pay attention mode”. MIT, 2000.. [12] Baladi, P. & Hornik K., “Neural networks and principal component analysis: learning from examples and local minima “, Neural Networks, 2,5358.1989. [13] Hsinchun Chen, Sherrilynne S. Fuller, Carol Friedman, and William Hersh, "Knowledge Management, Data Mining, and Text Mining In Medical Informatics", Chapter 1, eds. Medical Informatics: Knowledge Management And Data Mining In Biomedicine, New York, Springer, pp. 3-34, 2005. [14] S Stilou, P D Bamidis, N Maglaveras, C Pappas, “Mining association rules from clinical databases: an intelligent diagnostic process in healthcare”, Stud Health Technol Inform 84: Pt2. 1399-1403, 2001. [15] T Syeda-Mahmood, F Wang, D Beymer, A Amir, M Richmond, SN Hashmi, "AALIM:Multimodal Mining for Cardiac Decision Support", Computers in Cardiology, pp. 209-212,Sept. 30 2007-Oct. 3 2007. [16] Anamika Gupta, Naveen Kumar, and Vasudha Bhatnagar, "Analysis of Medical Data using Data Mining and Formal Concept Analysis", Proceedings Of World Academy Of Science, Engineering And Technology,Vol. 6, June 2005. [17] Sellappan Palaniappan, Rafiah Awang, "Intelligent Heart Disease Prediction System Using Data Mining Techniques", IJCSNS International Journal of Computer Science and Network Security, Vol.8 No.8, August 2008. [18] Andreeva P., M. Dimitrova and A. Gegov, “Information Representation in Cardiological Knowledge Based System”, SAER’06, pp: 23-25 Sept, 2006. [19] Tzung-I Tang, Gang Zheng, Yalou Huang, Guangfu Shu, Pengtao Wang, "A Comparative Study of Medical Data Classification Methods Based on Decision Tree and System Reconstruction Analysis", IEMS, Vol. 4, No. 1, pp. 102-108, June 2005. [19] M. Zekic-Susac, B. Klicek. A Nonlinear Strategy of Selecting NN Architectures for Stock Return Predictions. Finance, Proceedings from the
Page 99
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
50th Anniversary Financial Conference Svishtov, Bulgaria, 11-12 April, 2002, pp. 325-355. [20] M. Zekic-Susac, N. Sarlija, M. Bensic. Small Business Credit Scoring: A Comparison of Logistic Regression, Neural Network, and Decision Tree Models, Proceedings of the 26th International Conference on Information Technology Interfaces, June 7-10, 2004, Cavtat, pp. 265-270.
Velammal College of Engineering and Technology, Madurai
Page 100
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
GAP: Genetic Algorithm based Power Estimation Technique for Behavioral Circuits Johnpaul C. I1, Elson Paul2, Dr. K. Najeeb3 Department of Computer Science & Engineering Government Engineering College, Sreekrishnapuram Palakkad, India-679517 1
[email protected] [email protected] 3 [email protected]
2
Abstract— The problem of peak power estimation in digital circuits is essential for analyzing the reliability and performance of circuits at extreme conditions. The power estimation problem involves finding input vectors that cause maximum dynamic power dissipation (maximum toggles) in circuits. In this paper, an approach for power estimation for both combinational and sequential circuits at behavioral level is presented. The basic intuition behind the approach is to use the genetic algorithm based approach to find the vector pairs for maximum toggling in the circuit. Peak dynamic power consumption is strongly dependent on the stream of inputs applied to the circuit. Without any information about the input stream, it is impossible to accurately estimate the power consumption of a design. Thus, for a power estimation technique it is necessary to use approximation methods. The proposed technique was employed on the ISCAS'85, ISCAS'89 and ITC benchmark circuits. Keywords— CMOS circuits, power dissipation, power estimation, automatic power vector generation, genetic algorithms
VI. INTRODUCTION The high transistor density, together with the growing importance of reliability as a design issue, has made early estimation of worst case power dissipation (peak power estimation) [1] in the design cycle of logic circuits an important problem. High power dissipation may results in decrease in performance or in extreme cases cause burn out and damage to the circuit. The increased transistor density in the processors can be well defined by Moore's law which describes a long term trend in the computing hardware, in which the number of transistors that can be placed inexpensively on an integrated circuit has doubled approximately in two years. Whenever technology advances with new high performance processors, transistor density will obey Moore's law. As more and more transistors are put on a chip, the cost to make each transistor decreases, but the chance that the chip will not work due to a defect increases. Moreover due to the complexity of their operations their interrelationship between the transistors increases. This will increase the processor complexity both in its design and complexity. In this context it is essential to formulate the
Velammal College of Engineering and Technology, Madurai
power estimation methods that will become more helpful in the phase of designing complex digital circuits. The power is an important factor in designing the circuits [2]. In the case of processors containing billions and billions of transistors, power dissipated is more. Hence power management is essential in designing the circuits of a processor to optimize its performance. As technology advances, the bus width of transistor circuits evolves in the order of 32nm, 22nm, 16nm, 11nm, the implementation of power management techniques is a crucial part in the production of complex hardware circuits. Moreover when the bus width is reduced to 11nm range, the number of transistors on a single die dramatically increases. Another impact due to this phenomenon is that the clock frequency increases exponentially with increase in the number of transistors. Another dilemma due to increase in power is thermal run away. The power in the circuit can be expressed in the form of heat dissipated [3]. Due to increase in the heat, current in the circuit increases which will further increase the heat. This will cause destruction in the circuit. Parameter variations cause many hurdles for the CPU designers. The most critical sources of parameter variations are in process characteristics across dies, supply voltage drop and temperature variation [4]. These are generally known as PVT variations. Parameter variations causes the worst case delay in flow of current and, hence effects performance and power [5]. This will further increase heat. In other words the peak power determines the thermal and electrical limits of components and the system packaging requirements [6]. The peak power consumption corresponds to the highest switching activity generated in the circuit under the test during one clock cycle [7]. The dynamic power dissipated in the combinational portion of the sequential circuit [8] (PR ) can be computed as shown in equation (1):
P R=
1 × V 2DD× f 2
∑ [N
g ×C g
]
(1)
Page 101
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
where the summation is performed overall gates g, and N(g) is the number of times gate g has switched from 0 to 1 or vice versa within a given clock cycle, C(g) is the output capacitance of gate g, VDD is the supply voltage and f is the frequency of operation. In the proposed method the assumption taken for power estimation is that the output capacitance for each gate is equal to the number of fanouts. Therefore, the total switching activity is the parameter that needs to be maximized for maximum dynamic power dissipation. Accurate estimation of maximum power consumption involves finding a pair of input vectors which when applied in sequence, maximize the value of PR given by equation (1), among all possible input vector pairs.
power dissipation in the digital circuits. Rest of the paper is organized as follows. Section 2 describes the related work done in this field of power optimization. Section 3 describes the frame work developed with genetic algorithm approach. It includes algorithm used for optimization and toggle vector generation. Section 4 includes the implementation details and experimental results on ISCAS'85, ISCAS'89 and ITC benchmark circuits. Section 5 concludes the whole work. PREVIOUS WORK There are different techniques that have been developed to solve the problems related to peak power estimation. Each of these methods have their own peculiarities and limitations.
Power estimation of digital circuits can be viewed under various abstraction levels. By carefully analyzing complex circuits under these abstractions [3], it will be more convenient to determine the power dissipation in each stages. The different abstraction levels includes system, behavioral, RT level, gate level, circuit level and device level. Each of these levels contains notable peculiarities of different parameters which include estimation accuracy, speed, estimation capacity, complexity, ease of abstraction etc. A designer can choose any of these abstraction levels for conducting a detailed study of power estimation in a systematic pattern [7]. The figure 1 shows the complexity of power estimation for different abstraction levels.
Fig. 2 Classification of Peak Dynamic Power Estimation (PDPE) methods The figure 2 shows different methods of peak dynamic power estimation. The different models that are used for the power estimation in circuits are SAT based approach, constraint based approach, simulation based approach, ATPG based approach, genetic algorithm based and macro-modeling.
Fig. 1 Power estimation vs. different abstraction levels VII.
A. OUTLINE AND CONTRIBUTION OF THIS PAPER
This paper proposes a methodology for peak dynamic power estimation in behavioral circuits using genetic algorithm [9]. The paper proposes the generation of an intermediate file from Verilog input file which will finally fed into the proposed algorithm for optimization process. This optimization process is to obtain an optimized pair of input vectors that produce maximum toggle in the corresponding circuits. This input vectors can be otherwise called as Genetic Power Virus. Power virus in the sense that it will result in the maximum
Velammal College of Engineering and Technology, Madurai
Recent advances in Boolean satisfiability (SAT) models [10] and algorithms have made it tempting to use satisfiability based techniques in solving various VLSI design related problems such as verification and test generation [11]. Usually SAT solvers are used to handle zero-one integer linear programming problems [11]. The constraints are typically expressed in conjunctive normal form (CNF) [12]. This method works well for small and medium sized circuits. But when the size of the circuit increases it will work slower. Simulation based power estimation method can be applied after completing the design of a circuit [13]. It provides an accurate way to estimate power consumption of the design because the power estimation based on the method which reflects the actual design behavior. Since simulation happens later in the design cycle, this power estimation is generally used to verify the power consumption of a device on board. This method is also used as a tool to estimate the power
Page 102
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
dissipation in the portions of a larger design when integrating smaller designs into larger FPGAs. In the case of gate level approach the simulation can be of the form that different gates are simulated using the programming methods and the test vectors are fed to the circuit as input to analyze the functionality [14]. Automatic test pattern generation (ATPG) based techniques try to generate an input vector pair that produces the largest switched capacitance in the circuit [15]. The power consumption by the vector pair is calculated by counting the total number of circuit elements switched from one state to another [14]. The ATPG based techniques are very efficient and generate a tighter lower bound than that generated by random vector generation. ATPG based techniques can handle only simple delay models such as the zero-delay and unitdelay models [10]. Macro-modeling is a method of inclusion of large number of desired properties into a model, there by the worst case analysis of a complex system can be generated [16]. It can also be used to analyze the possible outcomes form a model. Several macro-modeling methodologies have been proposed in the recent past. The majority of these approaches targets specifically register-transfer level power estimation. The fundamental requisites for a power macro-models are evaluation efficiency and flexibility. In every macro-modeling techniques evaluation efficiency is dictated by the need of performing numerous power estimates while the synthesis tool is used for evaluating design alternatives [16]. Flexibility is imposed by the fact that the same macro can be instantiated in completely different designs or operated under widely varying conditions. Genetic algorithm (GA) based approach of power estimation is to determine the input vector pairs that can provide maximum toggles in the digital circuits [9]. Different design methodologies are considered in this approach. GA efficiently produced promising results in circuit partitioning problem and also the circuit element placement problem [9]. The prime aim of GA is optimization of the results that causes power dissipation and to find out the optimal pair of vectors that produces the maximum toggles in digital circuits [2] [7]. GA based approach is highly efficient in the design phase of complex digital circuits such as processors, co-processors etc. ALGORITHM FOR TOGGLE VECTOR GENERATION A. PROBLEM DEFINITION Given a description of circuit either in gate or behavioral level abstraction. The problem is to find out an optimal input test vector pairs v1 and v2 when applied in sequence, leads to maximum switching activity in the circuit. B. GENERATION OF INTERMEDIATE
FILE FROM VERILOG INPUT
Velammal College of Engineering and Technology, Madurai
The algorithm for the generation of intermediate file from Verilog description described in algorithm 1. The algorithm contains separate functions to isolate the input lines, output lines, assign lines, wire lines from the Verilog file. After the generation of corresponding arrays, the formatted data in the arrays are used for generating the intermediate file. For this, a function called tree() is used. The tree() function will traverse the elements of the assign array and replaces the various operations such as NEGATION, XOR, AND etc. to numerical form. The conversion is on the basis that each operations has given a specific number. The corresponding output is written to an output file, which is further used in this framework for the generation of test vectors. Algorithm 1 (INTERMEDIATE FILE) input: Verilog description of behavioral circuit. output: Numerical representation. assign: Assign statements. queue: Enqueue the assignment statements in the order of execution in the Verilog file. repeat Stmt= dequeue(Queue). Tree(stmt). Convert the statements to numerical representation. until Queue= NULL A typical module of the behavioral Verilog file is given below. The algorithm for the generation of the intermediate file scans the input in such a way that whenever it finds a module call from the main module it will execute that module functions and return to the main module. In this way the modules are interconnected. module PriorityA(E, A, PA, X11), Input[8:0] E, A; Output[8:0] X1; Wire [8:0] Ab, EAb; assign Ab = ~ A; assign EAb = ~ (Ab & E); assign PA = ~ EAb; assign X1 = ~ EAb ^ (9(PA)); endmodule. A behavioral level module of circuit 432b.v is shown above. The algorithm 2 explains the procedure used by the proposed approach. The core functions used in the approach are crossover and mutation. These basic functions drive the algorithm to produce the optimized results. Algorithm 2 (GENETIC ALGORITHM) Create initial population. repeat Select parent1 and parent2 from the population. Offspring= crossover (parent1,parent2).
Page 103
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
are randomly complemented. Mutation may sometimes result in the generation of vectors with better fitness value.
Mutation (offspring). until Stopping criteria is reached.
C . EXPLANATION
OF GENETIC ALGORITHM FOR THE FRAMEWORK
Updation Create Initial Population A population of size P is generated randomly to create the initial population. Here the population elements are the strings of binary numbers containing zeros and ones which are equal to the number of input lines in the circuit [9]. The size of the population is to be large enough so that variance in the strings can be easily formed. From this initial population optimization process is performed for the production of the toggle vectors. Selection of parents Each individual or a string has a fitness value, which is a measure of quality of the solution represented by the individual [18] [19]. The formula which is used to calculate the fitness value of each individual
F i =T i
The creation of two offspring increases the size of the population by two. Since it is necessary to maintain a constant population size, two vectors will need to be eliminated from the population [9]. For this elimination process fitness values are used. In the population there are vectors with the lowest fitness value. Two of such vectors are removed and the newly generated offspring are added to the population. Stopping Criteria Stopping criteria uses a limit value W to determine when the algorithm stops. If there is no improvement after W generations, then the algorithm stops. No improvement means that there are no changes in the maximum fitness value of the population. The final solution is the individual with the highest fitness value [9], [19].
T w− T b 3
(2)
IMPLEMENTATION AND EXPERIMENTAL RESULTS
where Fi represents the fitness value of the individual, Ti represents the toggle count of individual i. Tw represents the smallest toggle count in the population and Tb is the largest toggle count in the population. The fitness values are used for sorting the strings in the ascending order. From the middle of the population size two individuals are selected for performing crossover. Crossover After two parents are selected, crossover is performed on the parents to create two offspring [9]. A string split point is randomly selected, and is used to split each parent string into two. The first offspring is created by concatenating the left half of the first parent and the right half of the second parent, while the second offspring is created by concatenating the left half of the first parent and the complement of the right half of the second parent. Mutation Mutation is the phenomenon that produces sudden change in the bit sequence of each string [9]. The mutation operation is performed on each string in such a way that a bit position is selected randomly and that bit is complemented. Mutation can be designed in another way that from the position say b bits
Velammal College of Engineering and Technology, Madurai
Fig. 3 Toggle vector generation The implementation details can be visualized from figure 3. The input to the test vector generator algorithm is in the form of Verilog file. The genetic algorithm is applied over both gate level circuits and behavioral level circuit. The proposed intermediate file is generated from the Verilog input file. This proposed intermediate file is the key file input to the GA. The above mentioned intermediate file contains the information of fanout and the output lines, input lines, wires etc. of circuit elements. The intermediate file of behavioral level and gate level circuits are entirely different from each other. A gatelevel intermediate file is shown in figure 3. The interpretation
Page 104
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of this file is that the specific circuit described contains total number of 1145 gates. Second line till end contains the interconnection to different gates. The first number in each line determines the type of the gate. A typical gate level intermediate file is given below: 1145 2 333 45 -1 3 43 34 675 -1 7 23 56 763 -1 8 89 47 -1 2 44 231 673 89 -1 2 54 33 43 -1 2 22 66 332 90 -1 3 84 431 221 643 -1 7 432 123 243 745 -1 3 67 -1 2 87 336 89 -1 In the case of behavioral level intermediate file [20], it contains the specific bit positions of each module and also the total bit count in the output. Under genetic algorithm, various operations of each module is simulated and record the changes after a pair of input vector is fed into the genetic algorithm. By calculating the number of bits changed due to the application of second optimized input vector [11], toggle count is obtained. Optimized in the sense that applying various steps of the GA in a specified manner [11]. As toggle count is obtained from this experiments, the power dissipation from each circuit can be understood. From the different experiments conducted on ISCAS'85, ISCAS'89 and ITC benchmark circuits, it can infer that the toggle count in behavioral level circuit is better promising result than the gate level [17]. 31 135 36 3 1 70 56 -1 0 4 17 3 1 70 78 0 0 136 144 -1 0 70 78 -1 0 0 8 3 1 54 54 0 1 136 136 -1 0 79 87 2 0 46 44 -1 0 78 87 4 1 136 144 -1 0 54 54 3 1 88 96 -1 0 0 8 3 1 97 105 1 0 136 144 -1 0 88 96 -1 0 18 26 3 1 106 114 0 0 136 144 -1 0 36 44 -1 97 105 3 1 55 55 0 1 136 136 -1 0 106 114 2 0 45 53 -1 0 106 114 4 1 136 144 -1 0 55 55 3 1 88 96 -1 0 0 8 3 1 115 123 1 0 136 144 -1 0 88 96 -1 0 27 35 Toggling in the gate-level circuits are shown in the table I contain various benchmark circuits, with their corresponding number of gates in the circuit. Genetic algorithm is applied to all these circuits with a population size P. The number of generations (iterations) is represented as W. The number of gates in each circuit is also shown. The output produced by the genetic algorithm will be the test vector pairs that produces maximum toggling. The optimization of the test vectors are done by the GA. The generation of test vector pairs is shown in the figure below. The selection of the
Velammal College of Engineering and Technology, Madurai
optimal pair of vector pairs is based on the maximum toggle count. The first two test vectors are the fittest solution that produces the maximum toggling in a given circuit. Since the maximum toggling is found out from this experiments, the notion of maximum power dissipation can also be known because they are directly proportional. Thus the peak power can be estimated.
Fig. 4 Toggle vector generation Circuit
No. of gates
P
W
Toggle count
In lines
c17.v
6
10
100
5
5
c432.v
160
10
100
76
36
c7552.v
3513
10
100
1603
207
b01.v
70
10
100
37
4
b02.v
41
10
100
10
3
b03.v
312
10
100
96
6
b04.v
998
10
100
219
13
b05.v
1007
10
100
484
3
b06.v
73
10
100
37
4
b07.v
682
10
100
142
3
b08.v
261
10
100
109
11
b09.v
273
10
100
91
3
b10.v
288
10
100
103
13
b11.v
793
10
100
234
9
b11n.v
1145
10
100
284
9
b12.v
1084
10
100
537
7
b13.v
577
10
100
230
12
b14.v
6437
10
100
2398
34
b141.v
6934
10
100
2766
34
b15.v
12292
10
100
3649
38
b20.v
13980
10
100
5251
34
b211.v
14882
10
100
5646
34
b221.v
21671
10
100
7749
34
Page 105
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
boothG.v
1145
10
100
658
10
IdctG.v
7550
10
100
3086
41
TABLE I TOGGLE COUNT OF GATE-LEVEL CIRCUITS
The results of toggling in behavioral circuits is also represented as the number of bits toggled. The input vector pair that produced maximum toggling is also calculated. In the case of behavioral and gate level this input vector pair is in the form of long sequences of binary strings which will be equal to the number of input lines in gate level and it will be equal to the number of bits in the input lines of behavioral circuits. CIRCUIT
TOTAL BIT
P
W
TOGGLE
c432b.v
135
10
100
57
c74L85b.v
13
10
10
11
c74L81b.v
52
10
100
20
c6288.v
32
10
100
22
TABLE II TOGGLE COUNT OF BEHAVIOR LEVEL CIRCUITS
Tables I and II show the experimental results conducted on different ISCAS'85, ISCAS'89 and ITC benchmark circuits. The programming language used for the implementation of the proposed approach is C. It proved to be more flexible for accessing hardware resources such as memory, file operations etc. The intermediate file production is performed using scripting language Perl. It is more convenient to use this language in order to extract sufficient information from an input file. The context of using this language is that the input to the peak power estimation program [1] is a Verilog file. Most of the cases the data extracted from the input file is with the help of regular expressions. CONCLUSION AND FUTURE WORK
The proposed method of power estimation is a novel approach in the field of power estimation. GA based power estimation in digital circuits not only generates the required toggle vectors but also optimizes it. Experimental results using the ISCAS'85, ISCAS'89 and ITC benchmark circuits shows that the proposed method could produce an optimized results on behavioral circuits. The GA based power estimation can be extended for power estimation in complex processors, for analyzing their behavioral peculiarities. Future work shall involve modification of this method for behavioral level power estimation of many core architectures. VIII.
REFERENCES
[56] [1] E. M. Rudnick M. S. Hsiao and J. H. Patel, “Peak power estimation of [57] VLSI circuits: New peak power measures,” in IEEE Transactions on [58] VLSI Systems, 2000, p. 435439.
Velammal College of Engineering and Technology, Madurai
[59] [2] E. M. Rudnick M. S. Hsiao and J. H. Patel, “Peak power estimation of [60] VLSI circuits: New peak power measures,” in IEEE Transactions on [61] Very Large Scale Integration Systems., 2000, pp. 435–439. [62] [3] C.T. Hsieh C.S. Ding, Q. Wu and M. Pedram, “Statistical estimation of [63] the cumulative distribution function for power dissipation in VLSI [64] circuits,” in Proceedings of Design Automation Conference, 1997, pp. [65] 371–376. [66] [4] Liao and Lepak K. M, “Temperature and supply voltage aware [67] performance and power modeling at micro architecture level,” in IEEE [68] Trans on Computer-aided Design of Integrated Circuits and Systems, [69] 2005, pp. 1042–1053. [70] [5] M. Hunger and S. Hellebrand, “Verification and analysis of self[71] checking properties through ATPG,” in Proc. 14th IEEE Int. On-Line [72] Testing Symp (IOLTS08), 2008, pp. 25–30. [73] [6] S. Sheng A. P. Chandrakasan and R. W. Broderson, “Lowpower CMOS [74] digital design,” Journal of Solid-State Circuit, vol. 27, no. 4, pp. 473– [75] 483, April 1992. [76] [7] H. Takase K. Zhang, T. Shinogi and T. Hayashi, “A method for [77] evaluating upper bound of simultaneous switching gates using circuit [78] partition,” in Asia and South Pacific Design Automation Conf. [79] (ASPDAC), 1999, p. 291. [80] [8] M. Pedram C-Y. Tsui and A. M. Despain, “Efficient estimation of [81] dynamic power dissipation under a real delay model,” in IEEE [82] Transactions on Computer-Aided Design of Integrated Circuits and [83] Systems, 1993, pp. 224–228. [84] [9] Theodore W. Manikas and James T. Cain, “Genetic algorithms vs [85] simulated annealing: A comparison of approaches for solving the circuit [86] partitioning problem,” in IEEE Trans. on Computer-aided Design, 1996, [87] pp. 67–72. [88] [10] P. Schneider and U. Schlichtmann, “Decomposition of Boolean [89] functions for low power based on a new power estimation technique.,” [90] In Proceedings of 1994 International Workshop on Low Power Design, [91] 1994, pp. 123–128. [92] [11] S. Devdas F. Fallah and K. Keutzer, “Functional vector generation for [93] HDL models using linear programming and Boolean satisfiability,” in [94] IEEE Transactions on Computer-Aided Design of Integrated Circuits [95] and Systems, August 2001, pp. 994–1002. [96] [12] S. Ravi L. Lingappan and N. K. Jha, “Satisfiability-based test [97] generation for non separable RTL controller-datapath circuits,” in IEEE [98] Transactions on Computer-Aided Design of Integrated Circuits and [99] Systems., 2006, pp. 544–557.
Page 106
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[100] [13] Q. artas, Simulation Based Power Estimation, Altera Corporation, [101] 2004. [102] [14] A. K. Chandra and V. S. Iyengar, “Constraint solving for test case [103] generation: a technique for high-level design verification,” in In [104] Proceedings on International Conference on Computer Design: VLSI [105] in Computers and Processors., 1992, pp. 245–248. [106] [15] I. Ghosh and M. Fujiita, “Automatic test pattern generation for [107] functional register transfer level circuits using assignment decision [108] diagrams,” in IEEE Trans on Computer-aided Design of Integrated [109] Circuits and Systems, 2001, pp. 402–415. [110] [16] Noel Menezes Chandramouli Kashyap, Chirayu Amin and Eli Chiprou, [111] “A nonlinear cell macromodel for digital applications,” in in IEEE [112] Transactions on Computer-Aided Design of Integrated Circuits and [113] Systems, 2007, p. 678685. [114] [17] C. Y. Wang and K. Roy, “Cosmos: A continuous optimization [115] approach for maximum power estimation of CMOS circuits,” in [116] Proceedings of ACM/IEEE International Conference on Computer [117] Aided Design, 1997, pp. 45–50. [118] [18] Thomas Weise, Global Optimization Algorithms: Theory and [119] Application, www.it-weise.de/projects/book.pdf, Germany, 2006. [120] [19] Alex Fraser and Donald Burnell, Computer Models in Genetics, [121] McGraw-Hill, New York, 2002. [122] [20] R. Vemuri and R. Kalyanaraman, “Generation of design verification [123] tests from behavioral VHDL programs using path enumeration and [124] constraint programming,” in in IEEE Transaction on VLSI Systems, 1995, pp. 201–214.
Velammal College of Engineering and Technology, Madurai
Page 107
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Human Action Classification Using 3D Star Skeletonization and RVM Classifier Mrs. B. Yogameena#1, M. Archana*2, Dr. (Mrs) S. Raju Abhaikumar#3 #
Deptartment of Electronics and Communication, Affiliated to Anna University Thiagarjar College of Engineering, Madurai – 625015, Tamil Nadu, India. 1
[email protected] [email protected]
3
*
Deptartment of Electronics and Communication, Affiliated to Anna University 94/16, Irulandi st, Natraj nagar, Madurai – 625016, Tamil Nadu, India. 2
[email protected]
Abstract - This paper presents a real-time video surveillance system which classifies normal and abnormal action in crowds. Detection of individual’s abnormal action has become a critical problem in the events like terrorist attacks. The aim of this work is to provide a system which can aid in monitoring human actions, using 3D star skeleton technique, which is a suitable skeletonization for human posture representation and it reflects the 3D information of human posture. This 3D star skeleton technique can offer more accurate skeleton of posture than the existing star skeleton techniques as the 3D data of the object is concerned. Using these 3D features as input to the Relevance Vector Machine (RVM) classifier, the different human actions from a stationary camera is classified as normal or abnormal for a video surveillance application. Keywords – Video surveillance, Abnormal action, 3D star skeletonization, Relevance Vector Machine.
I. INTRODUCTION Security of citizens in public places such as Hotels, Markets, Airports and Train stations is increasingly becoming a crucial issue. The fundamental problem in visual surveillance system is detecting human presence, tracking human motion, analysing the activity and asses abnormal situations automatically. Based on this motivation, crowd modelling technology has been under development to analyse the video input which is constantly crowded with humans, as well as to ready to act against abnormal activities emerge. The aim of the paper is to classify human normal and abnormal actions in crowds using 3D star skeletonization for a surveillance system which is based on RVM. Human action and posture recognition is a significant part on human centred interface which is coming up to the resent issues now-a-days. In posture recognition application, the skeletal representation [1] captures the essential topology of the underlying object in a compact form which is easy to understand. There are three existing methods for skeletal construction such as distance transformation [2]-[5], voronoi diagram [6]-[8] and thinning [9]-[12]. These methods can generate an accurate skeleton but are computationally
Velammal College of Engineering and Technology, Madurai
expensive and cannot be used in real time. To complement these, a 2D star skeleton was proposed [13], [14], [15]. However it also has some limitations to describe the 3D information of the posture. Hence the 3D star skeletonization method is proposed for the feature extraction for the classifier. Then the normal and abnormal human action is classified in the real time surveillance applications to analyze the unusual activity of an individual from the normal ones. Many methods based on computer vision have been proposed in the literature to classify people’s action earlier. Ardizzonee et al. [16] has proposed a pose classification algorithm using support vector machine to classify different poses of the operator's arms as direction commands like turn-left, turnright, go-straight, and so on. In [17] the authors proposed a method that describes a system which delivers robust segmentation of the hands using a combination of colour and motion analysis, and the implementation of a multi-class classification of the hand gestures using a SVM ensemble. Cipolla and Philip et al. [18] estimate the pose of an articulated object from a single camera using relevance vector machine. In [19] Hui-Cheng Lian et al. presented a novel approach to multi-view gender classification considering both shape and texture information to represent facial image and the classification is performed using SVM. These papers dose not concentrate on full body pose. Hence Relevance Vector Machine is considered which utilizes only the relevance vectors for the training [15]. In this paper, an improved 3D star skeleton [20] is proposed which is based on the 3D information of the human posture whose purpose is to recognize human action as normal or abnormal. The 3D features are given as input to the Relevance Vector Machine for classification of human action. Experimental results demonstrate that the approach is robust in classifying human abnormal action. The remainder of this paper is organized as follows: Section 2 describes the methodology. Section 3 describes the 3D star skeletonization. Section 4 describes the RVM learning system for classification. Section 5 describes the experimental results. Finally, the conclusion is presented.
Page 108
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Video Sequence
Background Subtraction
3D Star Skeletonization
Classification Using Relevance Vector Machine
Classified Poses Fig 1: The block diagram of the Video Surveillance System I. METHODOLOGY The overview of the system is shown in fig.1. In the foreground detection stage, the blob detection system detects the foreground pixels using a statistical background model. Subsequently foreground pixels will be grouped into blobs. The 3D star skeleton features are obtained for the foreground image sequences. Using these features the Relevance Vector Machine classification scheme classifies the given image sequence into normal action or abnormal action sequence.
II.
BACKGROUND SUBTRACTION AND PROJECTION In this work, background subtraction is accomplished in real-time using the adaptive mixture of Gaussians method proposed by Atev et al. [21]. There are some of the practical issues concerning the use of the existing algorithm based on mixtures of Gaussians for background segmentation in outdoor scenes, including the choice of parameters [[22], [15]]. The proposed system analyses the choice of different parameter values and their performance impact is obtained to get robust background model. In addition, the motivation for dopting this method stems from its simplicity and efficiency in meeting with sudden global illumination changes based on the contrast changes over time. Subsequently, the individual is to be identified for further analysis leading to action classification. An extracted blob in a frame, representing an individual is subjected to action analysis described in subsequent sections. If there exists more than one blob, but with connectivity, there is likelihood to be considered as single entity. This results in the identification of a group as “individual.” This makes recognition of individual’s action in a crowd more difficult. Therefore, a geometric projection on the blob is proposed to separate an individual from the group for analysing his or her actions. The blob is projected to head and ground plane from the camera view point leading to intersected area in world coordinates. Such projection shown in Figure 2 eliminates the variation of area with the distance from the camera so that it identifies only humans [15]. The success of human identification lies on segmentation of individual human in a given frame as a single blob. However, there is a chance of multiple blobs representing an individual human due to over segmentation.
Velammal College of Engineering and Technology, Madurai
But, since the projection of a blob is accomplished from head plane to ground plane, any discontinuity in a blob representing an individual is achieved by linking discontinuous blobs covered by bounding rectangle. III. 3D STAR SKELETONIZATION Star skeleton, which is a “star” fashion, is a kind of representative features to describe a human action. 3D star skeleton technique extracts the 3D information of the image sequence. It is simple, real-time and robust technique. There are 4 steps involved id 3D star skeletonization as described below. A. Projection Map Generation The definition of a projection map is to project the 3D world onto the 2D image plane orthographically. A pixel value on the projection map generally contains 3D information of the target object, such as nearest boundary, furthest boundary and the thickness of the object.To generate the projection map, and initially the 8 projection plane is created. Then a virtual camera is set to the projection plane and perpendicular vectors are drawn along the viewing direction. These hit points along the viewing direction are obtained and is saved in a buffer. From this the maximum hit points are obtained. The depth information of the human posture is obtained from these projection maps. The 8 projection map is used as the input to construct the 3D star skeleton. Each of the projection map used has a different silhouette of human action according to where the projection map is generated and each pixel value means a distance from view point to a voxel of model surface, depth information. Through these properties star skeleton can diminish the loss of 3D human posture information. B. Candidate Extraction Candidate extraction means detection of gross extremities of posture in the projection map. The extremities of boundary in the projection map are called the “candidate”. Candidate is likely to be the final extremity of 3D star skeleton as a feature. As the existing star-skeleton is generated by the extremities which are from 2D boundary points, there are some limitations to describe a 3D human posture. To overcome these limitations, the 3D information losses of human posture is reduced by using the 3D boundary, in which depth, and x, y co-ordinates data of the posture in the projection map are included.Candidates are found from the distance from each boundary trace in clockwise or counter-clockwise order. The distance function has noise component of the boundary, which can be reduced by the noise reduction process executed using smoothing filters. Consequently, candidates are detected by finding zero crossings in the smoothed distance function.
Page 109
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1)
Candidate Extraction Process Boundary of posture in the projection map is given as input to the candidate extraction system. The centroid of the candidate of the posture boundary in a projection map (xc, yc, zc) are extracted using the following expressions.
xc =
1 N ∑ xi N j =1
1 Nb
(2)
Nb
∑z i =1
M rotation
( xi − xc ) + ( yi − yc ) + ( zi − zc ) 2
2
2
(4)
Then the distance function is declared as a one dimensional discrete function as d(i)=di. But this distance function is corrupted by noise of the boundary. This noise is reduced by a smoothing filter. The smoothed distance function is . C. dCoordinate Transformation (i ) In transformation phase, the candidates of each projection map are transformed into one particular coordinate. This is done because the coordinate systems of the extracted candidates are not identical. All candidates exist in their own projection map coordinate system. To construct 3D star skeleton, the candidates should be in the same coordinate. To transform all candidates into one particular coordinate, 3 phases are done: reflection, translation, rotation process. Reflection is to fit the two opposite projection map coordinates to one coordinate. The applied reflection matrix is,
M reflection
⎛ −1 ⎜ 0 =⎜ ⎜0 ⎜ ⎝0
0 0 w ⎞ ⎟ 1 0 0 ⎟ 0 −1 255 ⎟ ⎟ 0 0 1 ⎠
(5)
Where, w is the width of projection map and maximum values of z in the projection map is 255. To rotate the candidates to one particular coordinate, the axis of all candidates should be translated to one identical axis, which is
Velammal College of Engineering and Technology, Madurai
⎛ ⎜1 ⎜ ⎜0 =⎜ ⎜0 ⎜ ⎜0 ⎝
w⎞ 2⎟ ⎟ 1 0 0 ⎟ 256 ⎟ 0 −1 ⎟ 2 ⎟ 0 0 1 ⎟⎠ 0
0
−
(6)
The rotation of the 4 couples of projection map coordinates, which are fitted at reflection process, into one particular coordinate is performed by the process of rotation. The rotation matrix is given as,
(3)
i
Where, Nb is the number of boundary points. (xi, yi, zi) is a boundary point of the projection map. The distance di from each boundary point (xi, yi, zi) to the centroid (xc, yc, zc) is obtained using,
di =
M translation
(1)
1 N yc = ∑ yi N j =1
zc =
performed by the translation process. Translation matrix performs the axis shift operation which is given by,
⎛ cos θ ⎜ 0 =⎜ ⎜ − sin θ ⎜ ⎝ 0
0 sin θ 1 0 0 cos θ 0 0
0⎞ ⎟ 0⎟ 0⎟ ⎟ 1⎠
(7)
Where, θis the variable which is changed according to the projection map. D. Clustering Candidates and 3D Star Skeleton Construction To determine the extremities as features of the 3D star skeleton, the transformed candidates should be classified. After transformation process the transformed candidates are scattered in one particular coordinate indicating the some part of posture locates near others generated from the same part, not exist at one specific position together due to the thickness of human body. Hence, all the candidates are to be classified into several groups. The mean of each cluster becomes the extremities of star skeleton. ‘K’ mean clustering algorithm is used to classify the candidates. According to which the candidates are divided into k groups, defining K as 5. The centroid is the average of centroids on all the projection maps. Finally, the 3D star skeleton is constructed by connecting the extremities with the centroid. The features of 3D star skeleton are generated by calculating the distance from extremities to the centroid of 3D star skeleton. In some cases the candidates might not locate at the landmark points, such as feet or hands. For example, one knee point can screw up the foot canter if it is included in that cluster leading to error in clustering process. Hence K-Mean Clustering Algorithm is modified for fitting to the proposed method. To reduce the error candidates, noises are filtered from the cluster by using standard deviation. After clustering process, if some candidate is far from the mean of its cluster, it is removed and the mean of that cluster is recalculated. Thus the 3D star skeleton is constructed. IV. RELEVANCE VECTOR MACHINE Action classification is a key process for analyzing the human action. Computer vision techniques are helpful in automating this process, but cluttered environments and consequent
Page 110
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
occlusions often makes this task difficult [24]. There are numerous methods for incremental model based pose estimation where a model of an articulated structure (person) is specified [25]-[26]. Many types of neural networks are used for a binary classification problem like individual’s activity classification as normal or abnormal. By training the systems, the difference between normal and abnormal human actions, the computational action models built inside the trained machines can automatically identify whether the action is normal or abnormal. The action classification system proposed in this paper is trained for both normal and abnormal actions so that testing becomes a two class hypothesis problem. SVM is classical training algorithm because it has stronger theory-interpretation and better generalization than the other neural networks mentioned earlier. The decision function of the SVM classification system cannot be much sparser i.e., the number of support vectors can be much larger. This problem can be partially overcome by the state of the art model RVM. The proposed Relevance Vector Machine (RVM) classification technique has been applied in many different areas of pattern recognition, including functional neuro images analysis [27], facial expressions recognition [28] and pose estimation [29]. The RVM is a Bayesian regression framework, in which the weights of each input vector are governed by a set of hyper parameters. These hyper parameters describe the posterior distribution of the weights and are estimated iteratively during training. Most hyper parameters approach infinity, causing the posterior distributions of the corresponding weights to zero. The remaining vectors with non-zero weights are called relevance vectors. RVM does not need the tuning of a regularization parameter and also the inversion of a large matrix is not required during the training phase. This makes this methodology appropriate for large datasets. In this paper, Relevance Vector Machine technique is used for the classification of human action such as normal or abnormal. A. Classification of poses using multi-class Relevance Vector Machine The extracted 3D features are inputted to Relevance Vector machine for learning to classify poses. In our framework, we are focused our attention on the pose of the whole human body. This kind of gestures can be described very well through the analysis of the body contour. Moreover, to provide a detailed description of a shape, one has to take into account the whole body contour. The image features were shape-contexts descriptors of silhouette points and pose estimation was formulated as a one-to-one mapping from the feature space to pose space, as shown in Fig.4. The pose of an articulated object, a full human body, is represented by a parameter vector x given in eqn (8)
X = wk φ ( z ) + ξ K
k
w - is the weight of the basis function ф(z) - is the vector of the basis function k
ξ - is the Gaussian noise vector In order to learn about the multiple RVM technique an expectation maximization algorithm has been adopted. It is used to minimize the cost function of the multiple RVM regression shown in (9)
( )
Lk = ∑ C k( n ) Yk( n )
T
( )
S k Yk( n )
(9)
( )
Yk( n ) = x( n ) − wkφ Z ( n ) Here, Y
(n)
k
(10)
-is the output with n sample points belongs to the
mapping function k. k
W - is the weight of the basis function. (n)
Φ(z ) –is the design matrix of vector of the basis function. k
S – is the diagonal covariance matrix of the basis function. C
k
(n)
- is the probability that the sample point n belongs
to the mapping function k. V. RESULTS and DISCUSSION: The efficiency of the proposed algorithm has been evaluated by carrying out extensive work on the simulation of the algorithm using Matlab7.9. The proposed method process about 24 frames of any size each on PC with AVI file format. The Weizmann dataset for walking, running, bending and jumping are taken and CAVIAR dataset for a person lifting his hand is taken. In order to test for complexity, IBM dataset having multiple people walking and CMU database containing two persons fighting with each other in one sequence and a person pulling the hands of the other person in one sequence are taken. The video sequence is converted into frames and the foreground frames are obtained using GMM. As mentioned in Table.1. Dataset DS I, has a person bending down along with two other foreground blobs is detected. In DS II, there are multiple people with two groups, one group of people in the camera’s viewpoint and the other group of people away from camera’s viewpoint. Hence the people away from camera’s viewpoint are easily identified as individuals and the remaining as group. In DS III five foreground blobs are obtained including the running person’s blob. In all the three DS the car is considered to be a static background.
(8)
Where, x – is the input for the system
Velammal College of Engineering and Technology, Madurai
Page 111
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE 1: DATASETS FROM DIFFERENT OUTDOOR SEQUENCES Data set College Campus
Benchmark Dataset
IBM Dataset Weizmann Dataset
CMU Dataset CAVIAR
Indoor/outdoor
Sequence Length
Frame Size
Indoor/outdoor
1628
240X320
Indoor/outdoor
864
240X320
Running human in a group-DS-III
outdoor
1320
240X320
A person carrying a bar in a group DS- IV Two persons Walking
outdoor
920
240X320
Indoor
781
240X320
Eli-Walk
Outdoor
645
240X320
Eli-Run
Outdoor
712
240X320
Moshe-Bend
Outdoor
786
240X320
Eli-Jump
Outdoor
855
240X320
Indoor
1234
240X320
An individual bends down while most of walking-DS-I(DS represents Data Set) A person waving hand in a group-DS-II
Person A fights with person B Person A pulls the person B
Indoor
1065
240X320
An individual with his hand lifted up
Indoor
1187
240X320
Then the 3D star skeleton obtained for the different datasets using the projection map system. For the individual blob like walking, running and jumping the skeletal features are obtained even more clearly than the 2D star skeleton. For the bending action DS I, the skeleton features vary from the Weizmann dataset in 2D star skeleton. In the Weizmann dataset the skeleton points were depicted as a short human walking, were as it is a skeleton of a man “bending”. This problem is overcome in 3D star skeletonization. The bending person is identified correctly. In 2D skeleton of the CMU fighting datasets the person who holds the stool is skeletonised including the stool. But in 3D star skeletonization the stool is identified separately from the person holding it. Similarly for the CMU pulling dataset also the skeleton features are obtained clearly for the two individuals than the 2D star skeleton. Finally, the skeleton features are fed to the Relevance Vector Machine to classify the abnormal action from the normal action which is indicated in red color as shown in Figure (2.a.), Figure (2.b.) & Figure (2.c.) and also for benchmark datasets as shown in Figure (4).
a)One person bending down
Velammal College of Engineering and Technology, Madurai
b) One person waving hand
c) One person running in the crowd and another person half bending Fig.2. Results of classified abnormal actions for College Campus dataset
Relevance Vector Machine uses suitable kernel for the task of classification. Gaussian kernels are used in this proposed method. A ten-fold cross validation (CV) has been applied in the training dataset, to determine the fine-tuning parameters of the RVM classifier model for optimal performance. For each dataset, 80% of the sequence has been used for training and the remaining 20% for testing. A best error level of 6.89% is obtained by using the Gaussian Kernel. Gaussian Kernel is the best with around 100% for training, 94% for CV and 96% for testing the given feature vectors. The error rate of the Gaussian Kernel is lower than that of other kernels in terms of classification rate. The results of RVM classification is shown in table.2.
Page 112
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE 2: THE RESULTS OF RVM CLASSIFICATION Actions
Normal Abnormal
Benchmark Datasets
Running Carrying Bar Bending Waving hand IBM Eli-Walk Eli-Run MosheBend Eli-Jump CMU1 CMU2 CAVIAR
Vectors RVM Multi class RVM with skeleton Points 11 18 23 12 16 12 15 18 20 16 19 14 12
and false negative. Higher error level generally leads to poor classification of actions. From Figure (3) it is inferred that, as the number of iterations increases the better convergence of relevance vector is achieved for the experimental database. When the 2D star skeleton features were inputted to the RVM classifier the abnormal actions except the Weizmann bending and CMU pulling were classified. Weizmann bending was classified as normal action and in CMU pulling dataset both individuals were considered as walking. But this is overcome by using 3D star skeleton features as input to the RVM classifier. The Weizmann bending and CMU pulling datasets are classified as abnormal actions as shown in Figure (4).
VI. CONCLUSION In this paper, a novel and real-time video surveillance system for classifying human normal and abnormal action is described. Initially, the foreground blobs are detected using adaptive mixtures of Gaussians which is robust to illumination changes and shadows. Then the projection map system is generated to get the 3D information of the object in the foreground image. Then the 3D star skeleton features are extracted. These features reduce the training time and also improve the classification accuracy. The features are then learnt though a relevance vector machine to classify the individual’s actions into normal and abnormal behaviour. The number of relevance vectors obtained is less and it does not require the tuning of a regularization parameter during the training phase. The error rate is also reduced by selecting the appropriate Gaussian kernel which also reduces the computational complexity. The distict contribution of this proposed work is to classify the actions of individuals. The proposed system is able to detect abnormal actions of individuals such as running, bending down, waving hand while others walk, and people fighting with each other with high accuracy.
Fig 3: Relevance vector obtained after training input feature vectors
The percentage of error level considered for both training and testing sequences of the dataset are in terms of true positive
Velammal College of Engineering and Technology, Madurai
Page 113
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Dataset
Original Frame
Background Subtracted Image
Skeleton Image
Classified Image
IBM dataset
Weizmann dataset Eli-Walking
Eli-Running
MosheBending
EliJumping CMU Fighting
CMU Pulling
CAVIAR
Fig 4: Results for benchmark datasets
Velammal College of Engineering and Technology, Madurai
Page 114
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
ACKNOWLEDGMENT We sincerely thank our management and our respectful principal for providing us with all the facilities we required for carrying on our research. We also thank them for their support and encouragement in all aspects for our research. REFERENCES [1] Nicu D. Cornea, Deborah Silver, Patrick Min, “ Curve-Skeleton Properties, Application, and Algorithms,” IEEE Trans. Visualization and Computer Graphics, vol.13, 2007, pp. 530-548. [2] Gunilla Borgefors, “Distance transformations in digital images,” Computer Vision, Graphics, and Image Processing, vol. 34, 1986. pp. 344-371. [3] Gunilla Borgefors, “ Distance transformations in arbitrary Dimensions,” Computer Vision, Graphics, and Image Processing, vol. 27, 1984. pp. 321-345. [4] Gunilla Borgefors, “ On digital distance transforms in three dimensions,” Computer Vision and Image Understanding, vol. 64, 1996, pp. 368-376. [5] Frank Y.Shih and Christopher C.Pu, “A skeletonization algorithm by maxima tracking on Euclidean distance transform”, J.Pattern Recognition, vol.28, 1995, pp. 331-341. [6] Franz Aurenhammer, “Voronoi diagrams – A Survey of a fundamental geometric data structure,” ACM Commputing Survey, vol. 23,1991, pp. 345-405. [7] Jonathan W. Brandt and V. Ralph Algazi, “ Continuous skeleton computation by Voronoi diagram,” CVGIP: Image understanding, vol. 55, 1991, pp. 329-338. [8] Kenneth E. Hoff III, Tim Culver, John Keyser, Ming Lin and Dinesh manocha, “Fast computation of generalied Voronoi diagrams using graphic hardware,” in Proc. 26th annual Conf. Computer graphics and interactive tech-nique, 1999, pp. 277-286. [9] Kalman Palagyi, Erich Sorantin, Emese Balogh, Attila Kuba, Csongor Halmail, Balazs Erdohelyi, and Klaus Hausegger, “A Sequential 3D Thinning Algorithm and Its Medical Applications,” in Proc. 17th international Conf.IPMI, vol. 2082, 2001, pp. 409-415. [10] Kalman Palagyi and Attila Kuba, “ A 3D 6-subiteration thinning algorithm for extracting medial lines,” Pattern Recognition Letters, vol. 19, 1998, pp. 613-627. [11] Kalman Palagyi and Attila Kuba, “ Directional 3D thinning using 8 subiterations,” in proc. 8th international Conf. DGCI, vol. 1568, 1999, pp. 325-336. [12] Ta-Chin Lee, Rangasami L. Kashyap and Chong-Nam Chu, “ Building skeleton models via 3-D medical Surface/axis thinning algorithms,” CVGIP : Graphical Models and Image Processing, vol. 56, 1994, pp. 462-478. [13] H. Fujiyoshi and A.J Lipton, “Real-time human motion analysis by image skeletonization,” 4th IEEE Workshop on Application of Computer Vision, 1998, pp. 15-21. [14] Hsuan-Sheng Chen, Hua-Tsung Chen, Yi-Wen Chen and Suh-Yin Lee, “Human Action Recognition Using Star Skeleton,” in Proc. 4th ACM international workshop on video surveillance and sensor networks, 2006, pp.171-178. [15] B. Yogameena, S. Veeralakshmi,E. Komagal, S. Raju, and V. Abhaikumar,“ RVM-Based Human Action Classification in Crowd through Projection and Star Skeletonization,” in EURASIP Journal on Image and Video Processing, vol. 2009, 2009. [16] E. Ardizzone, A. Chella, R. Phone, “ Pose Classification Using Support Vector Machines,” International Joint Conference on Neural Networks, vol. 6, 2000. [17] D.F. Llorca, F. Vilarino, Z. Zhouand G. Lacey, “A multi-class SVM classifier ensemble for automatic hand Washing quality assessment,” Computer science Dep, Trinity College Dublin, Rep, of Ireland. [18] Cipolla and Philip et al,“ Pose estimation and tracking using Multivariate regression,” pattern recognition,2008.
Velammal College of Engineering and Technology, Madurai
[19] Hui – Cheng Lian and Bao – Liang Lu, “ Multi - View Gender Classification using Local Binary Patterns and Support Vector machines,” verlag, 2006, pp. 202-209. [20] Sungkuk Chun, Kwangjin Hong, and Keechul Jung, “ 3D Star Skeleton for Fast Human Posture Representation,” Proceedings of World Academy of Science, Engineering and Technology, vol. 34, 2008. [21] S. Atev, O. Masoud, N.P. Papanikolopoulos, “ Practical mixtures of Gaussians with brightness monitoring,” Proc. IEEE Int. Conf. Intel. Transport. Syst. pp.423–428, October, 2004. [22] C. Stauffer and W.E.L Grimson., “Adaptive background mixture models for real time tracking,” In proceedings of the IEEE Int’l conf.Computer Vision and Pattern recognition pp. 246 -252 1999. [23] G. Tzikas, A. Likas, N. P. Galatsanos, A. S. Lukic and M. N. Wernick, “Relevance vector machine analysis of functional neuroimages,” IEEE intel. Symposium on Biomedical Imaging, vol.1, pp. 1004-1007, 2004. [24] H.C.Lian, B.L. Lu, “Multi-view Gender Classification Using Local Binary Patterns and Support Vector Machines,”-verlag -pp. 202–209, 2006. [25] R. Chellappa, A. K. Roy-Chowdhury and S. K. Zhou, “Recognition of Humans and Their Activities Using Video,” First edition, Morgan and Claypool publishers, 2005. [26] H. Ren and G. Xu, “Human action recognition in smart classroom,” IEEE Proc. Int. Conf. on Automatic Face and Gesture Recognition, pp. 399404, 2002. [27] A. Mittal, L. Zhao, L.S. Davis, “ Human Body Pose Estimation Using Silhouette Shape Analysis,” – proceedings of the IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS'03), 2003 [28] D Rita Cucchiara , A. Prati, R. Vezzani, University of Modena and Reggio Emilia, Italy, “Posture Classification in a Multi-camera Indoor Environment,” – International conference on image processing..vol 1, 2005. [29] G. Tzikas, A. Likas, N. P. Galatsanos, A. S. Lukic and M. N. Wernick, “Relevance vector machine analysis of functional neuroimages,” IEEE intel. Symposium on Biomedical Imaging, vol.1, pp. 1004-1007, 2004.
Page 115
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Relevance Vector Machine Based Gender Classification using Gait Appearance Features Mrs. B. Yogameena#1, M. Archana*2, Dr. (Mrs) S. Raju Abhaikumar#3 #
Deptartment of Electronics and Communication, Affiliated to Anna University Thiagarjar College of Engineering, Madurai – 625015, Tamil Nadu, India. 1
[email protected] [email protected]
3
*
Deptartment of Electronics and Communication, Affiliated to Anna University 94/16, Irulandi st, Natraj nagar, Madurai – 625016, Tamil Nadu, India. 2
[email protected]
Abstract - Automatic gait recognition is important task in security surveillance, biomechanics, purloining behaviors and biometrics. The background subtraction for detecting foreground information, the reliable extraction of characteristic gait features from image sequences and their classification are three important issues in gait recognition. The person has to be found in the image plane first, using robust background subtraction by modifying the parameters in Gaussian Mixture Model (GMM), assuming a static camera.This paper describes the representation of consistent gait appearance features and Relevance Vector Machine (RVM) based gender classification. Experimental results on benchmark datasets demonstrate the proposed gender classification method is robust and efficient. A comparative study between RVM based classification and Support Vector Machine (SVM) based classification is also presented. Keywords - Gait recognition, Relevance Vector Machine, appearance features, background subtraction
I. INTRODUCTION Automatic visual surveillance systems could play an important role in supporting and eventually replacing human observers. To become practical, this system needs to distinguish people from other objects and to recognize individual persons with a sufficient degree of reliability, depending on the specific application and security level. These applications required the task of estimating the gait and automatically recognising the gender for high level analysis. The first step in the automatic gender classification system is background subtraction. Piccardi et al. have reviewed a number of background subtraction approaches [1]. Wren et al. [2] have proposed a statistical method, in which a single Gaussian function was used to model the distribution of background. Later Mittal et al. have proposed a novel kernel based multivariate density estimation technique that adapts the bandwidth according to the uncertainties [3].Yet there are issues like the robustness to illumination changes, the effectiveness in suppressing shadows and the smoothness of foreground’s boundary which need to be addressed in indoor and outdoor environments [4].
Velammal College of Engineering and Technology, Madurai
There have been a number of appearance-based algorithms for gait and activity recognition. Cutler and Davis [5] used self-correlation of moving foreground objects to distinguish walking humans from other moving objects such as cars. D.F. Llorca et al., [6] proposed a method that describes a system which delivers robust segmentation of the hands using a combination of color and motion analysis, and the implementation of a multi-class classification of the hand gestures using a SVM ensemble. Cipolla and Philip have been estimated the walking nature of an individual of an articulated object from a single camera using relevance vector machine [7]. In [8] Hui-Cheng Lian et al., presented a novel approach to multi-view gender classification considering both shape and texture information to represent facial image and the classification is performed using SVM. Nixon, et al., [10] used principal component analysis of images of a walking person to identify the walker by gait. Shutler, et al., [11] used higher-order moments summed over successive images of a walking sequence as features in the task of identifying persons by their gait. Johnson and Bobick [12] utilized static parameters of the walking figure, such as height and stride length, to identify individuals. However it has been reported that the feature extraction is an important issue for classification algorithm [9]. The consistency of the number of features for a machine learning classification algorithm also plays a vital role in gender classification. There are many types of neural networks that can be used for a binary classification problem, such as Support Vector Machines, Radial Basis Function Networks, Nearest Neighbor Algorithm, and Fisher Linear Discriminant and so on. Machine learning techniques can be considered as linear methods in a high dimensional feature space nonlinearly related to the input space. Using appropriate kernel functions, it is possible to compute the separating hyper plane which classifies the two classes. An automated system based on SVM classifier that classifies gender by utilizing a set of human gait data has been described by Yoo et al. [9]. The decision function of the SVM classification system can not be much sparser i.e., the number of support vectors can be much larger. SVM need the tuning of a regularization parameter
Page 116
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
during the training phase. Consequently a machine learning technique which minimizes the number of active kernel functions to reduce computation time is required ([18]-[20]). The objective of this paper is to classify gender using appropriate features and classification technique to improve the classification accuracy. Hence a novel system is proposed to classify the gender. First the foreground blobs are detected using a background subtraction technique which is to be robust to illumination changes and shadows. Then appropriate features and RVM are used to classify the gender and thereby reducing the computational complexity by selecting an appropriate kernel also by improving the classification accuracy. Experimental results demonstrated that the proposed approach is robust in classifying gender. The remainder of this paper is organized as follows: Section 2 describes the methodology. Section 3 describes background subtraction, section 4 presents the formulation of centroid, section 5 represents the elliptical view of body posture, section 6 presents the extraction of feature and section 7 depicts the experimental results and the classification of gender using the Relevance Vector Machine extraction method. Finally, the conclusion is presented. II. METHODOLOGY Figure 1 shows an outline of the proposed system. Each image from the camera is forwarded to the pre-processing module where the background subtracted image sequences are obtained. Then consistent features have been selected using silhouette segmentation and ellipse fitting. Finally, the Relevance Vector Machine classification scheme classifies the gender as male or female for the given image sequence.
III.BACKGROUND SUBTRACTION The first stage of video surveillance systems seeks background to automatically identify people, objects, or events of interest in various changing environments. The difficult part of background subtraction is not the differencing itself, but maintenance of a background model and its associated statistics. In this work, background substraction is accomplished in real-time using the adaptive mixture of Gaussians method proposed by Atev et al [22]. There are some of the practical issues concerning the use of the existing algorithm based on mixtures of Gaussians for background segmentation in outdoor scenes, including the choice of parameters [23]. The proposed system analyses the choice of different parameter values and their performance impact are obtained to get robust background model. In addition to that, the notion for adapting this method is because of its simplicity and also an efficient method for coping with sudden global illumination changes based on the contrast changes over time. It describes K Gaussian distributions to model the surface reflectance value and is represented by eqn (1). k
P ( X t ) = ∑ ω i ,t *η ( X t , μ i ,t , ∑ i, t ) Where K is the number of distributions,
ωi,t is an estimate of the weight of the ith Gaussian, μ is the mean,
∑ i, t
is the covariance matrix of the ith Gaussian, and
η is a Gaussian probability density function. where, η(X , μ , ∑i,t) = t i,t
Background subtraction
Silhouette segmentation using centroid
Ellipse Fitting
……. (1)
i =1
1 n
1 (2π) 2 |∑| 2
− 1 (Xt −μt )T ∑−1(Xt −μt ) l 2
..... (2)
Assume the covariance matrix is of the form
∑
k
= σ k2 I
…… (3)
This means that the red, green and blue reflectance components of the surface are independent and have the same variances, which can reduce costly matrix computation. The weight is adjusted as shown in eqn (4)
ω ω t =(1 − α )ω t −1 +α ( M t ) i, t
Feature Extraction .
Classification using RVM III.
…… (4)
Where α is the learning rate and
M t is 1 for the model which is matched and 0 for others. After ordering the Gaussians, the first B distributions are chosen as the background model as shown in eqn (5)
Classified Gender
Velammal College of Engineering and Technology, Madurai
Page 117
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
b
B = arg min b (∑ ω k ⟩T )
.…. (5)
k =1
Where, T is a measure of minimum models for background. The background subtracted image obtained using adaptive GMM is shown in Fig. 2.
Fig.2.An example of the background subtracted image of weizmann dataset
IV. CENTROID For each silhouette of a gait, the centroid (xc, yc) is determined by using the following eqns (6 ),(7) and is shown in Fig 3.
xc =
1 N
N
∑x j =1
1 yc = N
….. (6)
i N
∑y j =1
i
….. (7)
Where (xc, yc) represent the average contour pixel position, (xi, yi) represent the points on the human blob contour and there are a total of N number of points on the contour [16].
The frontal-parallel view of the silhouette is divided into the front and back sections (except for the head region) by a vertical line at the silhouette centroid. The parts above and below the centroid are each equally divided in the horizontal direction, resulting in 7 regions that roughly correspond to: head/shoulder region, front of torso, back of torso, front thigh; back thigh, front calf/foot, and back calf/foot [12]. These seven regions are chosen to be very easy to compute in contrast to methods that locate the joints and segment the body at those joints. These regions are by no means meant to segment the body parts precisely. V. ELLIPTICAL VIEW After getting the centroid of the full body region and the seven segments of a silhouette, for each of the 7 regions, an ellipse is fitted to the portion of foreground object visible in that region as shown in Figure 5. The fitting of an ellipse to an image region involves computing the mean and the covariance matrix for the foreground pixels in the region. Let I(x, y) be the binary foreground image of a region to which the ellipse is to be fitted [17]. Assume that the foreground pixels are 1 and the background pixels are 0, then the centroid of the region, is calculated using the following relations,
1 ∑ I ( x, y ) x N x, y 1 y = ∑ I ( x, y ) y N x, y x=
…….. (8)
……. (9)
where N is the total number of foreground pixels:
N = ∑ I ( x, y )
......... (10)
x, y
The covariance matrix of the foreground region is then obtained by, Fig.3.An example figure for pointing the centroid of an weizmann dataset
Then the silhouette is proportionally divided into 7 parts as shown in Fig 4.
⎡( x − x ) 2 ( x − x)( y − y )⎤ ⎡a c ⎤ 1 ⎥ ⎢b b ⎥ = N .∑ I ( x, y ).⎢ 2 x, y ⎣ ⎦ ⎣⎢( x − x)( y − y ) ( y − y ) ⎦⎥ ............(11) The covariance matrix can be decomposed into eigenvalues λ1, λ2 and eigenvectors v1, v2 which indicate the length and orientation of the major and minor axes of the ellipse which is obtained by eqn (12).
⎡λ1 0 ⎤ ⎡a c ⎤ ⎢c b ⎥ [v1 v 2 ]= [v1 v 2 ] ⎢0 λ ⎥ ⎣ ⎦ 2⎦ ⎣
……. (12)
Fig 4: An example figure for the segmented silhouette image of Weizmann dataset
Velammal College of Engineering and Technology, Madurai
Page 118
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The elongation of the ellipse, l, is given by
l=
λ1 λ2
.
…. (13)
and the orientation α of the major axis is given by the following relation
⎛ v .x ⎞ α = angle (v1 ) = arccos⎜⎜ 1 ⎟⎟ ⎝ v1 ⎠
Frame 113
Frame 125
Frame 145
….. (14)
where x is the unit vector [1, 0].
Frame 124
Frame 132
Frame 129
Fig.6: Human gait image from Weizmann dataset
Fig 5: An elliptical view for the segmented silhouette image of Weizmann dataset
I. FEATURE EXTRACTION After calculating the orientation and elongation of the human walking figure, an ellipse is fitted according to the features extracted from each frame of a walking sequence consists of features from each of the 7 regions. They are the relative height of the centroid of the whole body [12], for his or her body length which is taken as a single feature and then the four ellipse parameter consist of the x, y coordinates of the centroid, the orientation, and the elongation, which correspond to head region, chest, back, centroid of the head, orientation of the head, and finally, the mean and variance of the back calf. These six features constitute about the appearance of the gait and are shown in Table I.
In this proposed method, there are about six different image sequences of walking of both genders in an outdoor environment are taken from Weizmann dataset [16] for the classification of the gender using machine learning technique by analyzing their gait as shown in Figure 6.For this experimentation, there are about five subjects are walking in an outdoor environment consisting of three men and two women taken from weizmann dataset [15]. Totally six datasets have been taken for the classification of the gender by analyzing their gait as shown in fig (6). The detection of the moving objects from the foreground region, the background subtracted images are shown in fig (7). As shown in fig (3), the centroid has been taken for all the frames and the ellipse is fitted for each frame as shown in fig (4). Using the above information, the features which are used to classify the subject as a men or women are given as follows: front calf, back, head (centroid and orientation), mean and standard deviation of the back calf. Then these feature vectors are given for training and testing of the relevance vector machine.
Instead of using the twenty nine features, these primary six features are used for the gender classification.
II. CLASSIFICATION OF GENDER USING RELEVANCE VECTOR MACHINE In recent years, machine learning methods have become prevalent to examine the pattern of structural data. Kernel methods such as Support Vector Machines (SVM) are widely used to classify the poses of faces, hands, different human body parts and Robots. Relevance Vector Machine Yields a formulation similar to that of a Support Vector Machine and it also uses hyper parameters instead of Margin/Costs. The RVM is a Bayesian regression framework, in which the weights of each input example are governed by a set of hyper parameters [18]. These hyper parameters describe the posterior distribution of the weights and are estimated iteratively during training. Most hyper parameters approach infinity, causing the posterior distributions of the effectively setting the corresponding weights to zero. The remaining vectors with non-zero weights are called relevance vectors. It reduces the “inappropriate” parameters for the classification than SVM [[6], [17]]. In this framework, the attention is focused to the nature of the gait of a human to classify the subject as a man or woman.
Velammal College of Engineering and Technology, Madurai
Page 119
TABLE. I .TYPES OF FEATURES S.NO
REGION
1
front calf
FEATURES mean of orientation
2 3
Back Head
4 5
Head back calf
mean of orientation mean of x coordinate of centroid mean of orientation std of x of centroid
6
back calf
mean of x of centroid
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
x = Wk φ (z) +
ξk
……. (15)
Where, x – is the input for the system Wk - is the weight of the basis function ф(z) - is the vector of the basis function ξk - is the Gaussian noise vector It is used to minimize the cost function of the RVM regression using eqn (16)
( )
( )
LK = ∑ C K( n ) y k(n ) S k y k(n ) T
where, (n )
……. (16)
( )
y k = x (n ) − W K φ Z (n )
……… (17) Where, yk(n)-is the output with n sample points belongs to the mapping function k. Φ(z(n)) –is the design matrix of vector of the basis function. Sk – is the diagonal covariance matrix of the basis function. Ck(n) - is the probability that the sample point n belongs to the mapping function k.
III. RESULTS AND DISCUSSION OF RVM CLASSIFICATION The efficiency of the algorithm proposed has been evaluated by carrying out extensive works on the simulation of the algorithm .In this paper the video file used is Weizmann dataset [21]. The proposed method processes about 24 frames per second and the total number of frames is up to 460 at size of 240x320 each.
Frame 113
Frame125
Frame 145
features and are given as input for the RVM for the gender classification. The classification accuracy is measured with two types of kernels in this proposed method. The two types of feature vectors of each subject are used for testing, training and cross validation (CV) for both SVM and RVM. The experimental results are summarised in Table II. The accuracy is the average by the number of SVs, RVs, and classification rate. The test result of 6 selected features is a little higher than that of 29 original features. Also the numbers of relevant vectors are smaller when compared to the support vectors. The average accuracy of linear kernel was the best with around 100.0% for training, 94.0% for CV and 96.0% for testing in the 6 features. The error rate of the linear kernel was lower than that of other kernels in terms of classification rate and computational cost. The Fig.8 shows the initial stage of the training. The Fig. 9 show the result after the fifth iteration. It is to be noted that the 3 colors (Indigo, red and yellow) represents the gender as a male and the remaining 3 colors (Blue, pink and green) indicates the gender as a female by setting the threshold value. The color dots in the figure denote the relevance vectors of the corresponding dataset which is obtained after the fifth iteration. The maximum number of iterations used here is ten. Figure.10 shows the comparative representation of SVM and RVM vectors, for the gender classification. The vectors used for gender classification is reduced for RVM than in SVM. From the fig 10, the results for the benchmark datasets have been shown with the corresponding step by step procedure as explained by the proposed method. Finally the classified image is obtained as, the male sequences are identified through the green color rectangular bounding box and the female sequences a recognized through the red color bounding box. Output (Z+ weight)
These kinds of gestures can be described very well through the analysis of the gait. Moreover, to provide a detailed description of a shape, one has to take into account the gait. The walking pose of an articulated object, gait is represented by a parameter vector x and is given in eqn (15)
Input feature vector (Z) Fig.8.Initial Stage
Frame 124
Frame 132
Frame 129
Fig.7. background subtracted image of the weizmann dataset
The background subtracted images using GMM are shown in Fig 7. After background subtraction, the centroid is obtained and the silhouette is segmented into seven regions. Then ellipse is fitted for each segment and is shown in Figures (35). Then six primary features are extracted from the total 29
Velammal College of Engineering and Technology, Madurai
Page 120
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Output (Z+ weight)
Input feature vector (Z) Fig.9.Result for Classification
XI. CONCLUSION This paper has introduced the framework of relevance vector machine to classify the gender for the application of video
ACKNOWLEDGMENT We sincerely thank our management and our respectful principal for providing us with all the facilities we required for carrying on our research. We also thank them for their support and encouragement in all aspects for our research.
REFERENCES [1] M. Piccardi, “Background subtraction techniques: A review,” IEEE International Conference on Systems,Man and Cybernetics, vol.4, pp. 3099-3104, Oct. 2004. [2] C. Wren, A. Azarbayejani, T. Darrell, and A. Pentland, “Pfinder: Real-Time Tracking of the Human Body,” IEEE Trans.On Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp.780-785, July 1997. [3] A. Mittal, N. Paragios, “Motion-Based Background Subtraction using Adaptive Kernel Density Estimation,” In Proceedings of the Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2004), pp. 302-309, 2004 [4] J. Hu and T.Su “Robust background subtraction with shadow and Highlight removal for indoor surveillance,” In proceedings of the EURASIP Journal on Advances in signal processing, 14 pages,2007. [5] R. Cutler and L. Davis., “Robust real-time periodic motion detection, analysis, and applications”. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2000, 22(8):781 –796. [6] D.F. Llorca, F. Vilarino, Z. Zhou and G. Lacey, “A multi-class SVM classifier ensemble for automatic hand washing quality assessment”, British Machine Vision Conference proceedings 2007. [7] Cipolla and Philip et al “pose estimation and tracking using multivariate regression” pattern recognition Letters 2008. [8] Hui-Cheng Lian and Bao-Liang Lu “Multi-view Gender Classification Using Local Binary Patterns and Support Vector Machines”Verlag, pp. 202–209, 2006. [9] Yoo, D. Hwang and M. S. Nixon ,, “Gender Classification in Human Gait Using Support Vector Machine” ACIVS 2005, LNCS 3708, Springer-Verlag Berlin Heidelberg, pp. 138 – 145, 2005 [10] M. Nixon, J. Carter, J. Nash, P. Huang, D. Cunado, and S. Stevenage, “Automatic gait recognition”, IEE Colloquium on Motion Analysis and Tracking, pp. 3/1–3/6, 1999.
Velammal College of Engineering and Technology, Madurai
surveillance. The appearance features which are used to recognize the individual as a gender are extracted using elliptical view of seven segmented regions of a gait and fed into the RVM learning system. It reduces the number of vectors and improved the classification accuracy by choosing the appropriate kernel. To this end the gait sequence is classified as male or female. The method has been used as a component of a system for estimating the counts of gender for video surveillance. It is been noted that utilizing the six features instead of utilizing the twenty nine feature shows the good classification accuracy and these results were also compared with the SVM. Apart from this, the gender classification task can clearly handle a large number of subjects successfully. By this, these results show that people can be identified according to gender by their walking pattern. This accord with earlier psychological suggestions and buttressing other similar results.It is evident that RVM has good accuracy.
[11] J. Shutler, M. Nixon, and C. Harris. “Statistical gait recognition via velocity moments”, Visual Biometrics (Ref.No. 2000/018), IEE Colloquium, pp. 11/1–11/5, 2000. [12] A. Johnson and A. Bobick. “Gait recognition using static, activity specific parameters”, CVPR, 2001. [13] C. BenAbdelkader, R. Cutler, and L. Davis. “Motion-based recognition of people in eigengait space”, Automatic Face and Gesture Recognition Proceedings, Fifth IEEE International Conference, pp.267 – 272, May 2002. [14] Rong Zhang, Christian Vogler, Dimitris Metaxas , “Human gait recognition at sagittal plane” Image and Vision Computing, Vol 25, Issue 3, pp 321-330, March 2007. [15] Carl Edward Rasmussen, “The Infinite Gaussian Mixture Model” Advances in Neural Information Processing Systems, pp. 554–560, 2004. [16] Prahlad Kilambi, Evan Ribnick, Ajay J. Joshi, Osama Masoud, Nikolaos Papanikolopoulos, “Estimating pedestrian counts in groups”, Computer Vision and Image Understanding, pp. 43–59,2008. [17] Lily Lee, “Gait analysis for classification” – AI Technical Report Massachusetts Institute of Technology, Cambridge, USA. June 2003. [18] Tipping, M.E. “Sparse Bayesian learning and the relevance vector machine”, J. Mach. Learn. Res, pp. 211–244, 2001. [19] Tipping, M.E., Faul. A., “Fast marginal likelihood maximization for sparse bayesian models”, In: Proc. 9th Internat. Workshop on Artificial Intelligence and Statistics, 2003. [20] Agarwal, A., Triggs, B., “3D human pose from silhouettes by relevance vector regression”, In: Proc. Conf. on Computer Vision and Pattern Recognition, vol. II. Washington, DC, pp. 882–888, July 2004. [21] Weizmann dataset Downloaded from http://www.wisdom.weizmann.ac.il/~vision/SpaceTime Actions.html [22] S. Atev, O. Masoud, N.P. Papanikolopoulos, “Practical mixtures of Gaussians with brightness monitoring,”Proc. IEEE Int. Conf. Intel.Transport. Syst. pp.423–428, October, 2004.
C. Stauffer and W.E.L Grimson., “Adaptive background mixture models for real time tracking,” In proceedings of the IEEE Int’l conf.Computer Vision and Pattern recognition pp. 246 -252 1999
Page 121
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE II EXPERIMENTAL RESULTS FOR CLASSIFICATION Kernel
No. Of
SV
V
Classification Rate(%)
features
Training SVM
Guassian
29
78 54
8 19
29
64
6
57
6
94.4 93.4
Cross Validation RVM
SVM
Testing
RVM
SVM
RVM
94.6 96
92 91
92 92
93 94
94 96
34 100.0
100.0
95
96
94
95
28 100.0
100.0
95
96
95
97
Linear
Original image
Background subtracted image
Segmented image
Elliptical view of an image
Classified image
Denis
Eli walking
Ido
Daria
Lena_ walk1
Lena_ walk2
Velammal College of Engineering and Technology, Madurai
Page 122
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Survey on Gait Recognition Using HMM Model M.Siva Sangari M.Yuvaraju Lecturer / Dept of MCA Lecturer / Dept of CSE Sri Ramakrishna Engineering College Anna University Coimbatore Coimbatore [email protected] [email protected] serve as a useful filtering tool that allows us to narrow the Abstract Vision-based human identification at a distance in search down to a considerably smaller set of potential surveillance systems has attracted more attention recently, and its candidates. current focus is on face, gait or activity-specific recognition. The Gait of each person is identified to be identical which can be used for identification of a person. Potential sources for gait biometrics are two important aspects namely gait dynamics and gait shape. So set of individuals is being identified and their gaits are being stored in the form of databases. There are different types of databases available namely CMU Mobo Database set, UMD database set, SOTON Large data set, Human ID gait challenge dataset etc. This paper has concentrated on the aspects of various persons experimental results on these databases based on Hidden Markov Model.
Keywords: Gait Recognition, Hidden Markov Model (HMM), Human ID I
INTRODUCTION
Vision-based human identification at a distance in surveillance applications has recently gained more interests, e.g., the Human ID Program being supported by DARPA. GAIT refers to the walking style of an individual. Studies in psychophysics indicate that human have the capability of recognizing people from even impoverished displays of gait, indicating the presence of identity information of gait. A gait cycle corresponds to one complete cycle from rest position toright-foot-forward-to-rest-to-left-foot-forward-to-rest position. In comparison with other first-generation biometrics such as fingerprints and iris, gait has the advantage of being noninvasive and non-contact, and it is also the only perceivable biometric for personal identification at a distance. Given the video of an unknown individual, we wish to use gait as a cue to find who person is among the individuals in the database. For a normal walk, gait sequences are repetitive and exhibit nearly periodic behavior. As gait databases continue to grow in size, it is conceivable that identifying a person only by gait may be difficult. However, gait can still
Velammal College of Engineering and Technology, Madurai
Furthermore, unlike faces, gait is also difficult to conceal. Similar in principle to other biometrics, gait is affected by some physical factors such as drunkenness, pregnancy and injuries involving joints. Ideally, the recognition features extracted from images should be invariant to factors other than gait, such as color, texture, or type of clothing. In most gait recognition approaches recognition features are extracted from silhouette images. Approaches in computer vision to the gait recognition problem can be broadly classified as being either modelbased or model-free. Both methodologies follow the general framework of feature extraction, feature correspondence and high-level processing. The major difference is with regard to feature correspondence between two consecutive frames. Model based approach assumes a priori models that match the two- dimensional (2-D) image sequences to the model data. Feature correspondence is automatically achieved once matching between the images and the model data is established. Examples of this approach include the work of Lee et al., where several ellipses are fitted to different parts of the binarized silhouette of the person and the parameters of these ellipses such as location of its centroid, eccentricity, etc. are used as a feature to represent the gait of a person. Recognition is achieved by template matching. Model-free methods establish correspondence between successive frames based upon the prediction or estimation of features related to position, velocity, shape, texture, and color. Examples of this approach include the work of Huang et al., whose optical flow to derive a motion image sequence for a walk cycle. Principal Components Analysis (PCA) is then applied to the binarized silhouette to derive what are called Eigen gaits. Alternatively, they assume some implicit notion of what is being observed Little and Boyd describe the shape of the human motion with the scale-independent features from moments of the dense optical flow, recognized the individuals by phase vectors estimated from the feature sequences. Sundaresan et al proposed a Hidden Markov Model (HMM) based on individual recognition by gait. We considered two
Page 123
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The third aspect is the distance measure used, which can be Euclidean simple dot product-based based on probabilistic models. The second class of approaches opts for parameters that can be used to characterize gait dynamics, such as stride length, cadence, and stride speed. Sometimes static body parameters, such as the ratio of sizes of various body parts are considered in conjunction with these parameters. However, these approaches have not reported high performances on common databases, partly due to their need for 3D calibration information. The third class of approaches emphasizes the
silhouette shape similarity and disregards or underplays temporal information. While extracting the silhouette shape there are two options. 1) To use the entire silhouette 2) To use only the outer contour of the silhouette. The choice of using either of the above features depends upon the quality of the binarized silhouettes. If the silhouettes are of good quality, the outer contour retains all the information of the silhouette and allows a representation, the dimension of which is an order of magnitude lower than that of the binarized silhouette. However, for low quality, low resolution data, the extraction of the outer contour from the binarized silhouette may not be reliable. In such situations, direct use of the binarized silhouette may be more appropriate. When compared with other two approaches the silhouette shape-based approaches have been used by various researchers in determining the gait of a person. Sundaresan et al proposed a HMM on individual recognition by gait. II. COMPARISON OF USE OF HMM In this paper we compared how the HMM is used to determine the Gait of each person and to recognize them. Zongyi Liu and Sudeep Sarkar have used a generic walking model as captures by a population Hidden Markov Model (pHMM) defined for a set of individuals. The states of this pHMM represent gait stances over one gait cycle and the observations are the silhouettes of the corresponding gait stances. For each sequence, they first used Viterbi decoding of the gait dynamics to arrive at one dynamics-normalized, averaged, gait cycle of fixed length. The distance between two sequences is the distance between the two corresponding dynamics-normalized gait cycles, which they quantify by the sum of the distances between the corresponding gait stances. Distances between two silhouettes from the same generic gait stance are computed in the linear discriminant analysis space so as to maximize the discrimination between persons, while minimizing the variations of the same subject under different conditions. The distance computation is constructed so that it is invariant to dilations and erosions of the silhouettes. Thus they handled variations in silhouette shape that can occur with changing imaging conditions. The results were presented on three different, publicly available, data sets. First, the HumanID Gait Challenge data set, which is the largest gait benchmarking data set that is available (122 subjects), exercising five different factors, i.e., viewpoint, shoe, surface, carrying condition, and time. It was significantly improve the performance across the hard experiments involving surface change and briefcase carrying conditions. Second, improved performance on the UMD gait data set that exercises time variations for 55 subjects. Third, on the CMU Mobo data set, we show results for matching across different walking speeds. It is being concluded that worth noting that there was no separate training for the UMD and CMU data sets. Ju Han and Bir Bhanu, they have consider individual recognition by activity specific human motion, i.e., regular human walking, which is used in most current approaches of
Velammal College of Engineering and Technology, Madurai
Page 124
image features, one being the width of the outer contour of the binarized silhouette, and the other being the binary silhouette itself. A set of exemplars that occur during a gait cycle is derived for each individual. To obtain the observation vector from the image features we employ two different methods. In the indirect approach the high-dimensional image feature is transformed to a lower dimensional space by generating the frame to exemplar (FED) distance. The FED vector captures both structural and dynamic traits of each individual. For compact and effective gait representation and recognition, the gait information in the FED vector sequences is captured using a HMM model for each individual. In the direct approach, we work with the feature vector directly and train an HMM for gait representation. The difference between the direct and indirect methods is that in the former the feature vector is directly used as the observation vector for the HMM whereas in the latter, the FED is used as the observation vector. In the direct method, we estimate the observation probability by an alternative approach based on the distance between the exemplars and the image features. In this way, we avoid learning highdimensional probability density functions. The performance of the methods is tested on different databases. II. HIDDEN MARKOV MODEL (HMM) Gait recognition approaches are basically of three types: 1) temporal alignment-based, 2) static parameterbased, and 3) silhouette shape-based approaches. Here we have discussed the silhouette shape-based approach. The first stage is the extraction of features such as whole silhouettes, principal components of silhouette boundary vector variations, silhouette width vectors, silhouette parts or Fourier Descriptors. The gait research group at the University of Southampton (Nixon et al.) has probably experimented with the largest number of possible feature types for recognition. This step also involves some normalization of size to impart some invariance with respect to distance from camera. The second step involves the alignment of sequences of these features, corresponding to the given two sequences to be matched. The alignment process can be based on simple temporal correlation dynamic time warping hidden Markov models phase locked-loops or Fourier analysis.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of images on the stance set creating a vector (Frame-to-Stance Distance or FSD) representation for each frame and use these samples to train an HMM model using the Baum-Welch algorithm .The viterbi algorithm is used in the evaluation phase to compute the forward probabilities. The absolute values of the log probability values are recorded as the similarity scores. IV .CONCLUSIONS This paper gives an analysis of how Hidden Markov Model (HMM) used for gait identification by various researcher. This provides the effectiveness and computational efficiency for real world applications. It’s being proved that the HMM is mostly used to obtain the silhouette from the images. Though there are different models are available, HMM provides an efficient way to extract the silhouette.
individual recognition by gait. They have proposed a new spatio-temporal gait representation, called Gait Energy Image (GEI), to characterize human walking properties for individual recognition by gait. To address the problem of the lack of training templates, they also proposed a novel approach for human recognition by combining statistical gait features from real and synthetic templates. They directly compute the real templates from training silhouette sequences, while generating the synthetic templates from training sequences by simulating silhouette distortion. They have used a statistical approach for learning effective features from real and synthetic templates. We compare the proposed GEI-based gait recognition approach with other gait recognition approaches on USF HumanID Database. Experimental results show that the proposed GEI is an effective and efficient gait representation for individual recognition, and the proposed approach achieves highly competitive performance with respect to the published gait recognition approaches. The fundamental assumptions made here are: 1) the order of poses in human walking cycles is the same, i.e., limbs move forward and backward in a similar way among normal people, and 2) differences exist in the phase of poses in a walking cycle, the extend of limbs, and the shape of the torso, etc. Under these assumptions, it is possible to represent spatiotemporal information in a single 2D gait template instead of an ordered image sequence. Naresh Cuntoor, Amit Kale and Rama Chellappa has analyzed three different sets of features are extracted from the sequence of binarized images of the walking person. Firstly, we investigate the swing in the hands and legs. Since gait is not completely symmetric in that the extent of forward swing of hands and legs is not equal to the extent of the backward swing, we build the left and right projection vectors. To match these time varying signals, dynamic time warping is employed. Secondly, fusion of leg dynamics and height combines results from dynamic and static sources. A Hidden Markov model is used to represent the leg dynamics. While the above two components consider the side view, the third case explores frontal gait. We characterize the performance of the recognition system using the cumulative match scores computed using the aforesaid matrix of similarity scores. As in any recognition system, we would like to obtain the best possible performance in terms of recognition rates. Combination of evidences obtained is not only logical but also statistically meaningful. We show that combining evidence using simple strategies such as Sum, Product and MIN rules improves the overall performance. To compare the truncated width vectors that contain the information about leg dynamics, we use HMM, which is a generalization of the DTW framework. There exists Markovian dependence between frames since the way humans go through the motion of walking has limited degrees of freedom. K-means clustering is used to identify ’key frames’ or ’stances’ during a half-cycle. We found that a choice of is justified by the rate-distortion curve. We project the sequence
REFERENCES [1]. Zongyi Liu and Sudeep Sarkar,”Improved Gait Recognition by Gait Dynamics normalization” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 28, No. 6, June 2006. [2]. Murat EK_INC_I,” Human Identification Using Gait”, Turk J Elec Engin, VOL.14, NO.2 2006, c T¨UB_ITAK. [3].Naresh Cuntoor, Amit Kale and Rama Chellappa, Combining Multiple Evidences for Gait Recognition. [4]. Rong Zhang, Christian Vogler, Dimitris Metaxas, “Human Gait Recognition” [5]. S. Sarkar, P.J. Phillips, Z. Liu, I. Robledo-Vega, P. Grother, and K.W. Bowyer, “The Human ID Gait Challenge Problem: Data Sets, Performance, and Analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 27, no. 2, pp. 162-177, Feb. 2005 [6]. Z. Liu and S. Sarkar, Simplest Representation yet for Gait Recognition: Averaged Silhouette, Proc. Int’l Conf. Pattern Recognition, vol. 4, pp. 211-214, 2004. [7]. D. Tolliver and R. Collins, Gait Shape Estimation for Identification, Proc. Int’l Conf. Audio and Video-Based Biometric Person Authentication, pp. 734-742, 2003. [8]. L. Wang, T. Tan, H. Ning, and W. Hu, “Silhouette Analysis-Based Gait Recognition for Human Identification,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 25, no. 12, pp. 15051518, Dec. 2003 [9]. R. Collins, R. Gross, and J. Shi, “Silhouette-Based Human Identification from Body Shape and Gait,” Proc. Int’l Conf. Automatic Face and Gesture Recognition, pp. 366-371, 2002.
Velammal College of Engineering and Technology, Madurai
Page 125
V. FUTURE WORKS HMM can be compared with the other models namely Principal Component Analysis (PCA),Baseline Algorithm etc their efficiency can be observed with respect with the other two model which is being discussed here. Also efficiency of the algorithm can be checked with various factors like time, surface, carrying luckages, under various climate conditions can be analyzed with the same Model (HMM).
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[10].L. Lee and W. E. L. Grimson, “Gait analysis for recognition and classification,” in Proc. IEEE Conf. Face and Gesture Recognition, 2002, pp.155–161. [11]. J. Little and J. Boyd, “Recognizing people by their gait: The shape of motion,” Videre, vol. 1, no. 2, pp. 1–32, 1998. [12]. J. Cutting and L. Kozlowski, “Recognizing friends by their walk: gait perception without familiarity cues,” Bull. Psycho. Soc., vol. 9, pp. 353–356, 1977. [13]. M. P. Murray, “Gait as a total pattern of movement,” Amer. J. Phys. Med., vol. 46, pp. 290–332, June 1967. [14]. M. P. Murray, A. B. Drought, and R.C. Kory, “Walking patterns of normal men,” J. Bone and Joint Surgery, vol. 46A, no. 2, pp.335–360, 1964.
Velammal College of Engineering and Technology, Madurai
Page 126
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Congestion Management Routing Protocol In Mobile Adhoc Networks A. Valarmathi1 and RM. Chandrasekaran2 1Department of Computer Applications, Anna University Tiruchirappalli, Tiruchirappalli -620 024, India E-mail: [email protected] 2Department of Computer Science and Engineering, Anna University Tiruchirappalli, Tiruchirappalli-620 024, India :[email protected] Abstract Streaming multimedia applications over Mobile Ad hoc Networks (MANETs) require a guaranteed or controlled load service. Network congestion is the dominant reason for packet loss, longer delay and jitter in streaming multimedia applications. Most of the present routing protocols are not designed to adapt congestion. In the present paper, the original DSR protocol was modified to monitor the occurrence of congestion by using multiple resource utilization thresholds as QoS attributes and trigger multi-pathrouting during the periods of congestion to improve QoS in MANETs. This paper also deals with the effects of mobility on the modified routing protocol. Simulation experiments are conducted using NS-2 network simulator to evaluate the performance of the modified DSR in terms of jitter and percentage packet loss and compared with the original DSR. The results showed that a significant improvement in performance of modified DSR was achieved and decreases the network congestion with the use of multi-path routing. Keywords: Mobile ad hoc network, DSR, Congestion, Multi-path routing, NS-2.
1. INTRODUCTION Mobile ad-hoc networks (MANETs) are autonomous dynamic networks which provide high flexibility. These networks uses multi-hop radio relay that operate without the support of any fixed infrastructure [1]. Streaming multimedia or CBR applications over MANET require minimal delay and packet loss. To meet these critical requirements, a MANET inherently depends on the routing scheme employed. Routing protocols for Ad hoc networks can be broadly classified as proactive and reactive. Proactive (or) table-driven routing algorithms employ distance vector based or link state based routing strategies. However, the main drawback of these algorithms is that the need for frequent table updation consumes significant amount of memory, bandwidth and battery power [2]. Example of such protocols is Optimized Link State Routing (OLSR) [3] and Destination Sequenced Distance Vector routing (DSDV) [4]. In reactive routing
Velammal College of Engineering and Technology, Madurai
protocols, a node searches and maintains a route to another node only when it requires transmitting data to that particular node. This on-demand reactive approach minimizes bandwidth utilization compared to proactive routing protocols. Ad-hoc On-Demand Distance Vector (AODV) [5], Dynamic Source Routing (DSR) [6], Temporally Ordered Routing Algorithm (TORA) [7] and Associatively-Based Routing (ABR) [8] are examples of such protocols. Among these, DSR is a simple and efficient routing protocol which is widely used in MANET. It operates based on two mechanisms such as route discovery and route maintenance. All these protocols rely on single-path routing algorithms, where all traffic to a destination is required to be routed through a single successor and when a link become congested, the entire traffic has to be rerouted. Routing protocols for MANET also have another dimension as congestion-un-adaptive routing and congestion adaptive routing. Existing routing protocols belong to the first group where congestion nonadaptiveness in routing may result in longer delay, higher overhead and many packet losses for traffic intensive data like multimedia applications. DSR does not make dynamic routing adjustments based on memory, delay, number of service requests or battery utilization information which often indicates certain levels of usage and possible occurrences of congestion. Detection of undesired paths enables the ability to implement route admission control and to perform multi-path routing scheme which could further improve the performance of the network in a significant manner. In the present paper, a modified DSR protocol is proposed to monitor congestion or multiple resource utilization levels, notify the routing process of congestion and invoke multi-path routing during the periods of congestion for streaming multimedia applications. In addition, this paper also determine the effects of various mobility(low, medium, high) models on the modified routing protocol and compare the performance metrics of jitter and percentage packet loss of standard DSR protocol with modified version of DSR protocol.
Page 127
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2. RELATED WORK The research on congestion for MANET is still in the early stage and there is a need of new techniques. In this section, the research work related to congestion control and multipath routing is presented. Round trip time measurements are used to distribute load between paths in Multi-path Source Routing (MSR) [9]. A distributed multi-path DSR protocol (MP-DSR) was developed to improve QoS with respect to end-to-end reliability [10]. The protocol forwards outgoing packets along multiple paths that are subjected to an end-to-end reliability model. Split Multi-path Routing (SMR) utilized multiple routes of maximally disjoint paths which minimize route recovery process and control message overhead [11-12]. The protocol uses a per-packet allocation scheme to distribute data packets into multiple paths of active sessions which prevents nodes from being congested in heavily loaded traffic situations. Kui Wu and Janelle Harms [13] proposed a path selection criteria and an on-demand multi-path calculation method for DSR protocol. Peter Pham and Sylvie Perreau [14] proposed a multi-path DSR protocol with a load balancing policy which spreads the traffic equally into multiple paths that are available for each source-destination pair. A dynamic load-aware based load-balanced routing (DLBL) algorithm was developed which considers intermediate node routing load as the primary route selection metric [15]. This helps the protocol to discover a route with less network congestion and bottlenecks. When a link breaks because of the node mobility or power off, DLBL provides efficient path maintenance to patch up broken links to help to get a robust route from the source to the destination. A simple Loop-Free Multi-Path Routing (LFMPR) with QoS requirement was developed from AODV and DSR protocol [16]. In the route request phase, intermediate nodes record multi-reverse links which is applied to construct multiple-paths during the route reply phase. Each path is assigned unique flow identification in order to prevent routing loop problems. Rashida Hashim et al [17] proposed an adaptive multi-path QoS aware DSR Protocol. The protocol collects information about QoS metrics during the route discovery phase and uses them to choose a set of disjoint paths for data transmission. De Rango et al [18] proposed an energy aware multi-path routing protocol by considering minimum drain rate as a metric. An update mechanism and a simple data packet scheduling among the energy efficient paths have also been implemented to update the source route cache and for improving the traffic and energy load balancing. Raghavandra et al [19] proposed Congestion Adaptive Routing in Mobile Ad Hoc Networks (CRP). In CRP every node appearing on a route warns its previous node when prone to be congested. The previous node then uses a “bypass” route bypassing the potential congestion to the first non-congested node on the route. Traffic will be split probabilistically over these two routes, primary and bypass, thus effectively lessening the chance of congestion occurrence. Zhang XiangBo and Ki-Il Kim [20] proposed a multi-path routing protocol based on DSR which uses Multi-Probing and Round-Robin mechanisms (MRDSR) for updating alternate multiple paths.
Velammal College of Engineering and Technology, Madurai
The different multi-path schemes proposed earlier concentrates few of the performance aspects. For example, the measurement of jitter is very scarce, which is a critical performance metric for multimedia traffic. Moreover, the performance of a multi-path routing scheme should be quantitatively evaluated in different mobility scenarios to completely assess the applicability. 3. CONGESTION MANAGEMENT The existing DSR protocol was modified to perform multipath routing with congestion adaptive feature. In the proposed protocol, the battery level and queue length are used as the key resource utilization parameters. In the present paper, the congestion was assumed to be in existence when queue length was near capacity or when battery level fell below a predefined threshold. The proposed protocol works similar to normal DSR protocol if the current energy level is between 10 to 100 % of initial energy level and queue length is less than 50. The modified DSR algorithm invokes multi-path routing when it exceeds these threshold values. 4. SIMULATION SET-UP NS-2 version 2.30 [23] was used to simulate the performance of the modified DSR under different mobility scenarios. The 802.11 MAC protocol is defined to be the wireless channel. Each mobile node used a Two-Ray Ground radio propagation model with an Omni antenna. The initial battery capacity of each node assumed to be 100%. The battery capacity of a mobile node was decremented in a predefined manner by the txPower and rxPower levels which remains constant throughout the simulation. The CTS / RTS process used in IEEE 802.11b within NS2 normally generate huge volumes of data that made processing the trace files complicated. In order to restrict the processing of unnecessary data, the RTSThreshold_ variable was set to 3,000 bytes which restrict the processing of packets that have a size greater than 3,000 bytes. The random topology consists of 800 x 800 meter grid with 20 nodes. The packet sizes of CBR multimedia streams were fixed at 512 bytes with each node had a maximum queue size of 50 packets. The simulation time was set to be 400 seconds and the speed of the node is 10 m/s. Three different simulation scenarios of high, medium and low were created by varying pause times at 4, 10 and 20 seconds. The workload and data rate was kept constant at 5 CBR and 200 PPS, respectively. 5. PERFORMANCE METRICS In the present paper, the performance metrics such as average jitter and percentage packet loss was calculated and evaluated for both normal DSR and modified DSR using the following relationships. Average Jitter Jitter is a measure of variation in delay across multiple packets associated with a given traffic flow. In streaming multimedia applications, only a small and limited amount of jitter is tolerable.
Page 128
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Percentage Packet loss % Packet loss = Number of packets sent - Number of packets received / Number of packets sent *100
reduction in percentage packet loss in high mobility than in low and medium mobility scenario further confirms the effective usage of multi-path routing in high mobility scenario.
6. RESULTS AND DISCUSSION 6.1
AVERAGE JITTER
Durkin [22] stated that excessive jitter normally caused by unavailable data may lead to discontinuity in streaming multimedia applications. For example, an excessive jitter might result a video application to experience frozen screens. The computed average jitter values for different mobility scenarios are shown in Figure 1. In low and high mobility scenarios, the performance of modified DSR is only in comparable with original DSR. In medium mobility scenario, the measured jitter for modified DSR is higher than the original DSR.
Fig. 2. Percentage packet loss of original and modified DSR at different mobility scenarios. 7. CONCLUSION In this paper, Multiple Thresholds based Congestion-adaptive Multi-path Dynamic Source Routing for MANET is proposed which avoids the occurrence of congestion and effective distribution of data traffic in an optimal manner. This protocol exhibits substantial reduction in % packet loss and jitter than original DSR. For original DSR, the percentage packet loss lies between 18 20% at low and medium mobility scenarios. This increases to 36.3 % at high mobility. In modified DSR, the percentage packet loss for low and medium mobility scenarios are approximately 13 %, 5 - 7 % lower than original DSR. REFERENCES
The percentage packet loss of original and modified DSR at different mobility scenarios is shown in Figure 2. For original DSR, the percentage packet loss lies between 18 - 20% at low and medium mobility scenarios. This increases to 36.3 % at high mobility. This clearly indicates the effect of route failure at high mobility which causes excessive packet loss. In modified DSR, the percentage packet loss for low and medium mobility scenarios are approximately 13 %, 5 - 7 % lower than original DSR. In high mobility, the percentage packet loss is 20 % lower than original DSR. This drastic
[1] C. S. R. Murthy, B. Manoj, “Ad hoc wireless networks Architectures and protocols,” special edition, Printice Hall, 2004. [2] J. Macker, M. Corson and V. Park, “Mobile and wireless internet services: Putting the pieces together,” IEEE CommunicationsMagazine.36, 2001, pp. 146-155. [3] T. Clausen and P. Jacquet, “Optimized link state routing protocol,” IETF RFC 3626, Network Working Group, October 2003. [4] C. E. Parkens and P. Bhagwat, “Highly dynamic Desination-Sequenced Distance-Vector Routing (DSDV) for mobile computers,”Computer Communications Review 24 (1994) ,234-244. [5] C. Perkins, E. Belding-Royer and S. Das, “Ad hoc ondemand Distance Vector Routing,” RFC 3561, July 2003. [6] D. B. Johnson, D. A. Maltz and J. Broch, “DSR- The dynamic source routing protocol for multi hop wireless ad hoc networks,” in: C. E. Perkins (Eds.), Ad hoc Network, Chapter 5, Addison-Wesley, 2001.
Velammal College of Engineering and Technology, Madurai
Page 129
Fig. 1. Average Jitter of original and modified DSR at different mobility scenarios. 6.2
PERCENTAGE PACKET LOSS
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[7] S. Corson and V. Park, “Temporally-Ordered Routing Algorithm (TORA) Version 1 Functional Specification,” Mobile Ad Hoc Network (MANET) Working Group, IETF, Oct. 1999. [8] C. K. Toh, “A novel distributed routing protocol to support ad-hoc mobile computing,” Proceeding of IEEE 15th Annual International Phoenix Conference on Computers and Communication, 1996, pp. 480-486. [9] Lei Wang, Lianfang Zhang, Yantai Shu, Miao Dong, Multi-path source routing in wireless ad hoc networks, 2000 Canadian Conference on Electrical and Computer Engineering. 1 (2000) 479-483. [10] R. Leung, Jilei Liu, E. Poon, A.-L.C. Chan and Baochun Li, “MP-DSR: a QoS-aware multi-path dynamic source routing protocol for wireless ad-hoc networks”, Proc. 2001 Local Computer Networks, 26th Annual IEEE Conference, 2001, pp.132-141. [11] Sung- Ju Lee and Mario Gerla, “Dynamic load-aware routing in ad hoc networks‟, IEEE Conference on Communications. 10 ,2001,pp. 3206-3210. [12] Sung- Ju Lee and Mario Gerla, “Split Multi-path Routing with Maximally Disjoint Paths in Ad Hoc Networks,” Proceedings of IEEE International Conference on Communications, 2001, pp. 3201-3205. [13] K. Wu and J. Harms, “Performance Study of a Multi-path Routing Method for Wireless Mobile Ad Hoc Network,“ Proceedings of 9th IEEE International Symposium on modeling, Analysis, and Simulation of Computer and Telecommunication Systems MASCOTS'01),2001, PP. 1-7. [14] Peter Pham and Sylvie Perreau, “Multi-path routing protocol with load balancing policy in mobile ad hoc network,” 4th International Workshop on Mobile and Wireless Communications Network ,2002, pp. 48-52. [15] Xiangquan Zheng, Wei Guo, Renting Liu and Yongchun Tian, “A New Dynamic Loadaware Based Load-balanced Routing for Ad Hoc Networks,” IEEE, 2004, pp. 407-411. [16] Cheng-Ying Yang, Yi-Wei Ting and Chou-Chen Yang, “A Loop-Free Multi-Path Routing With QoS for Wireless Ad Hoc Network,” 20th International Conference on Advanced Information Networking and Applications. 2, 2006, pp. 179185. [17] Rashida Hashim, Qassim Nasir, Saad Harous, Adaptive Multi-path QoS Aware Dynamic Source Routing Protocol for Mobile Ad-Hoc Network, Innovations in Information Technology. 2006, pp. 1-5. [18] F. De Rango, P. Lonetti S. Marano, “Energy-aware metrics impact on Multi-path DSR in MANETs environment,“ International Symposium on Performance Evaluation of Computer and Telecommunication Systems, 2008,pp. 130137. [19] Duc A. Tran and Harish Raghavendra, “Congestion Adaptive Routing in Mobile Ad Hoc Networks, “ IEEE transactions on parallel and distributed systems. 17, 2006, pp. 1294-1305.
[20] Hang Xiang Bo and Ki-Il Kim, “Load-aware metric for efficient balancing on multi-path DSR protocol in Mobile Ad hoc Networks,” International Conference on Advanced Technologies for Communications. 2008, pp. 395-398. [21] K. Samdanis, A. Aghvami and V. Friderikos, “Quality of service routing (QoSR) interactions with the BRAIN candidate micro-mobility protocol (BCMP),“ Proceedings of 2nd International Workshop on Mobility Management and Wireless Access Protocols, 2004, pp. 35-42. [22] Durkin, Voice-enabling the data network: H.323, MGCP, SIP, QoS, SLAs, and security, first ed., Cisco Press, Indianapolis, IN, 2003. [23] The Network Simulator NS-2, http://www.isi.edu/nsnam/ns/ [24] N. Reide and R. Seide, 802.11 (Wi-Fi) Networking Handbook, first ed., McGraw-Hill, CA, 2003.
Velammal College of Engineering and Technology, Madurai
Page 130
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Human Motion Tracking And Behaviour Classification Using Multiple Cameras
M.P.Jancy#, B.Yogameena$ Research Scholar, $Assistant Professor Dept. of Electronics and Communication Engineering, Thiagarajar College of Engineering, Madurai.India [email protected], [email protected] #
Abstract- This paper presents a framework for human behaviour understanding using multi-camera system. The proposed methodology classifies human behaviour as normal or abnormal, by treating short-term behaviours such as walking, running, standing still, etc. The analysis includes foreground segmentation, tracking and feature extraction from the image sequences followed by the classification of short term behaviours. A set of calculated feature vectors are input to the Relevance Vector Machine which works on human behaviour classification. Experimental results demonstrate the proposed method to be highly accurate, robust and efficient in the classification of normal and abnormal human behaviours. Keywords - surveillance, feature extraction,, relevance vector machine, behaviour understanding.
I. INTRODUCTION Multicamera surveillance system aims at tracking people or object of interest. The fundamental problem in video surveillance is automatic detection of human motion, tracking human and classifying abnormal situations. Based on this motivation, this paper proposes an efficient approach for tracking people through multicamera system. It creates a model for normal human behaviour and any deviation from the basic model is analysed. Our methodology applies on short term behaviour referring those that can be localized in a spatio-temporal sense, i.e. brief and within a restricted space. Examples of such behaviours are walking, standing still, running, moving abruptly, waving hand etc. The algorithm provides clear discrimination of such anomalies. The paper is organized into five sections. Section II discusses the algorithms available in the present day. Section III describes the methodology Section IV describes background subtraction, extraction of feature and RVM learning system for classification are explained. Section V depicts the experimental results and the classification of human using the Relevance Vector Machine extraction method. Finally, the conclusion is presented.
Machines [4], Radial Basis Function Networks, Nearest Neighbor Algorithm, and Fisher Linear Discriminant and so on. Machine learning techniques are linear methods in a high dimensional feature space, nonlinearly related to the input space. Using appropriate kernel functions, it is possible to compute the hyper plane which classifies the two classes. An automated system based on SVM classifying the human activities as normal or abnormal, utilizes a set of feature vector data has been described by Scholkopf [5]. SVM needs the tuning of a regularization parameter during the training phase, requiring a technique which minimizes the number of active kernel functions to reduce computation time ([6]-[8]). Appropriate features for RVM are used to classify and thereby reduces the computational complexity by selecting an appropriate kernel to improve the classification accuracy. Experimental results demonstrated that the proposed approach is robust in classifying human behaviour. III. PROPOSED METHOD Surveillance system is divided into low level and high level process. The proposed methodology is shown in fig.1 based on the fusion of data collected from several cameras with overlapping fields of view. The low level addresses the problem of motion detection and blob analysis, whereas the upper level addresses the feature vectors. Background subtraction is applied for motion detection to obtain the foreground blob and its corresponding bounding box is extracted. Additional object information namely the object’s blob size, shape, centroid, edge, area, and skeleton are obtained. From the object information the feature vectors are calculated which are input to the Relevance Vector Machine Classifier.Motion-based techniques are mostly used for shortterm activity classification. Examples are walking, running, fighting. These techniques calculate features of the motion itself and perform recognition of behaviours based on these feature values.
II. EXISTING APPROACHES There have been a number of video surveillance systems, for multiple people detection and tracking [1], [2] in a crowded Environment [3]. All these systems can detect and track multiple people. There are many types of neural networks that can be used for binary classification problem, such as Support Vector
Velammal College of Engineering and Technology, Madurai
Page 131
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Video
Segmentation Frame 10
Foreground Blob
Frame 90
Frame 115
Fig.2 Background subtracted image for camera1
Low level High level
Feature Vectors
Frame 05
A. FEATURE CALCULATION Feature vectors are computed by taking into account both the background subtraction and the ground plane information. Object’s centroid is determined by using the following equations
Final Decision [Normal/Abnormal] Fig.1 System Overview
Motion-based techniques are mostly used for shortterm activity classification. Examples are walking, running, fighting. These techniques calculate features of the motion itself and perform recognition of IV. BACKGROUND SUBTRACTION The first stage of video surveillance system seeks background to automatically identify people, objects, or events of interest in various changing environments. Background substraction is accomplished in real-time using the adaptive mixture of Gaussians method proposed by Stauffer [9]. It describes K Gaussian distributions to model the surface reflectance value. k
P ( X t ) = ∑ ω i ,t *η ( X t , μ i ,t , ∑ i, t )
…(1)
where K is the number of distributions, ωi,t is an estimate of Gaussian, μ is the mean, th
∑ i, t
is the
covariance matrix of the i Gaussian. After ordering the Gaussians, the first B distributions are chosen as the background model as shown in eqn (2) b
B = arg min b (∑ ω k ⟩T )
xc = yc =
1 N 1 N
N
∑x j =1
N
∑
j =1
…(3)
i
yi
…(4) where (xc,
yc) represent the average contour pixel position, (xi, yi) represent the points on the human blob contour and there are a total of N number of points on the contour [16]. Boundary extraction is the minimal rectangle that surrounds the borderline. The rectangle’s height is expressed as
Y
max
−Y min
...(5)
And its width is expressed as
i =1
the weight of the i
Frame 15
Fig.3 Background subtracted image for camera2
Relevance Vector Machine
th
Frame 85
...(2)
k =1
where, T is a measure of minimum models for background. The background subtracted image obtained using adaptive GMM is shown in Fig.2 and Fig.3.
X
−X
...(6) Skeleton feature is obtained by calculating the distance max
min
di
from the centroid to contour points. 2 2 d i…(7) = ( xi − xFrom ( yi d− yplot c ) +the c ) the local maximum and i
minimum points are collected and star skeletonization is formed. These points are selected as features and fed into RVM learning system to classify normal and abnormal behaviour of human. B. CLASSIFICATION USING RELEVANCE VECTOR MACHINE The RVM is a Bayesian regression framework, in which the weights of each input example are governed by a set of hyper parameters [8]. These hyper parameters describe the posterior
Velammal College of Engineering and Technology, Madurai
Page 132
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
distribution of the weights and are estimated iteratively during training. Most hyper parameters approach infinity, causing the posterior distributions of the effectively setting the corresponding weights to zero. The remaining vectors with non-zero weights are called relevance vectors. It reduces the “inappropriate” parameters for the classification than SVM [[11], [12]]. In this framework, the attention is focused to the nature of the behaviour of a human to classify his activities as normal or abnormal. The walking pose of a person is considered as an articulated object, and is represented by a parameter vector x, given in eqn (15) x = Wk φ (z) + ξ k …. (15) where, x is the input for the system, Wk is the weight of the basis function, ф(z) is the vector of the basis function and ξk is the Gaussian noise vector. It is used to minimize the cost function of the RVM regression using eqn (16)
Fig.4 Edge image for camera1 and camera 2
Fig.5 Centroid and Bounding Box for camera1 and camera2.
LK = ∑ C K( n ) ( y k(n ) ) S k ( y k(n ) ) ……. (16) T
( )
y k(n ) = x (n ) − W K φ Z (n )
……… (17) where, yk is the output with n sample points belonging to the mapping function k. Φ(z(n)) is the design matrix vector of the basis function, Sk is the diagonal covariance matrix of the basis function and Ck(n) is the probability that the sample point n belongs to the mapping function k. (n)
Fig.6 skeleton feature for human
V.EXPERIMENTAL RESULTS FOR CLASSIFICATION In this work, the video is taken in Indoor and Outdoor . The proposed method processes about 24 frames per second for RGB images and the total number of frames is equal to 3600, and the frame size is 152 × 172 on PC with a 2.5Ghz in a Pentium IV CPU to demonstrate the performance of the proposed method. It classifies the behaviour as normal or abnormal efficiently, by considering the human postures or actions that are explicitly visible over a larger field of view with the multi-camera system. Thus it overcomes the constraint of a single camera system where the human postures or actions may be misinterpreted due to a single field of view which leads to incorrect classification. In this proposed method, the input video file is first converted into frames. Then the Gaussian Mixture Model background subtraction method is used to separate the foreground objects from the background scenes. After extracting the background image, the edges are extracted using the canny filter and its shown in fig.4. The features such as boundary, centroid, area, edge and skeleton are calculated and are shown in fig.5 and fig.6. These features are input to the classifier. Finally the RVM classifier distinguishes the normal and abnormal behaviour.
Velammal College of Engineering and Technology, Madurai
Fig.7 Training stage
Fig.8 Classified RVM output
Page 133
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The Fig.7 shows the initial stage of the training. Then the Fig 8 shows the result after the iteration. The circles in the figure denote the relevance vector obtained after the iteration. The maximum number of iterations used here is Twenty Five. Thus the resultant vector obtained here is closer to the original template and there is no probability of misclassification of vectors. This is the main advantage in this proposed method. Normal behaviour classification is shown in Fig 9. and the datasets were taken from indoor and from caviar benchmark dataset. The algorithm has been tested using 2100 samples. Abnormal behaviour classification of a person bending down and a person waving his hand are shown in Fig 10. . VIII. CONCLUSION This paper has introduced the framework of relevance vector machine for the classification of different kind of human activities for the application of video surveillance. To this end we have classified the given image sequence as standing, running and so on. Results demonstrate the proposed method to be highly accurate, robust and efficient in the classification of normal and abnormal human behaviours. REFERENCES 1. M. A. Ali, S. Indupalli and B. Boufama,” Tracking Multiple People for Video Surveillance”, School of Computer Science, University of Windsor,Windsor, ON N9B 3P4, 2004 2. Mykhaylo Andriluka Stefan Roth Bernt Schiele,”People-Trackingby-Detection and People-Detection-by-Tracking” Computer Science Department Technische Universit¨at Darmstadt,2007. 3. Tao Zhao Ram Nevatia,“Tracking Multiple Humans in Crowded Environment”, Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’04)1063-6919/04 $20.00 © 2004 IEEE. 4. Xinyu WU, Yongsheng OU, Huihuan QIAN, and Yangsheng XU, “Detection System for Human Abnormal Behavior”,2004. 5. B. Scholkopf, J.C. Platt, J. Shawe-Taylor, A.J. Smola, and R.C. Williamson.Estimating the Support of a High-Dimensional Distribution. Neural Computation, 13(7):1443–1471, 2001. 6. Rong Zhang, Christian Vogler, Dimitris Metaxas , “Human gait recognition at sagittal plane” Image and Vision Computing, Vol 25, Issue 3, pp 321-330, March 2007. 7. Carl Edward Rasmussen, “The Infinite Gaussian Mixture Model” Advances in Neural Information Processing Systems, pp. 554–560, 2004. 8. Prahlad Kilambi, Evan Ribnick, Ajay J. Joshi, Osama Masoud, Nikolaos Papanikolopoulos, “Estimating pedestrian counts in groups”, Computer Vision and Image Understanding, pp. 43–59,2008. 9. C. Stauffer, W. Eric L. Grimson. Learning patterns of activity using real-time tracking. IEEE Trans. on Pattern Analysis and Machine Intelligence, Volume 22, Issue 8, pp.747-757, 2000. 10. C. BenAbdelkader, R. Cutler, and L. Davis. “Motion-based recognition of people in eigengait space”, Automatic Face and Gesture Recognition Proceedings, Fifth IEEE International Conference, pp.267 – 272, May 2002 11. C.Chang, C.Lin, LIBSVM: a library for support vector machines, Software available at: http://www.csie,ntu.edu.tw/cjlin/libsvmi, vol. 80, 2001, pp.604–611. 12. D. Kosmopoulos, P. Antonakaki, K. Valasoulis, D. Katsoulas, “Monitoring human behavior in an assistive environment using multiple views”, in: 1st International Conference on Pervasive Technologies Related to Assistive Environments PETRA’ 08, Athens, Greece, 2008.
Velammal College of Engineering and Technology, Madurai
Page 134
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hierarchical Zone Based Intrusion Detection System for Mobile Adhoc Networks. D G Jyothi #1, S.N Chandra shekara*2 #Department of Computer Science and Engineering, Bangalore Institute of Technology, Visveswaraiah technological University, Bangalore, India. 1
[email protected]
*Department of Computer Science and Engineering, SJC Institute of Technology , Visveswaraiah technological University, Chickballapur, India 2
[email protected]
ABSTRACT. Intrusion Detection Systems (IDSs) for Mobile Ad hoc Networks (MANETs) are indispensable since traditional intrusion prevention based techniques are not strong enough to protect MANETs. However, the dynamic environment of MANETs makes the design and implementation of IDSs a very challenging task. In this paper, we present a hierarchical ZoneBased Intrusion Detection System (HZBIDS) model that fits the requirement of MANETs. The model utilizes intelligent light weight mobile agent which collects data from the different mobile nodes(Audit data), preprocess the data, alert and alarm messages among HZBIDS local mobile agents and gateway nodes. With alert information from a wider area, gateway nodes’ IDS can effectively suppress many falsified alerts and provide more diagnostic information about the occurring attacks. The model can adjust itself dynamically to adapt to the change of the external environmental. Also the model is robust and Scalable. Keywords – MANET, ADHOC network, Mobile agents,
MANET requirements which is efficient and flexible and ensures scalability of the IDS. In our proposed model, we use a Local IDA which are attached to every node in the zone and global IDA to the gateway nodes of the zone. The rest of this paper is organized as follows: In the next section, we review some related work in intrusion detection for mobile Ad Hoc networks. In Section 3, we present and explain our intrusion detection scheme.
INTRODUCTION The unique characteristics of Mobile Ad hoc Networks (MANETs), such as arbitrary node movement and lack of centralized control, make them vulnerable to a wide variety of outside and inside attacks [1]. How to provide effective security protection for MANETs has become one of the main challenges in deploying MANET in reality. Intrusion prevention techniques, such as encryption and authentication, can determine attackers from malicious behavior. But prevention based techniques alone cannot totally eliminate intrusions. The security research in the Internet demonstrates that sooner or later a smart and determined attacker can exploit some security holes to break into a system no matter how many intrusion prevention measures are deployed. Therefore, intrusion detection systems (IDSs), serving as the second line of defense, are indispensable for a reliable system. IDSs for MANETs can complement and integrate with existing MANET intrusion prevention methods to provide highly survivable networks [1]. Nevertheless, it is very difficult to design a once-for-all detection model. Intrusion detection techniques used in wired networks cannot be directly applied to mobile Ad Hoc networks due to special characteristics of the networks. In this paper, we are concerned with the design of intrusion detection system for MANET. Our goal is to design a new Hierarchical ZBIDS model, derived from
2.RELATED WORK. An IDS for MANET The Intrusion Detection systems architecture for a wireless ad-hoc network may depend on the network infrastructure itself. Wireless ad-hoc networks may be configured in either a flat or multi-layered network infrastructure. In a flat network infrastructure, all nodes are equal; therefore it may be suitable to civilian applications such as virtual classrooms or conferences [5]. On the contrary, in the multi-layered network some nodes considered different. Nodes can be divided into clusters, with one cluster head in each cluster. In traditional wired networks, many intrusion detection systems have been implemented where switches, routers or gateways play key role to make IDS implemented in these devices easily [3]. But on the other hand these devices do not exist in MANET. For that reason several intrusion detection architectures proposed for ad hoc network which include stand-alone IDS architecture, distributed and cooperative IDS architecture, and hierarchical IDS architecture [4]. In a stand-alone Intrusion Detection System architecture, each node runs an IDS independently to detects malicious attacks and determine intrusions, Since stand-alone IDS do not cooperate or share information with other detection systems running on other nodes , all intrusion detection decisions are based on information available to the individual node. Furthermore nodes do not know anything about the situation on other nodes in the same network [5]. Zhang and Lee [1], proposed a distributed and cooperative intrusion detection architecture that is shown in figure 1, in this architecture each node runs an IDS agent and makes local detection decision independently, by monitoring activities such as user and system activities and the communication activities within the radio range. At the same time, all the nodes cooperate in a global detection
Velammal College of Engineering and Technology, Madurai
Page 135
IDS, Gateway node.
1.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
decision making. [7] In this architecture, if IDS agent on any node detects an intrusion with weak or inconclusive evidence, it can initiate a cooperative global intrusion detection procedure or if a node detects locally an intrusion with strong evidence, it can independently determine an attack on the network. Similarly to stand-alone IDS architecture, distributed and cooperative intrusion detection system architecture is more suitable for flat network infrastructure and not multilayered network. Although distributed and cooperative IDS architecture can solve some limitations that exist in the stand-alone IDS architecture, it has the following problems. First, cooperative intrusion detection may lead to unnecessary communication and calculation between nodes in MANET, causing decrease in network performance. Second, the sharing data between trusted nodes is not in general true since there are a lot of possible threats in a wireless network environment [8].
Figure 1: Distributed and Cooperative IDS in MANETs [1].
Hierarchical Intrusion Detection Systems architectures have been designed for multi-layered ad hoc network infrastructures where the network is divided into clusters. Cluster heads (CH) of each cluster usually have more functionality than other node in the clusters, for example cluster-head nodes centralized routing for all nodes in the cluster and routing packets across clusters. Therefore, these cluster heads, act as manage points which are similar to switches, routers, or gateways in traditional wired networks. Each IDS agent runs on every node. Also it is responsible for detecting intrusion locally by monitoring local activities. A cluster head is responsible locally for its node as well as globally for its cluster, e.g. monitoring network packets and initiating a global response when network intrusion is detected [2,6]. The two-level non overlapping Zone-Based Intrusion Detection System (ZBIDS) is proposed to meet the unique requirement of MANETs. With ZBIDS, the network is divided into non overlapping zones and each IDS agent broadcasts the locally generated alerts inside the zone. Gateway nodes (also called inter zone nodes, which have physical connections to nodes in different zones) are responsible for the aggregation and correlation of locally generated alerts. Furthermore, in MANET intrusion detection system there are two types of decision making including
Velammal College of Engineering and Technology, Madurai
collaborative decision making and independent decision making. In collaborative decision making, each node participates actively in the intrusion detection procedure. Once one node detects an intrusion with strong confidence, this node can start a response to the intrusion by starting a majority voting to determine whether attack happens, like in [1]. On the other hand, in the independent decision making framework, certain nodes are assigned for intrusion detection .These nodes collect intrusion alerts from other nodes and decide whether any node in the network is under attack [7, 5]. These nodes do not need other nodes’ participation in decision making. 2.1 ZBIDS Framework It is obvious that local detection alone cannot guarantee satisfactory performance because of limited security information obtained by each IDS agent. What’s more, we may experience alert flooding problems given the distributed nature of MANETs. Therefore, a suitable framework is needed to integrate the alert information from a wider area. Moreover, attacks are likely to generate multiple related alerts. For example, because of the broadcast nature of radio channel, there may exist many victims suffering from same falsified routing control packets. The triggered alerts should have high correlations correspondingly. Therefore, it is desirable to treat them together. Based on the above considerations, we adopt a non overlapping zone based framework. The whole network is divided into non overlapping zones. We assume the existence of such a framework. This could be done easily based on techniques like geographic partitioning [8]. As illustrated in Fig. 2, there are two categories of nodes in ZBIDS: intra zone nodes and gateway nodes (also called inter zone nodes). If one node has a physical connection to a node in a different zone, this node is called a gateway node, for example, node 4, 7, 8 in Fig. 2. Otherwise, it is called an intra zone node. An agent is attached to each mobile node, and all agents collaboratively perform the intrusion detection task. Each IDS agent runs independently to monitor local node’s activities, such as the user behavior, system behavior, radio communication activities, etc. and performs intrusion detection tasks locally. Intra zone nodes will report their locally generated alerts to the gateway nodes in the same zone, and the gateway nodes will aggregate and correlate the received alerts. Gateway nodes in neighboring zones can further collaborate in order to perform intrusion detection tasks in a wider area. There may Fig. 2. The Zone-Based Intrusion Detection System (ZBIDS) for MANETs.
Page 136
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
exist many gateway nodes in a zone, thus avoiding the issue of single point of failure. 2.2 ) Mobile agents for IDS: Mobile agents have been deployed in many techniques for IDSs in MANETs. Due to its ability of moving in network, each mobile agent is considered for performing just one special task and then one or more mobile agents are distributed amongst network nodes. This operation allows the distributed intrusion detection in the system. Some responsibilities are not delegated to every node, and so it helps in reducing the energy consumption, which is also an important factor in MANET network. It also provides for fault tolerance in such a way that if the network is segmented or some of the agents break down, they can still continue to function. In addition, they can work in big and different environments because mobile agents can work irrespective of their architecture, but these systems require a secure module that enables mobile agents to settle down. Moreover, Mobile agents must be able to protect themselves from secure modules on remote hosts. 3.Proposed Intrusion Detection System In this section we describe the intrusion detection system and procedure for intrusion detection. Fig3. IDS for HZBIDS
To other GIDS
LID S
Response Agent
GIDS
Global Association Agent
Data Association Agent Global Detection Agent Detection Agent
To illustrate the system structure in our design of the HZBIDS, there are two major components for each IDS. Local IDS (LIDS) and Global IDS (GIDS). The LIDS Data collection module collects local data from the different mobile nodes. Local Detection Agent (LDA), will preprocess the audit data received, performs the intrusion detection locally. When it detects the intrusion, it initiates the local alarm. Local Data Association Agent (LDAA), aggregates the detection result from different LDA.LRA (Local Response Agent) sends aggregated results to the GIDS of that zone. Only gateway nodes run the GIDS and they are organized in multiple layers iDs GAA, GDA and GRA perform communication between the LIDS and GIDS. GAA maintains the configuration and status of each agent in the framework. .1 Procedure for Intrusion Detection 1. The local alarm is received by the GIDS as an indication of attack which are recognized and sent by the local agents. 2. GDAA aggregates the detection results received from all other mobile node in the same zone and send the combined result to GDA. 3. GDA will invoke some intrusion detection procedures to know the type of intrusion and determine the attacking node and sends intrusion detection result to GRA and LRA. 4. The GRA will keep the attacking node away from the network with the identity based Cryptosystem scheme [9] and broadcast it to all other nodes in the same zone and with the neighboring gateway nodes. 4. Conclusion We have proposed Hierarchical Zone Based Intrusion Detection System for MANETS. Since it uses hierarchical approach, it can be extended to deal with MANETS of any sizes. This fits distributed feature of MANET. As this model uses gateway nodes, it is supposed to make the adhoc network more efficient limiting the resources. The model represented in this paper represents the effective effort for the development of a sophisticated intrusion detection system for MANETS. As a future work, we plan to develop and improve the model and implement it using NS2 simulator to identify and eliminate the attacks which are vulnerable in MANETS due to open medium. 5. References
Data Collection
Global Response Agent
[1] Y. Zhang and W. Lee, “Intrusion Detection in Wireless Ad Hoc Networks,” the 6th Annual International Conf. on Mobile Computing andNetworking (ACM MobiCom’00), Boston, MA, Aug., 2000, pp. 275-283. [2] Fu, Yingfang; He, Jingsha; Li, Guorui,”A Distributed Intrusion Detection Scheme for Mobile Ad Hoc Networks”, IEEE Computer Software and Applications Conference, 2007.COMPSAC 2007 - Vol. 2. 31st Annual International Volume 2, Issue, 24-27 July 2007 Page(s):75 – 80 [3] Y. Hu and A. Perrig, “A Survey of Secure Wireless Ad Hoc routing”, IEEE Security and Privacy, IEEE press, pp: 28-39, 2004
Audit data
Velammal College of Engineering and Technology, Madurai
Page 137
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[4]Li Y.,Wei J., “Guidelines on Selecting Intrusion Detection Methods in MANET” Proc of ISECON 2004, v 21 [5]H.Deng, R. Xu, J. Li, F. Zhang, R. Levy, W. Lee, "Agentbased Cooperative Anomaly Detection for Wireless Ad Hoc Networks," in Proc. the IEEE Twelfth International Conference on Parallel and Distributed Systems (ICPDS'06), 2006 [6] Y. Huang and W. Lee, “A Cooperative Intrusion Detection System for Ad Hoc Networks”, Proc. 1st ACM Workshop Security of Ad Hoc and Sensor Networks,Virginia, ACM press, pp:135-147, 2003. [7] Anantvalee, Tiranuch and Wu, Jie, A survey on “Intrusion detection in mobile ad hoc networks”. In: Xiao, Y., Shen, X., Du, D.-Z. (Eds.), Wireless/Mobile Network Security, Springer. pp. 170-196. [8] Fu, Yingfang; He, Jingsha; Li, Guorui,”A Distributed Intrusion Detection Scheme for Mobile Ad Hoc Networks”, IEEE Computer Software and Applications Conference, 2007. COMPSAC 2007 - Vol. 2. 31st Annual International Volume 2, Issue, 24-27 July 2007 Page(s):75 – 80 [9] X. Lin and X. Zen, “Mobile Ad Hoc Network -Self-Organization Wireless Network Technology”,Publishing House of Electronics Industry, Beijing, China,2006.
Velammal College of Engineering and Technology, Madurai
Page 138
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Implementing High Performance Hybrid Search Using CELL Processor Mrs.Umarani Srikanth Assistant Professor, Computer Science and Engineering Department, S.A.Engineering College, Affiliated to Anna University Thiruverkadu Post, Chennai, Tamilnadu State, India [email protected]
Abstract— The slowing pace of commodity microprocessor performance improvements combined with ever-increasing chip power demands has become of utmost concern to computational scientists. As a result, the high performance computing community is examining alternative architectures that address the limitations of modern cache-based designs. Multicore processors are an architectural paradigm shift that promises a dramatic increase in performance. But they also bring of complexity in algorithmic design and software development. In this paper the challenges involved in designing the Hybrid Search which is the combination of both Breadth-First-Search algorithm (BFS) and Depth-First-Search (DFS) algorithm for the Cell/BE processor is discussed. The proposed methodology combines a high-level algorithmic design that captures the machineindependent aspects to guarantee portability with performance to future processors. The Hybrid Search helps searching terminology to be made simpler. There are a lot of searching terminologies in practice, which have its own advantages and disadvantages. The performance factor plays a key role in searching algorithms. The Cell broadband engine uses one PPE and eight SPEs thereby making the performance issues to be resolved and helps in faster computations. The Hybrid-Searching algorithm performs well on the multicore processor. This work combines the logics of both breadth first search and depth first search algorithms. It is proved that the parallelization improves the performance of the hybrid searching. Keywords— Multicore Processor, Parallel Computing, Cell Broadband Engine processor, Parallelization techniques, Depth First Search algorithm(DFS) and Breadth-First Search Algorithm(BFS).
I. INTRODUCTION Over the last decade, high-performance computing has ridden the wave of commodity technology, building clusters that could leverage the tremendous growth in processor performance fuelled by the commercial world. As this pace slows down, processor designers are facing complex problems when increasing gate density, reducing power consumption and designing efficient memory hierarchies. Traditionally, performance gains in commodity processors have come through higher clock frequencies, an exponential increase in number of devices integrated on a chip, and other architectural improvements. Power consumption is increasingly becoming the driving constraint in processor design. Processors are much more power limited rather than area limited. Current general purpose processors are optimized for single-threaded
Velammal College of Engineering and Technology, Madurai
workloads and spend a significant amount of resources to extract parallelism. Common techniques are out-of-order execution, register renaming, dynamic schedulers, branch prediction, and reorder buffers. Experiencing diminishing returns, processor designers are turning their attention to thread level, VLIW and SIMD parallelism [1]. Explicitly parallel techniques, where a user or a compiler expresses the available parallelism as a set of cooperating threads, offer a more efficient means of converting power into performance than techniques that speculatively discover the implicit and often limited instruction-level parallelism hidden in a single thread. Another important trend in computer architectures is the implementation of highly integrated chips. Several design avenues have been explored such as AMD Opteron, IBM Power5, SunNiagara, Intel Montecito, Itanium etc[2],[3],[4][5]. The Cell Broadband Engine (Cell/BE) processor, jointly developed by IBM, Sony, and Toshiba is a heterogeneous chip with nine cores capable of massive floating-point processing and is optimized for computationally intensive workloads and broadband rich media applications. The processing power of the single-precision Gflops/second has not passed unobserved.[30],[36]. To fully exploit the potential of multicore processors, we need a significant paradigm shift in the software development process. Unfortunately, for some classes of applications, this implies the radical redesign of algorithms. Together with the initial excitement of early evaluations [6], several concerns have emerged. More specifically, there is an interest in understanding what the complexity is of developing new applications and parallelizing compilers [9], whether there is a clear migration path for existing legacy software [7],[8], and what fraction of the peak performance can actually be achieved by real applications. Several recent works provide insight into these problems [10],[11],[12],[13]. In fact, to develop efficient multicore algorithms, one must understand in depth multiple algorithmic and architectural aspects. The list includes the following: • identifying the available dimensions of parallelism, • mapping parallel threads of activity to a potentially large pool of processing and functional units,
Page 139
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
• using simple processing cores with limited functionalities, • coping with the limited on-chip memory per core, • coping with the limited off-chip memory bandwidth The programming landscape of these advanced multicore processors is rather sophisticated. Many similarities appear with cluster computing: in both fields, we need to extract explicit parallelism, deal with communication, and take care of how threads of computation are mapped onto the physical machine [14],[15],[16]. There are also differences, mostly in the data orchestration between processing cores and main memory, which demand a fresh look at the problem and the design of new algorithms. 1.2 .Graph Exploration Algorithms By a graph search we mean some process that visits each node of a graph (arcs are merely the highways by which we perform the search). These algorithms known as breadth first search (BFS) and depth first search (DFS) are polar opposites. BFS is timid, it does not go ten yards from the den until it has explored everything within five yards. DFS is adventurous, as it goes as far as it can in any one direction before it tries a different direction. Many areas of science demand techniques for exploring large-scale data sets that are represented by graphs. Among graph search algorithms, Breadth-First Search (BFS) is probably the most common and the building block for a wide range of graph applications. For example, in the analysis of semantic graphs, the relationship between two vertices is expressed by the properties of the shortest path between them, given by a BFS search [17], [18], and [19]. BFS is also the basic building block for best-first search, uniform-cost search, greedy search, and A*, which are commonly used in motion planning for robotics [20], [21]. A good amount of literature deals with the design of BFS or dedicated hardware [22], [24]. But no studies have investigated how effectively the Cell/BE can be employed to perform BFS on large graphs and how it compares with other commodity processors and supercomputers. The Cell/BE, with its impressive amount of parallelism [36], [37] promises interesting speedups in the BFS exploration of the many graph topologies that can benefit from it. Searching large graphs poses difficult challenges, because the vast search space is combined with the lack of spatial and temporal locality in the data access pattern. In fact, few parallel algorithms outperform their best sequential implementations on clusters due to long memory latencies and high synchronization costs. These issues call for even more attention on multicores like the Cell/BE because of the memory hierarchy, which must be managed explicitly [38], [39], [40]. Breadth-first search is complete. Breadth-first Search (BFS) [25] is a fundamental graph theory algorithm that is extensively used to abstract various challenging computational problems. Due to the fine-grained irregular memory accesses, parallelization of BFS [41], [42] can exhibit limited performance on cache-based systems. If the goal node is at
Velammal College of Engineering and Technology, Madurai
some finite depth say d, BFS will eventually find it after expanding all shallower nodes. However the time taken to find out a solution is large. Whereas, DFS progresses by expanding the first child node of the search tree that appears and thus going deeper and deeper until a goal node is found, or until it hits a node that has no children. Then the search backtracks, returning to the most recent node it hasn't finished exploring. If the key to be found seems to reside at very high depths, the DFS algorithm may run into infinite looping whereby the searching process becomes incomplete. Breadthfirst search is ideal in situations where the answer is near the top of the tree and Depth-first search works well when the goal node is near the bottom of the tree. Depth-first search has much lower memory requirements. The combination of these two algorithms makes searching in an efficient manner in the sense that it takes the only the merits of both the algorithms and new parallelized algorithm is formulated called Hybrid Search and it is implemented which works very fast in searching the elements using the Cell Broadband Engine (Cell BE) Processor. As the modern processors moving more towards improving parallelization and multithreading, it has become impossible for performance gains in older compilers as technology advances. Any multicore architecture relies on improving parallelism than on improving single core performance. The main advantage of Hybrid Searching is that combining both of these traditional search strategies, one overcomes the disadvantage faced by the other. Experiments show an almost linear scaling over the number of used synergistic processing elements in the Cell/BE platform and compares favourably against other systems. II. MULTICORE PROCESSORS In computing, multi-core processor is a processing system composed of two or more independent cores. One can describe it as an integrated circuit to which two or more individual processors (cores) have been attached. The cores are typically integrated onto a single integrated circuit die [26] (known as a chip multiprocessor or CMP), or they may be integrated onto multiple dies in a single chip package. A many-core processor is one in which the number of cores is large enough that traditional multi-processor techniques are no longer efficient. With so many different designs it is nearly impossible to set any standard for cache coherence, interconnections, and layout. The greatest difficulty remains in teaching parallel programming techniques and in redesigning current applications to run optimally on a multicore system. Multicore processors are an important innovation in the microprocessor timeline. With skilled programmers capable of writing parallelized applications multicore efficiency could be increased dramatically. In years to come we will see much in the way of improvements to these systems. These improvements will provide faster programs and a better computing experience.
Page 140
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
As personal computers have become more prevalent and more applications have been designed for them, the end-user has seen the need for a faster, more capable system to keep up. Speedup has been achieved by increasing clock speeds and, more recently, adding multiple processing cores to the same chip. Although chip speed has increased exponentially over the years, that time is ending and manufacturers have shifted toward multicore processing. However, by increasing the number of cores on a single chip challenges arise with memory and cache coherence as well as communication between the cores. Coherence protocols and interconnection networks have resolved some issues, but until programmers learn to write parallel applications, the full benefit and efficiency of multicore processors will not be attained[28],[29]. The trend of increasing a processor’s speed to get a boost in performance is a way of the past. Multicore processors are the new direction manufacturers are focusing on. Using multiple cores on a single chip is advantageous in raw processing power, but nothing comes for free. With additional cores, power consumption and heat dissipation become a concern and must be simulated before lay-out to determine the best floor plan which distributes heat across the chip, while being careful not to form any hot spots. Distributed and shared caches on the chip must adhere to coherence protocols to make sure that when a core reads from memory it is reading the current piece of data and not a value that has been updated by a different core. 2.1.Moore’s Law One of the guiding principles of computer architecture is known as Moore’s Law. In 1965 Gordon Moore stated that the number of transistors on a chip wills roughly double each year and computer performance will double every 18 months [28]. The graph in Figure 1 plots many of the early microprocessors briefly discussed against the number of transistors per chip.
instructions are pre-fetched, split into sub-components and executed out-of-order. A major focus of computer architects is the branch instruction. Branch instructions are the equivalent of a fork in the road; the processor has to gather all necessary information before making a decision. In order to speed up this process, the processor predicts which path will be taken; if the wrong path is chosen the processor must throw out any data computed while taking the wrong path and backtrack to take the correct path. Often even when an incorrect branch is taken the effect is equivalent to having waited to take the correct path. Branches are also removed using loop unrolling and sophisticated neural network-based predictors are used to minimize the misprediction rate. Other techniques used for performance enhancement include register renaming, trace caches, reorder buffers, dynamic/software scheduling, and data value prediction 2.2 The Need for Multicore Speeding up processor frequency had run its course in the earlier part of this decade; computer architects needed a new approach to improve performance. Adding an additional processing core to the same chip would, in theory, result in twice the performance and dissipate less heat, though in practice the actual speed of each core is slower than the fastest single core processor. Multicore is not a new concept, as the idea has been used in embedded systems and for specialized applications for some time, but recently the technology has become mainstream with Intel and Advanced Micro Devices (AMD) introducing many commercially available multicore chips. In contrast to commercially available two and four core machines in 2008, some experts believe that by 2017 embedded processors could sport 4,096 cores, server CPUs might have 512 cores and desktop chips could use 128 cores[29]. This rate of growth is astounding considering that current desktop chips are on the cusp of using four cores and a single core has been used for the past 30 years[43]. 2.3 . MultiCore Framework The MultiCore Framework (MCF) is an API for programming a heterogeneous multicore chip oriented toward n-dimensional matrix computation. An MCF application consists of a manager program (that runs on the Cell’s PPE) and one or more workers (SPEs on the Cell). MCF’s most important features include the ability to:
To gain performance within a single core many techniques are used. Superscalar processors with the ability to issue multiple instructions concurrently are the standard. In these pipelines,
• define teams of worker processes, • assign tasks to a team of workers, • load and unload functions (called plugins) by worker programs, • synchronize between the manager and worker teams, • define the organization and distribution of n-dimensional data sets through data distribution objects, and • Move blocks of data via channels.
Velammal College of Engineering and Technology, Madurai
Page 141
Fig(1) Moore’s law
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
This last feature is perhaps the most important. It facilitates multi-buffered strip mining of data between a large memory (XDR memory on the Cell) and small worker memories (SPE local store on the Cell) in such a way as to insulate worker code from the details of the data organization in main memory. III.CELL BROADBAND ENGINE ARCHITECTURE The Cell architecture was patented by Ken Kuturagi (Sony) in 1999[30], [31], [32], [33], [34]. He differentiated between software and hardware cells. A software cell is a program with the associated data, and the hardware cell is an execution unit with the capability to execute a software cell. A cell processor consists of a so called PPE (PowerPC Processing Element) which acts as the main processor to distribute tasks (software cells), a MFC (Memory Flow Controller) which interfaces between the computing and memory units, many SPE’s (Synergistic Processing Elements) which are the hardware cells with their own memory. The Cell CPU is essentially a combination of a Power Processor with 8 small Vector Processors (cmp. Co-Processors). All these Units are connected via an EIB (Element Interconnect Bus) and communicate with peripheral devices or other CPU’s via the FlexIO Interface. Unfortunately, the Cell design incorporates hardware DRM features, but the positive aspects outbalance this easily. A single Cell, essentially a Network on Chip, offers up to 256 GFlop single precision floating point performance. A block-diagram of the processor is shown in figure 2.
transistors where 2/3 of them are dedicated to the SRAM(memory). The processor has no branch prediction or scheduling logic, and relies on the programmer/compiler to find parallelism in the code. As the PPE, it uses two independent pipelines and issues two instructions per cycle, one SIMD computation operation and one memory access operation. All instructions are processed strictly in-order and each instruction works with 128 Bit compound data items. 4 single precision floating point units and 4 integer units offer up to 32GOps each. The single precision floating point units are not IEEE754 compliant in terms of rounding and special values. The single precision units can also be used to compute double precision floating point numbers which are compliant to the IEEE754 standard. But their computation is rather slow (3-4GFlops). The memory layout of the SPE is also quite special, each SPE has its own 256kB RAM which is called Local Storage (LS). This SRAM storage can be accessed extremely fast in 128 bit lines. Additionally, each SPE has a large register file of 128, 128 bit registers which store all available data types. There is no cache, virtual memory support or coherency for the Local Storage, and the data can be moved with DMA from/to main memory via the EIB. The Memory Flow Controller (MFC) acts like a MMU in this context and provides virtual memory translations for main memory access. 3.3 .The Element Interconnect Bus
The SPE is essentially a full blown vector CPU with own RAM. Its ISA is not compatible to VMX and has a fixed length of 32 Bit. Current SPEs have about 21 million
The EIB is the central communication channel inside a Cell processor, it consists of four 128 bit wide concentric rings. The ring uses buffered point to point communication to transfer the data and is therewith scalable. It can move 96 bytes per cycle and is optimized for 1024 bit data blocks. Additional nodes (e.g. SPEs) can be added easily and increase only the maximal latency of the ring. Each device has a hardware guaranteed bandwidth of 1/numDevices to enhance the suitability for real time computing. The I/O Interconnect connects the Cell processor to other cell processors. It offers 12 unidirectional byte-lanes which are 96 wires. Each lane may transport up to 6.4GB/s, which make 76.8 GB accumulated bandwidth. 7 lanes are outgoing (44.8 GB/s) and 5 lanes incoming (32 GB/s). There are cache coherent, non coherent links and two cell processors can be connected glueless. The Memory Interface Controller (MIC) connects the EIB to the main DRAM memory, which is in this case Rambus XDR memory which offers a bandwidth of 25.2 GB/s. IV CELL PROGRAMMING The programming of the cell will be as special as the architecture. The big advantage is that there is no abstraction layer between an external ISA and the internal core (cmp. x86) [8]. But the strict RISC design moves the effort to generate optimal code up, to the programmer or compiler. And the general problems of parallel or streaming application development stay the same as for multi-core or multi processor machines. The SPEs are programmed in a direct
Velammal College of Engineering and Technology, Madurai
Page 142
Fig 2. Cell Processor Architecture 3.1. The Power Processing Element
The Power Processing Element (PPE) [36] offers the normal PowerPC (PPC) ISA. It is a dual threaded 64 bit power processor which includes VMX. Its architecture is very simple to guarantee high clock rates. Thus, it uses only in order execution with a deep super scalar 2-way pipeline with more than 20 stages. It offers a 2x32kB L1 split cache, a 512kB L2 cache and virtualization. Altogether the PPE can be seen as a simplified Power processor 3.2 .The Synergistic Processing Element
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
manner, as own autonomous processors with their 256kB Local Storage and 128 Registers. An Assembly language specification is available from IBM but higher level languages such as C/C++ are also supported for application development. The task distribution and allocation of SPEs is fully done in software. The operating system could use them as a shared resource and virtualize them for many tasks (each task sees their own SPEs, in the whole more than available). The PPE is programmed like a standard PPC970 and Linux is running as is (without SPE support, but a patch is available from IBM). The SPEs can be used in different scenarios. A job queue can be used to processes a fixed amount of independent jobs as fast as possible, the SPEs could also be used in a self multitasking manner, as if the cell were a multicore or multi CPU system. Stream processing (Pipelining) is also supported and especially very reasonable for media data (e.g. HDTV). The Local Storage could be used as a cache, but has to be managed by the software for this task. Additionally, MPI support is thinkable, where each SPE is a single node. All these different programming models are just ideas, the future will show which models will be used on the new Cell processors. V.SEQUENTIAL BREADTH-FIRST SEARCH ALGORITHM First the notation employed throughout the rest of this paper and sequential version of BFS is introduced. Then a simplified parallel algorithm as a collection of cooperating shared memory threads is described. Finally the level of detail with a second parallel algorithm that explicitly manages a hierarchy of working sets is described. Given a graph G(V ,E) and a root vertex r Є V , the BFS algorithm explores the edges of G to discover all the vertices reachable from r, and it produces a breadth-first tree rooted at r, containing all the vertices reachable from r. Vertices are visited in levels: when a vertex is visited at level l, it is also said to be at distance l from the root. A graph G = (V, E) is composed by a set of vertices V and a set of edges E. Define the size of a graph as the number of vertices |V|. Given a vertex vЄV, we indicate with Ev the set of neighboring vertices of v such that Ev = {wЄV:(v,w) Є E}, and with dv the vertex degree, i.e. the number of elements |Ev|. Denote D as the average degree of the vertices in a graph, D = vЄV |Ev|/|V|. Given a graph G(V,E) and a root vertex r ЄV, the BFS algorithm explores the edges of G to discover all the vertices reachable from r, and it produces a breadth-first tree, rooted at r, containing all the vertices reachable from r. Vertices are visited in levels: when a vertex is visited at level l it is also said to be at distance l from the root. The pseudo-code description of a sequential BFS algorithm is given below. Breadth-first search (BFS) is a graph search algorithm that begins at the root node and explores all the neighboring nodes. Then for each of those nearest nodes, it explores their unexplored neighbour nodes, and so on, until it finds the goal node[35].
Velammal College of Engineering and Technology, Madurai
Breadth First Search (G, S) Input : A graph G and a vertex. Output: Edges labelled as discovery and cross edges in the connected component. Create a Queue Q. ENQUEUE (Q, S) // Insert S into Q. While Q is not empty do for each vertex v in Q do for all edges e incident on v do if edge e is unexplored then let w be the other endpoint of e. if vertex w is unexpected then mark e as a discovery edge insert w into Q else mark e as a cross edge VI. SEQUENTIAL DEPTH-FIRST SEARCH ALGORITHM The algorithm starts at a specific vertex S in G, which becomes current vertex. Then algorithm traverse graph by any edge (u, v) incident to the current vertex u. If the edge (u, v) leads to an already visited vertex v, then backtrack to current vertex u. If, on other hand, edge (u, v) leads to an unvisited vertex v, then go to v and v becomes the current vertex. Proceed in this manner until reach to "deadend". At this point start back tracking. The process terminates when backtracking leads back to the start vertex. Edges leads to new vertex are called discovery or tree edges and edges lead to already visited are called back edges. Depth-first search (DFS) is an algorithm for traversing or searching a tree, tree structure, or graph. One starts at the root (selecting some node as the root in the graph case) and explores as far as possible along each branch before backtracking. DFS is an uninformed search that progresses by expanding the first child node of the search tree that appears and thus going deeper and deeper until a goal node is found, or until it hits a node that has no children. Then the search backtracks, returning to the most recent node it hasn't finished exploring. In a non-recursive implementation, all freshly expanded nodes are added to a stack for exploration. We will separate the DFS algorithm into four distinct procedures: Get Root, Go Forward, Go Backward and DFS. Note that once the algorithm is underway, there is a current node, C. If a node N has label (x,y) we write V(N) = x and L(N) = y. At any given time the algorithm is designed to go forward (that is further from the root). When it can’t go further it backs up to the prior node and tries to go further down another path. When the algorithm cannot go further and cannot back up then the algorithm looks for another root. When the algorithm looks for another root and fails to find one, it terminates. The DFS algorithm is as follows: Get Root {This procedure is only run when it is called from DFS} If there is a node labeled (0,0) with no antecedent
Page 143
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Then make this node C {it is current} and set V(C)←T; {T for time} Else If there is a node labeled (0,0) {and all such nodes have antecedents} Make it C and set V(C) ←T; Else Stop {algorithm is finished}; T←T+1. {end Get Root} Go Forward {This procedure is only run when it is called from DFS} Iterate2 ← True; Iterate←False While Iterate2 If an arc goes from C to a node labeled (0,0) Then make it C and set V(C)←T; Iterate←True; T←T+1; Else Iterate2 = False; {end Go Forward}. Go Backward {This procedure is only run when it is called from DFS} Iterate←False; L(C)←T If there is an antecedent node Then Find an antecedent node, N, with V(N) = V(C)-1; Set L(C)←T; C←N; {make N current} Else Iterate←False; T←T+1.{End Go Backward}.
for(i=1;i<spe count;i++) { Start:=spe[i]->thread[no of ele , key ,spe id] //invoke other SPEs to search in DFS } Found:=spu_read_out_mbox(); // Read status from outbound mailbox of SPEs. If(Found) Print “Key found ” Else Print “Key not found ” Terminate(spe[i]->thread) exit(0) end
DFS {This is the main procedure} Mark each node (0,0); Set T←0 Repeat Get Root {termination occurs in Get Root} Iterate←True While Iterate Go Forward Go Backward {end DFS}.
algorithm for the SPEs doing DFS create Tree(no of ele/7) dfs search tree(node* root , int key) if key found Found =1 Spu_write_out_mbox(spe id, found); // write the result to the ppe
algorithm for SPE doing BFS create Tree(no of ele); bfs search tree(node*root, int key) If key found // if key found , mark found as 1 Found=1 Spu_write_out_mbox(spe id, found); // write the result to the ppe Else Found =0 Spu_write_out_mbox(spe id, found);
VII. PARALLELIZED HYBRID SEARCH ALGORITHM Here it is proposed to combine two of the most popular searching algorithms used. The disadvantage of one algorithm is overcome by the other and thus making the searching algorithm to be developed in a better way. Parallel algorithm used on these searching techniques make them to run on the different cores thereby making effective utilization of the resources. And also the drawbacks of each of the algorithm are overcome by the other thereby reducing the time taken for the overall throughput. Algorithm for PPU No_of_ele:= read no of elements Spe_count:=read no of spes to be used Key:=key to be searched Start: spe[0]->thread(no of file , key ,spe id) //invoke first SPE to search in BFS
Velammal College of Engineering and Technology, Madurai
Fig3 pictorial representation of hybrid search
Page 144
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VIII. IMPLEMENTATION OF PARALLEL HYBRID SEARCH FOR CELL PROCESSOR To implement the hybrid search The IBM Full System simulator CELL SDK3.1 is used. The Cell Broadband Engine is a generalized simulator that can be configured to simulate a broad range of full-system configurations. It supports full functional simulation, including the PPE, SPEs, MFCs, PPE caches, bus, and memory controller. It can simulate and capture many levels of operational details on instruction execution, cache and memory subsystems, interrupt subsystems, communications, and other important system functions. It also supports some cycle-accurate simulation (performance or timing simulation). Where KEY is the element to be searched against the tree, spe_count is the number of SPEs to be utilized in the searching process, TREE SIZE is the total number of elements to be searched. The PPE determines the tree size for the SPEs performing DFS, depending on two factors namely, the input tree size and the number of SPEs to be utilized. PPE initiates the first SPE which is supposed to search its tree in BFS fashion by sending the key element to be searched. The entire tree is given as input to this SPE as this SPE will search the entire set of elements in BFS fashion. The PPE then initiates all other seven SPEs and it fixes the tree size for each SPE which is the total number of elements to be processed divided by the number of SPEs to be utilized. It then sends these trees to the respective SPE’s and it also sends the key element to be searched to each SPE. The number of SPEs invoked depends on the spe_count entered by the user. The processing of the SPEs starts now. Each SPE creates a binary tree with its own prescribed set of elements. The creation of nodes in the binary tree takes place in breadth first fashion. For the last SPE, the tree size differs slightly as the relation mentioned above doesn’t stand perfectly divisible at all times. So for the last SPE, the quotient obtained from the above relation is summed up with the remainder left out after the division is carried out, so that no element is left out within the tree_size prescribed. The nodes are created and added to the tree in level order fashion. Trees get constructed and stored in each of the SPEs local store. As the SPEs local store memory is very limited, this poses a limit on the number of elements that can be processed by this cell BE architecture. The actual searching process starts after the binary trees are created by all of the SPEs. The first SPE makes use of BFS strategy to search the elements of its tree. The rest of the SPEs search their own trees in Depth first search fashion. This is pictorially represented in the fig 3. Once a SPE finds the key, it immediately writes the key status to its outbound mailbox. The PPE keeps listening to the outbound mailbox of all of the SPEs continuously, The PPE
Velammal College of Engineering and Technology, Madurai
stalls until the outbound mailbox channel remains empty. When this channel gets filled up with the key status reported by any of the SPEs, the PPE immediately reads it and displays the result to the user. Once the key is reported, the PPE immediately terminates all SPE threads, to prevent them from searching further and all resources held are freed up. When the key element is not found by any of the SPEs, the SPEs performing DFS terminates with no signal of response. The responsibility of reporting that the key isn’t found in the tree prescribed is allotted to the SPE performing BFS. This is because BFS is always complete by which there is 100% guarantee of searching all the elements where DFS doesn’t stand complete, there is a possibility of this search running into infinite depths. So if the key to be searched is not found by the SPE performing BFS, then it writes the “KEY NOT FOUND” status to its outbound mailbox, which will be read by the PPE and reported to the user. The main advantage of combining both of these traditional search strategies is that, one overcomes the disadvantage faced by the other. If the key to be searched is found very deep in the tree, then it will be searched and reported by the SPE performing DFS. Where as if the key to be searched is found at very shallow depths, then the SPE performing BFS reports it first, by which time taken for searching the key element is greatly minimized IX.PERFORMANCE ANALYSIS The overall process and control flow of the parallelized and synchronized hybrid search is implemented by using cell sdk
.
3.1 simulator Performance in parallel algorithms is measured in terms of accuracy of the results and speed of execution of the algorithm. In this case the key element is searched parallely by all SPEs using both the graph searching algorithms. Since each SPE is working on its own set of nodes the result is obtained hardly within few seconds as soon as the process is started. The performance graph shown in figure 4 is meant for the key value 500, where the number of elements to be searched is 7000. Here we could see SPE7 doing the BFS, while all other SPEs doing DFS as shown in the figure. X. CONCLUSION Together with an unprecedented level of performance, multicore processors are also bringing an unprecedented level of complexity in terms of software development. A clear shift of paradigm from classical parallel computing, where parallelism is typically expressed in a single dimension, to the complex multidimensional parallelization space of multicore processors, where several levels of control and data parallelism must be exploited in order to gain the expected performance is seen. With this work, it is proved that for the specific case of the Hybrid Search graph exploration, it is possible to tame the algorithmic and software development process and achieve, at the same time, an impressive level of Performance. Thus the performance issues encountered using the existing searching algorithms are overcome by the hybrid search algorithm. The hybrid algorithms work fine and good
Page 145
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
with all the cores of the cell broadband engine. The accuracy of the results is also greatly improved. We can select an application for which this algorithm can be used to find optimal solution and enhance the performance by reducing the running time of the application. Also we can find paths from the root to the solution. We can also add weight corresponding to each edge.
[5]String Searching with Multicore Processors, Oreste Villa, Politecnico di Milano/Pacific Northwest National Laboratory
Daniele
Paolo Scarpazza and Fabrizio Petrini, IBM T.J. Watson Research Center,2009 Published by the IEEE Computer Society [6] S. Williams, J. Shalf, L. Oliker, S. Kamil, P. Husbands, and K. Yelick, “The Potential of the Cell Processor for Scientific Computing,” Proc. ACM Int’l Conf. Computing
Frontiers (CF ’06), May 2006. Fig.4 graph shows functioning of all SPEs . XI.REFERENCES [1] Optimizing a Fast Stream Cipher for VLIW, SIMD, and Superscalar Processors, Craig S.K. Clapp Proceedings of Fast Software Encryption Workshop 1997 [2] R. Kota and R. Oehler, “Horus: Large-Scale Symmetric Multiprocessing for Opteron Systems,” IEEE Micro, vol. 25, no. 2, pp. 30-40, Mar./Apr. 2005. [3]A Brief History of Microprocessors, Microelectronics Industrial Centre, Northumbria University, 2002, http://mic.unn.ac.uk/miclearning/modules/micros/ch1/micro01 hist.html [2] B. Brey, “The Intel Microprocessors”, Sixth Edition, Prentice Hall, 2003 [4]Video Transcript, “Excerpts from a Conversation with Gordon Moore: Moore’s Law”, Intel Corporation, 2005
Velammal College of Engineering and Technology, Madurai
[7]J.A. Kahle, M.N. Day, H.P. Hofstee, C.R. Johns, T.R. Maeurer, and D. Shippy, “Introduction to the Cell Multiprocessor,” IBM J. Research and Development, pp. 589604, July-Sept. 2005. [8]M. Ohara, H. Inoue, Y. Sohda, H. Komatsu, and T. Nakatani, “MPI Microtask for Programming the Cell Broadband Engine Processor,” IBM Systems J., vol. 45, no. 1, pp. 85-102, Jan. 2006. [9] K. Fatahalian, T.J. Knight, M. Houston, M. Erez, D.R. Horn, L. Leem, J.Y. Park, M. Ren, A. Aiken, W.J. Dally, and P. Hanrahan, “Sequoia: Programming the Memory Hierarchy,” Proc. Int’l Conf. High-Performance Computing, Networking, Storage and Analysis (SuperComputing ’06), Nov. 2006. [10] M.C. Smith, J.S. Vetter, and X. Liang, “Accelerating Scientific Applications with the SRC-6 Reconfigurable Computer: Methodologies and Analysis,” Proc. 19th IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’05), vol. 4, Apr. 2005.
Page 146
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[11] M. Guo, “Automatic Parallelization and Optimization for Irregular Scientific Applications,” Proc. 18th IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’04), Apr. 2004. [12] P. Bellens, J.M. Perez, R.M. Badia, and J. Labarta, “CellSs: A Programming Model for the Cell BE Architecture,” Proc. Int’l Conf. High-Performance Computing, Networking, Storage and Analysis (SuperComputing ’06), Nov. 2006. [13] D. Kunzman, G. Zheng, E. Bohm, and L.V. Kale`, “Charm++, Offload API, and the Cell Processor,” Proc. Workshop Programming Models for Ubiquitous Parallelism (PMUP ’06), Sept. 2006. [14]R. Drost, C. Forrest, B. Guenin, R. Ho, A. Krishnamoorty, D. Cohen, J. Cunningham, B. Tourancheau, A. Zingher, A. Chow, G. Lauterbach, and I. Sutherland, “Challenges in Building a Flat-Bandwidth Memory Hierarchy for a LargeScale Computer with Proximity Communication,” Proc. 13th IEEE Symp. High-Performance Interconnects (Hot Interconnects ’05), Aug. 2005. [15] J. Duato, “A New Theory of Deadlock-Free Adaptive Routing in Wormhole Networks,” IEEE Trans. Parallel and Distributed Systems, vol. 4, no. 12, pp. 1320-1331, Dec. 1993. [16]G. Bell, J. Gray, and A. Szalay, “Petascale Computational Systems,” Computer, vol. 39, no. 1, pp. 110-112, Jan. 2006 [17]M.E.J. Newman, “Detecting Community Structure in Networks,” European Physical J. B, vol. 38, pp. 321-330, May 2004. [18]M.E.J. Newman, “Fast Algorithm for Detecting Community Structure in Networks,” Physical Rev. E, vol. 69, no. 6, p. 066133, June 2004. [19] M.E.J. Newman and M. Girvan, “Finding and Evaluating Community Structure in Networks,” Physical Rev. E, vol. 69, no. 2, p. 026113, Feb. 2004 [20]A. Sud, E. Andersen, S. Curtis, M.C. Lin, and D. Manocha, “Real-Time Path Planning for Virtual Agents in Dynamic Environments,”Proc. IEEE Virtual Reality Conf. (VR ’07), Mar. 2007. [21]L. Zhang, Y.J. Kim, and D. Manocha, “A Simple Path Non-Existence Algorithm Using C-Obstacle Query,” Proc. Int’l Workshop Algorithmic Foundations of Robotics (WAFR ’06), July 2006. [22] M. deLorimier, N. Kapre, N. Mehta, D. Rizzo, I. Eslick, R. Rubin, T.E. Uribe, T.F.J. Knight, and A. DeHon, “GraphStep: A System Architecture for Sparse-Graph Algorithms,” Proc. 14th IEEE Symp. Field-Programmable Custom Computing Machines (FCCM), 2006. [23] D. Bader, V. Agarwal, and K. Madduri, “On the Design and Analysis of Irregular Algorithms on the Cell Processor: A Case Study on List Ranking,” Proc. 21st IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’07), Mar. 2007. [24] F. Blagojevic, A. Stamatakis, C. Antonopoulos, and D. Nikolopoulos, “RAxML-Cell: Parallel Phylogenetic Tree Inference on the Cell Broadband Engine,” Proc. 21st IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’07), Mar. 2007.
[25] Efficient Breadth-First Search on the Cell/BE Processor Daniele Paolo Scarpazza, Oreste Villa, and Fabrizio Petrini, IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 19, NO. 10, OCTOBER 2008 [26] W. Knight, “Two Heads Are Better Than One”, IEEE Review, September 2005 [27]R.Merritt, “CPU Designers Debate Multi-core Future”, EETimes Online, February 2008, http://www.eetimes.com/showArticle.jhtml?articleID=206105 179 [28] P. Frost Gorder, “Multicore Processors for Science and Engineering”, IEEE CS, March/April 2007 [29] D.Geer, “Chip Makers Turn to Multicore Processors”, Computer, IEEE Computer Society, May 2005 [30]D. Pham et al, “The Design and Implementation of a First-Generation CELL Processor”, ISSCC [31]P.Hofstee and M. Day, “Hardware and Software Architecture for the CELL Processor”, CODES+ISSS ‟05, September 2005 [32]J. Kahle, “The Cell Processor Architecture”, MICRO-38 Keynote, 2005 [33]D.Stasiak et al, “Cell Processor Low-Power Design Methodology”, IEEE MICRO, 2005 [34]D.Pham et al, “Overview of the Architecture, Circuit Design, and Physical Implementation of a First-Generation Cell Processor”, IEEE Journal of Solid-State Circuits, Vol. 41, No. 1, January 2006 Ma [35]J.B. Orlin, K. Madduri, K. Subramani, and M. Williamson. A faster algorithm for the single source shortest path problem with few distinct positive lengths. Journal of Discrete Algorithms, 2009. [36]Programming the Cell Broadband Engine Architecture Examples and Best Practices: www.redbooks.ibm.com/redbooks/pdfs/sg247575.pdf [37]D.Bader,V.Agarwal and K.Madduri, “On the Design and Analysis of Irregular Algorithms on the Cell Processor: A Case Study on List Ranking,”Proc. 21st IEEE Int’l Parallel and Distributed Processing Symp(IPDPS ’07), Mar. 2007. [38]P.Bellens,J.M.Perez, R.M.Badia and J. Labarta, “CellSs: A Programming Model for the Cell BE Architecture,” Proc. Int’l Conf. High-Performance Computing, Networking, Storage and Analysis (SuperComputing ’06), Nov. 2006. [39]Parallel exact inference on the cell broadband engine processor, Proceedings of the 2008 ACM/IEEE
Velammal College of Engineering and Technology, Madurai
Page 147
conference on Supercomputing,2008 [40]K. Subramani and K. Madduri. A Randomized Queueless Algorithm for Breadth-First Search. Int'l. Journal of Computers and their Applications, 15(3):177{186, 2008.} [41]String Searching with Multicore Processors, Oreste Villa, Politecnico di Milano/Pacific Northwest National Laboratory
Daniele
Paolo Scarpazza and Fabrizio Petrini, IBM T.J. Watson Research Center,2009 Published by the IEEE Computer Society
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[42]Approximating Decision Trees with Multiway Branches, Venkatesan T. Chakaravarthy, Vinayaka Pandit and Sambuddha International Colloquium on Automata, Languages and Programming (ICALP) 2009 [43]B.Bouzas, R. Cooper, J. Greene, M. Pepe, and M. J. Prelle. Multicore framework: An API for programming heterogeneous multicore processors. In STMCS. Mercury Computer Systems, 2006.
Velammal College of Engineering and Technology, Madurai
Page 148
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Phase Based Shape Matching For Trademarks Retrieval B.Sathya BamaLecturer, M.AnithaStudent, Dr.S.RajuProfessor Electronics and Communication Engineering, Thiagarajar College of Engineering Thiruparankundram, Tamil Nadu, India-625015 [email protected],[email protected],[email protected] Abstract: This paper proposes a phase based method for shape matching for retrieving trademarks in digital databases. The proposed approach utilizes phase as a key to develop shape descriptors. The phase combined with magnitude information improves the recognition rate in terms of discrimination power. Experiments were carried out with 30 samples of trademarks and results were examined. Comparative study with existing algorithm shows that the proposed method is highly efficient in elimination of irrelevant matches. Keywords—Shape matching, Phase discrimination, Chord distribution, Radon transform, Chamfer distance transform.
II. INTRODUCTION SHAPE matching is a challenging problem particularly by huge proliferation of images on the Internet, mainly in computer databases containing thousands or millions of images. Applications of shape recognition can be found in Computer Aided Design/Computer Aided Manufacturing (CAD/CAM), virtual reality (VR), medicine, molecular biology, security, and entertainment [1]. Existing approaches can be divided into [1]: Statistical descriptors, like for example geometric 3D moments employed by [2] and the shape distribution [3]. Extensionbased descriptors, which are calculated from features sampled along certain directions from a position within the object [4]. Volume-based descriptors use the volumetric representation of a 3D object to extract features (examples are Shape histograms [5], Model Voxelization [6] and point set methods [7]). Descriptors using the surface geometry compute curvature measures the distribution of surface normal vectors [8]. Image based descriptors reduce the problem of 3D shape matching to an image similarity problem by comparing2D projections of the 3D objects [9]. Methods matching the topology of the two objects (for example Reeb graphs, where the topology of the 3D object is described by a graph structure [10]). Skeletons are intuitive object descriptions and can be obtained from a 3D object by applying a thinning algorithm on the voxelization of a solid object like in [11]. Despite such a long history, interest in the problem has been recently reignited [12] by the fantastic proliferation of
Velammal College of Engineering and Technology, Madurai
electronic imaging and digital color images Although many typical image database algorithms, such as finding two matching or cropped photographs, may rely on a variety of shading, texture, and color attributes, there are significant opportunities for the use of shape as a discriminator. The problem of trademark infringement is a particularly fitting example: most trademarks are high-contrast images from which a shape is readily acquired, and where shading and texture, play only a secondary role, or none at all. Fourier shape descriptors have been studied extensively; however the descriptor phases have been mostly neglected. The phase has either been ignored entirely, leading to highly ambiguous matches, or treated in a simple manner which is reasonable for complex, natural shapes, but which fails for simply-structured shapes common among trademarks [13]. We emphasize that the purpose of this paper is a theoretical analysis of phase in the context of shape matching. We will derive highly-robust phase tests, invariant to shift, rotation, scale, and orientation. However this leaves us with a class of tests, of which the optimum, found through detailed experiments with other existing algorithms, is the subject of ongoing research. The rest of this paper is organized as follows: Section 2 discusses about the background in the context of shape matching. Section 3 describes the implementation of the proposed work. Section 4 gives the experimental results. Finally, Section 5 concludes this paper. III. BACKGROUND IN SHAPE MATCHING A. Phase in Shape Descriptors Given a huge number of binary images, which are to be tested on the basis of shape, there are two, obvious criteria: a) Comparison Speed, and b) Compactness of Representation. Considering two-dimensional shapes bounded by simple, closed curve may be a much more compact description than the huge two-dimensional array of pixels making up the shape.
Page 149
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Indeed, chain codes been developed for such a purpose, in which x(t), y(t) clockwise trace the outline of a shape, where parameter t measures the boundary length from some arbitrary starting point t = 0. Since x(t) and y(t) are clearly periodic, Fourier approaches have long been used in compacting and analyzing shapes. In particular, let
z ( t ) = x ( t ) + iy ( t )
(1) represent the shape outline in the two-dimensional, complex domain where t is not a continuous variable. Using the FFT, z(n) is easily converted to the Fourier domain
f ( k ) = F { z ( n )}
(2) where the complex f(k) are known as the Fourier shape descriptors. The entire challenge, then, of shape matching is the interpretation and comparison of the Fourier descriptors fm(k) from shape m with those of some other shape n.
Chamfer distance transform: The similarity between two shapes can be measured using their chamfer distance [15].
U = {u }n andV = {v }m
i i =1 j j =1 .the Given the two point sets chamfer distance function is the mean of the distances between each point, ui€U and its closest point in V:
d cham (U , V ) =
1 n
∑ min
u i ∈U
|| u i − v j ||
v j ∈V
(4) The symmetric chamfer distance is obtained by adding
d cham (U , V )
.
IV. PHASE BASED FOURIER SHAPE MATCHING Figure 3.1 shows the complete Phase based shape matching for trademark-matching problem. Given some model shape outline z(n), we can consider four perturbations which do not affect the inherent shape: _
z ( n ) = Δ + e i φ rz (( n − δ ) mod N ) (5) where r is scaling,δ is orientation, φ is rotation and Δ is shift and N is the length of the boundary in pixels. Applying the Fourier transform, we find that _
− i 2 πkδ / N
iφ
f (k ) = e {Δ .( k = 0 ) + e rf ( k )} (6) Since z1(n), z2(n) are equivalent shapes, the goal is to determine how to find descriptors invariant to r, φ, Δ, and δ, and thus variant only to inherent changes in shape. Fig 2.1. Sensitivity of shape to the phase information
Because of the sensitivity of the phase (the complex angle component) of f (k) to image rotations and changes in the shape origin (where we define our “start” point for tracing the outline), often only the magnitude components are examined. However the pitfalls of ignoring phase are long established in image processing; indeed, each column in Figure 2.1 has constant magnitude characteristics but with varying phase. The situation becomes progressively worse for more complex shapes, which would include many trademark images, where wildly differing shapes may be naively matched by magnitude comparators. B.
Input trademark Fourier spectrum using FFT Magnitude Normalization Phase Discrimination
ℜ-Transform
The ℜ-Transform to represent a shape is based on the Radon transform. The Radon transform is the projection of the image intensity along a radial line oriented at a specific angle. We assume that the function f is the domain of a shape and its Radon transform is defined by:
TR ( ρ , θ ) =
−∞ −∞
∫ ∫ f ( x, y )δ ( x cosθ + y sin θ − ρ )dxdy
− ∞− ∞
(3)
Selecting common reference
Phase Comparison Fig 3.1. Flow Chart for proposed work
First, the term f(0) is the only one sensitive to Δ, so we ignore it. Next, we normalize to f(1), =
f (k ) =
_
f (k ) _
f (1)
= e − i 2 π ( k −1) δ / N
f (k )
(6)
f (1)
where δ(.) is the Dirac delta-function.
Velammal College of Engineering and Technology, Madurai
Page 150
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
which eliminates the effects of r and φ, leaving us with magnitude independent of the perturbations r, φ, Δ, δ leading to the usual temptation to ignore the phase terms. =
f (k )
| f ( k ) |=
Step 2: Separation of magnitude and phase is performed and magnitude normalization is done.
(7)
f (1 )
But Phase component contributes more for shape information and from Figure 2.1 with varying phase for complex shapes results may be naively matched by using magnitude comparators. If we consider two shapes, z1(n), z2(n) which are identical except for the phase in the qth Fourier coefficient (8) f ( q ) = a∠ φ f ( q ) = a∠ φ 1
1,
2
2
We have to work with phase differences other than those from adjacent Fourier coefficients. So we generalize to normalize out δ for all coefficients. Given a reference index j ≠1, we can create _ =
=
| f (k ) |=| f (k ) | _
=
=
=
∠ f (k ) = ∠ f (k ) − ∠ f ( j). = ∠f (k ) −
k− j j −1
(9)
k− j k −1 ∠f ( j) + ∠f (1) j −1 j −1
where j is selected to be a meaningful reference that is, =
| f ( j ) | | is not small. Thus we have constructed a δ-invariant phase sequence. Given two shapes f1(k), f2(k), the phase comparison thus involves selecting a good common reference j, typically by _
_
=
=
f maximizing | f1 ( j ). f 2 ( j ) | , computing
1
. f
2
and discriminating as _
_
=
∑| f
=
1
_
=
Fig 4.1. Input datasets
Step 3: Phase discrimination is done by three steps which involves selection of common reference j is by • Taking maximum value as Common reference. Step 4: Phase Comparison with the query image by minimizing the discrimination value representing relevant match. In Phase based Fourier Descriptors, a pixel with maximum value in the image is taken as common reference. Results obtained are compared with existing approaches of chord distribution and randon transform. A sample of database and query image with perturbations is shown in Figure 4.2.
_
=
( k ). f 2 ( k ) |( 2 − 2 cos( ∠ f 1 ( k ) − ∠ f 2 ( k ))) (10)
k
Obtaining minimum values in objective function for discrimination represent effective matching of two shapes. 4. EXPERIMENTAL RESULTS Fig 4.2.Database image and perturbed query image
In the experimental activity, a database with 30 samples of different trademarks from internet has been taken. Nine samples in the database are shown in Figure 4.1. Matlab software is used for simulation. The size of the trademark images in database are of 50 X 50 pixels. A.
Fig 4.3 shows the randon transformation of the two input trademarks shown above. It is clearly seen that when the input rotates there is an equivalent amount of difference in frequency domain.
Algorithm
Step 1: Input trademark is read and its Fourier spectrum is obtained using FFT.
Velammal College of Engineering and Technology, Madurai
Page 151
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
B.
Precision, Recall, Efficiency graph
Many different methods for measuring the performance of shape matching algorithm have been created. This paper uses the most common evaluation methods namely, Precision and Recall usually presented as a Precision vs Recall graph. We can always make recall value 1 just by retrieving all images. In a similar way precision value can be kept in a higher value by retrieving only few images or precision and recall should either be used together or the number of images retrieved should be specified. Thus we have plotted two graphs: Precision vs No of retrievals; Recall vs No of retrievals.
Fig 4.3.Randon transforms for above images
Fig 4.6(a).Precision Fig 4.4.Fourier transform. Magnitude normalization and phase comparison
Magnitude comparison shows similar results such that we cannot have complete discrimination. But when we consider the phase information there is a significant change while discriminating inputs based on phase comparison shown in fig 4.4. The discrimination values for three approaches for six sample trademarks have been compared in the Figure 4.5. From the Figure we obtain minimum phase discrimination value using Phase based shape matching compared to other methods. The matching accuracy using phase has been improved twofold times and outperforms existing method in eliminating irrelevant matches.
7
Phase- F F T
5
R and o n t r ansf o r m
4 3 2 1 0 l ogo1
l ogo2
l o go 3
l ogo4
l o g o5
Phase-FFT
0.8
Randon transform 0.6 0.4 0.2 0 6
12
18
24
30
N o . o f R e t r i e v a l s
Fig 4.6(b).Recall
Comparision
6
Recall 1
l o go 6
l ogo7
l o g o8
l og o 9
I np u t t r a d e m a r k
Fig 4.5.Comparision of proposed method with existing algorithm
Velammal College of Engineering and Technology, Madurai
As can be seen in the Fig 4.6 (a), (b) the precision and recall values by Phase based shape matching for different number of retrievals for all the database trademarks are greater than the existing method. Thus it automatically improves the retrieval efficiency and reduces the error rate. V. CONCLUSION A novel phase based shape descriptor for trademark-matching problem has been proposed in this paper. The phase-based discrimination has been implemented and proved to be effective in recognition of trademarks from large databases. The performance has been improved greatly, particularly in the elimination of highly irrelevant matches which appeared in earlier
Page 152
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
ACKNOWLEDGMENT The authors would like to thank the Management and Department of Electronics and Communication Engineering of Thiagarajar College of Engineering for their support and assistance to carry out this work. REFERENCES [23] B. Bustos, D. A. Keim, D. Saupe, T. Schreck, and D. V. Vranic. “Feature-based similarity search in 3d object databases”, ACM Computer Survey, 37(4):345–387, 2005. [24] M. Elad, A. Tal, and S. Ar. “Content based retrieval of vrml objects: an iterative and interactive approach. In Euro graphics Workshop on Multimedia”, . Springer, pages 107–118, New York, 2001. [25] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin. Shape distributions. ACM Trans.Graph., 21(4):807–832, 2002. [26] D. Vranic and D. Saupe. “Description of 3d-shape using a complex function on the sphere”, In International conference on Multimedia and Expo., pgs 177–180. IEEE, 2002. [27] M. Ankerst, G. Kastenmller, H.-P. Kriegel, and T. Seidl. “3d shape histograms for similarity search and classification in spatial databases”. In 6th Int. Symp. on Advances in Spatial Databases, Springer, pages 201–226, London, UK, 1999. [28] D. Vranic and D. Saupe. “3d shape descriptor based on 3d fourier transforms”. In EURASIP Conference on Digital Signal Processing for for Multimedia Com munications and Services, pages 271–274. Comenius University, 2001 [29] J. W. H. Tangelder and R. C. Veltkamp.” Polyhedral model retrieval using weighted point sets”. In Shape Modeling International, pages 119–129, IEEE. Seoul, Korea, 2003. [30] E. Paquet and M. Rioux. Nefertiti: A tool for 3-dshape databases management. SAE transactions, 108:387–393, 1999. [31] C. M. Cyr and B. B. Kimia. A similarity based aspect graph approach to 3d object recognition. International Journal of Computer Vision, 57(1):5–22, 2004. [32] Y. Shinagawa, T. Kunii, and Y. Kergosien. Surface coding based on morse theory. Computer Graphics and Applications, IEEE, 11(5):66–78, September 1991. [33] H. Sundar, D. Silver, N. Gagvani, and S. Dickinson. Skeleton based shape matching and retrieval. Shape Modeling International, pages 130– 139, 12-15 May 2003. [34] T. Bui, G. Chen, “Invariant Fourier-wavelet descriptor for pattern recognition,” Pattern Recognition, 1999. [35] Paul Fieguth, Paul Bloore Adrian Domsa,” Phase-Based Methods For Fourier Shape Matching”, ICASSP,2004 [36] S. P. Smith and A. K. Jain, “Chord distribution for shape matching,” Computer Graphics and Image Processing, vol. 20, pp. 259-271, 1982. [37] Thayananthan, B. Stenger, P. H. S. Torr, and R. Cipolla, “Shape context and chamfer matching in cluttered scenes,” in Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. [38] S. Berretti, A. D. Bimbo, and P. Pala, “Retrieval by shape similarity with perceptual distance and effective indexing,” IEEE Trans. on Multimedia, vol. 2(4), pp. 225-239, 2000. [39] D. Guru and H. Nagendraswam, “Symbolic representation of twodimensional shapes,”Pattern Recognition Letters, vol. 28, pp. 144-155, 2007 [40] N. Alajlan, M. S. Kamel, and G. Freeman, “Multi-object image retrieval based on shape and topology,” Signal Processing: Image Communication, vol. 21, pp. 904-918, 2006. [41] N. Alajlan, I. E. Rube, M. S. Kamel, and G. Freeman, “Shape retrieval using triangle-area representation and dynamic space warping,” Pattern Recognition, vol. 40(7), pp. 1911-1920, 2007. [42] S. Han and S. Yang, “An invariant feature representation for shape retrieval,” in Proc. Sixth International Conference on Parallel and Distributed Computing, Applications and Technologies, 2005.
Velammal College of Engineering and Technology, Madurai
Page 153
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Enhanced Knowledge Base Representation Technique for Intelligent Storage and Efficient Retrieval Using Knowledge Based Markup Language A. Meenakshi1, V.Thirunageswaran2, M.G. Avenash3, Dept of CSE, KLN College of Information Technology, Sivagangai Dist., Tamil Nadu. [email protected],[email protected]
Abstract— Knowledge Engineering is an engineering discipline that involves integrating knowledge into computer systems in order to solve complex problems normally requiring a high level of human expertise. Knowledge Engineering is the technique applied by knowledge engineers to build intelligent systems: Expert Systems, Knowledge Based Systems, Knowledge based Decision Support Systems, Expert Database Systems etc. This system builds a Knowledge Base using Knowledge Base Markup Language (KBML) which is derived from XML architecture. All the Meta information is stored in a KBML file whereas the actual data may be available in any data source. Knowledge sharing is achieved, as this system can retrieve information from heterogeneous data source through this KBML file. This system also provides facilities to search/add the contents to and from the Knowledge Base though mobile phones and Windows Mobile phones without using GPRS. Keywords— Knowledge – Knowledge Base Markup Language – Meta information – data source.
I. INTRODUCTION IVING in a fast moving world, it is natural to expect things faster. Similarly in our quest for data search we need fast and efficient retrieving methodologies. With the evolution of new technology and new products in various domains, researchers are focusing more on exploring new techniques of storage, managing and retrieval of data and knowledge from a repository which has been acquired from various sources. Only having a repository of data or efficiently organizing the data cannot guide decision makers or management to make accurate decisions as humans do. The best approach is to integrate and manage the data in the form of knowledge. Retrieving of exact knowledge through online is increasing and it requires more amount of time to retrieve from different data sources and creating knowledge from information available. Knowledge searching through mobile phones does not exist through SMS. To defeat these anomalies we have built a Knowledge Base using Knowledge Base Markup Language (KBML) which is derived from XML architecture. This system also provides facilities to search/add the contents to and from the Knowledge Base though mobile phones and Windows Mobile phones without using GPRS. The aim of our project is to build a secured intelligent storage mechanism, which can store the information in the form of knowledge using the knowledge
L
Velammal College of Engineering and Technology, Madurai
based representation technique with the help of KBML tags, also the retrieval process is also simplified as it can just refer the KBML file which contains the Meta information about the Knowledge which is going to be stored in distributed data sources. 1.1. EXISTING SYSTEM AND ITS LIMITATIONS In any knowledge based system, the first step is to model the domain knowledge collected from experts, so as to enable effective retrieval of knowledge. Some of the existing knowledge based systems have employed the data structure termed as K-graphs, tree data structure for representing the expert knowledge in their domain of interest. The K-graphs was able to represent the expert knowledge about domains in problem-solving, minimizing the semantic loss that would occur if production rules were used instead. A tree data structure very much resembles a graph, and it can also be defined as an acyclic connected graph where each node has a set of zero or more children nodes, and at most one parent node. Existing Data Structure suffers from the following limitations: y Ambiguous design in storage of data. y Slower performance. y Insecure data storage. y Complexity in retrieving the appropriate data. y Increased Space and Time Complexity. 1 .2 CONVERSION OF META INFORMATION IN KBML This system performs the searching of knowledge across several data sources. The searching facilitates the user to select particular data sources from a list and to get the search result. During the search process, the existence of the search string is first looked up into a KBML file. This KBML file has the Meta information about all the knowledge in a particular data source, the KBML file which is a feature, enabled with derived XML architecture so that our goal of efficient retrieval of exact data is achieved. Information is stored in form of knowledge (KBML Tags). Security is a factor of major concern in any mode of data storage, so we take immense care of this stored information. Here, each knowledgebase is provided with a unique ID, which is stored in an encrypted format. When a search is
Page 154
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
made, the title is first picked and the corresponding IDs are used to navigate for retrieving the description. If the search keyword is not found in the KBML file, the control is passed to search the knowledge in the database for retrieval. End-users may wish to create knowledge when the particular search is not present. In such cases, users should create an account and enter the knowledge along with the title and description. At this stage, a KBML file is created for this newly created knowledge with all its constraints.
1.
ENHANCED KNOWLEDGE BASE REPRESENTATION Our vision of efficient retrieval comes true by the fast fetching of information stored in the form of knowledge. This system builds a Knowledge Base using Knowledge Base Markup Language (KBML) which is derived from XML architecture. All the Meta information is stored in a KBML file whereas the actual data may be available in any data source. Knowledge sharing is achieved, as this system can retrieve information from heterogeneous data source through this KBML file. This system also provides facilities to search/add the contents to and from the Knowledge Base though mobile phones and Windows Mobile phones without using GPRS. 2.1. DERIVATION OF KBML USING XML Extensible markup language (XML) is a fast emerging technical tool. It has several advantages to its name. The limelight feature of “user defined tags” makes this technology worth its salt. The sample structure of KBML is represented below. We use the structure of XML and derive our KBML file which is purely knowledge based. XML has a predefined structure whereas KBML has exclusively user defined structure, this is the major reason why we prefer KBML to XML. Apart from this KBML has a hierarchical structure which superimposes the characteristics of its predecessor XML. The exciting feature that KBML possesses is user can add his own information and by giving it with a unique id, which serves as a Meta information. This forms a part of the knowledge store that can be modified accordingly. Search of information which is our ultimate aim is also made easy.
Fig.1 Structure of KBML
2.2. KBML IN EDAPHOLOGY The KBML file is used to store the Meta information about the soil corresponding to each plant. The edaphology deals with the plant and its classification, here edaphology is a best suited case where we can impart the Knowledge base to represent the details of plants and its description as knowledge and this process is enhanced by the use of KBML file which contains the Meta information of the respective data stored in it. In order to avoid the complexities in using the tree structure we go for KBML. 2. STORAGE 3.1. ADDING THE KNOWLEDGE It is a key feature of the project that makes our database a dynamic one. Knowledge can be represented in any form such as text, documents etc. When a particular search is out of reach of the data sources, users may create their own knowledge to the data source. To facilitate the users, a wizard is designed which contains the simple steps to add the knowledge to the specified Knowledge base and the corresponding data source, provided the users are already registered. The system is also extended to add the user’s knowledge to the knowledge base through Mobile phones. Through the use of this feature our project seems to be “userfriendly”.
Fig. 2 Plant Data Source
3.2. UPDATING THE KNOWLEDGE Day in and day out we have new information. Any knowledge created is not constant, changes are mandatory so as to maintain the efficiency. Changes or updating are done when new techniques are introduced. To carry out these updates, users are allowed to update the existing knowledge that helps
Velammal College of Engineering and Technology, Madurai
Page 155
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
other searchers to gain knowledge. Also we can comment/ question on the particular knowledge of the user.
4.
EXPERIMENTAL RESULTS Users have the option of adding data in the already existing data base. While adding data about a plant the fields like name of the plant, taxonomy, geology should be filled in the appropriate text boxes as given in the form.
Fig. 3 Adding Plant Data Source Fig. 5Adding Description Source
3. RETRIEVAL 4.1 SEARCHING THE KNOWLEDGE it is the main part of the project, by which the data that a user needs is retrieved unto him. It is easier for us to search the relevant information which is available from the selected knowledge base through windows/mobile applications. When the user searches a particular knowledge, they are allowed to select knowledge base and data sources from which they need to search the knowledge from a varied variety of options. All the relevant results are displayed as a list and the knowledge is obtained by navigating to the specified data source.
6.1
CONCLUSION
Development of the Knowledge Base Markup Language (KBML) is used to improve the performance of the search process for a particular knowledge by selecting data sources. Each data source contains a set of knowledge descriptions to serve the requests. Using this, the search can be made efficient through both Windows/ Mobile applications. With world moving too fast all that a user expects is an efficient time constraint and we have strived to impose the same and have achieved a reasonable success. 6.2 FUTURE WORKS Collaborate through mobile: Depending upon the number of search results required, the user will get in their mobile phones. And through mobile system, new users can be connected to the knowledge base. Reception of updates performed (for e.g., daily – at the end of the day, the subscribed user will get the updates in their mobile phones) in the knowledge base. The boom of technology demands such enhancements which if done would be a great success.
REFERENCES [1] “Knowledge Engineering” from http://pages.cpsc.ucalgary.ca/~kremer/courses/CG/. [2] Jay Whittaker, Michelle Burns, John Van Beveren, "Understanding and measuring the effect of social capital on knowledge transfer within clusters of small-medium enterprises" in proceedings of the 16th Annual Conference of Small EnterpriseAssociation of Australia and New Zealand, 28 September – 1 October 2003. Fig. 4 Retrieval
Velammal College of Engineering and Technology, Madurai
Page 156
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[3] Brian D. Newman, Kurt W. Conrad,” A Framework for Characterizing Knowledge Management Methods, Practices, and Technologies" n Proc. of the Third Int. Conf. on Practical Aspects of Knowledge Management (PAKM2000) Basel, Switzerland, 30-31 Oct. 2000. [4] Venkatesh Mahadevan, Robin Braun and Zenon Chaczko,” A Holistic approach of Knowledge management Initiative for a Telecollaboration Business system” in Proceedings of International Business Research Conference, Victoria University of Technology, Level 12, 300 Flinders Street,Melbourne, November 15-16, 2004. [5] Kurt Conrad, Brian Newman,” A Framework for characterizing knowledge management methods, Practices & Technologies”, in proceedings of white paper published in TDAN.com, April 1, 2000. [6] Minsoo shin, “A framework for evaluating economics of knowledge management systems” in proceedings of Information and management, vol.42, no.1, pp: 179-196,2004. [7] Thi Bich Ngoc Pham, “Intra organizational knowledge Transfer process In vietnam’s Information technology companies”. [8] Gail-Joon Ahna, Badrinath Mohana, Seng-Phil Hongb, “Towards secure information sharing using role-based delegation”, Journal of Network and Computer Applications archive, vol 30 , pp: 42-59, 2007. [9]National Treatment Agency,” Confidentiality and information sharing”, September 2003. [10] Ravi Sandhu, Kumar Ranganathan, Xinwen Zhang, “Secure Information Sharing Enabled by Trusted Computing and PEI Models”, Ravi Sandhu and Kumar Ranganathan and Xinwen Zhang, in the Proceedings of ACM Symposium on Information, Computer, and Communication Security (ASIACCS), Taipei, Taiwan, pp: 2-12, 2006. [11] Ganesh Godavari, Edward Chow, “Secure Information Sharing Using Attribute Certificates and Role Based Access Control”, Security and Management, pp: 209-276,2005.
Velammal College of Engineering and Technology, Madurai
Page 157
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Semantic Web Based Personalization Of E-Learning Courseware Using Concept Maps And Clustering Anitha D Department of Computer Science, Lady Doak College, Madurai, Tamilnadu, India 625002 [email protected] Abstract: E-learning has now been regarded as a global learning system with its potential to reach the unreachable. Knowledge based e-learning tools are the key areas of inventions in the present educational scenario to improve the higher education. The recent research works have realized the web services to be invincible in providing the best services to e-learning. These web services are now made more intelligent as Semantic web and have proved themselves to be the best practices that can be adopted in e-learning system and the development of the learning materials. This paper proposes a systematic construction of e-learning courseware based on the semantic web in three phases, domain knowledge base representation, course delivery and evaluation of learners. The approach recommends the conceptual representation of domain knowledge and confirms the effectiveness of concept maps in accessing and delivering it. The paper finds the application of k means clustering and rule based inference to evaluate the learner performance and redesign the course delivery to the learners’ pace. Keywords -- E-learning, Semantic Web, Ontology, Concept Maps, Clustering
I. INTRODUCTION E-Learning, as a package of technology enhanced education tends to replace the standard practices of learning of board teaching and boring lectures. The concept of “Study anywhere, anytime and at anyone’s own pace” has made it a generally acceptable learning system. During the last decade, the introduction of information technology in the educational domain has resulted in a tremendous change in reshaping the method of delivery of academic knowledge. The World Wide Web has increased the intensity of technological play in the development of higher education especially through e-learning by making very fast access to relevant educational resource at any time and place. The web enabled e-learning system enables socially excluded communities to access higher education and connects different societies, communities, resources and learners. The very important resource of e-learning systems is the learning materials and so the development of knowledge based educational resources is the major area of research in the present educational scenario. The e-learning courseware is always expected to satisfy the knowledge needs of different types of learners and provide personalized delivery of academic knowledge. Recent research works in the field of e-learning shows that the main focus is on developing intelligent e-learning systems i.e. Web based systems that are more understandable by machines. The intelligent web based services are now rendered through Semantic web [1][2][5], which has been hottest topic of research now.
Velammal College of Engineering and Technology, Madurai
This paper proposes a method for intelligent semantic web based e-learning courseware development using concept maps and personalization based on clustering. The paper is organized as follows: Section II presents an overview of Semantic web and concept maps and their extensibility to build knowledge base systems , Section III introduces the semantic web model of e-learning courseware design using concept maps and its performance evaluation and Section IV pertains to the future issues regarding the work. II. SEMANTIC WEB AND CONCEPT MAPS - OVERVIEW A. Semantic Web Semantic web is a web architecture i.e. (common-sharedmeaning and machine-process able metadata), enabled by a set of suitable agents. The Semantic web is an extension of the World Wide Web, where content is expressed in a language with enough semantics so software tools can locate exchange and interpret information more efficiently. In the Semantic Web, machines will be able to navigate, integrate and process information in a meaningful way. The semantic retrieval of knowledge will focus searches in a much more concise manner. The Semantic web application will thrive through two key technologies, ontology (formal, explicit specification of a shared conceptualization) as well as reasoning and inference services. Ontologies are introduced to construct, share and reuse knowledge bases contextually which is the crucial application of the Semantic web [1][2][6][7]. They provide a common reference frame for communication. The ontologies are defined and developed using some specific languages among which OWL (Ontology Web Language) [11] being the latest and standardized one. OWL is a language for defining and instantiating Web ontologies. It makes the Web resources more readily accessible to automated processes by adding information about the resources that describe or provide Web content. It provides representation of classes and relationships with its constituent elements “class, instance, datatype property, object property, value and type of values”. OWL represents a domain clearly by defining classes and properties of those classes, defines instances and asserts properties about them, and reasons about these classes and instances to the degree permitted by the formal semantics of OWL. Data property is the relation between instances of classes and the data types of the individual members of the classes. Object property is the specification of relations between instances of two classes. It can assign value to a data property or
Page 158
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
represent an instance property with its syntactic structures. OWL is suggested for defining the domain knowledge as it preserves and enhances the conceptual modeling environment demanded by . They also serve as an educational tool by demonstrating the Semantic Web. the subject concepts with a diagrammatic representation thereby improving the learner’s ability to comprehend. B. Concept Maps Concept map is a flexible and informal form of knowledge Learner assessment could be made more efficient by representation. It is a tool that helps to organize knowledge for mapping the learners’ acquired knowledge into concepts meaningful learning. The concepts are represented in nodes on the and comparing them with the stored concept maps built by map, and the relation between the concepts is represented with experts of the domain. III. PROPOSED MODEL OF E-LEARNING COURSEWARE linking words. The linking words that are used to connect concepts provide meanings for the proposition and hence the DESIGN domain of knowledge. The fundamental components of concept The model is based on semantic construction of learning maps are: nodes (internal and external concepts), categories with materials and meaningful navigation through them, enabled which concepts are qualified (concept types), the graphic aspect with an ontological background. Semantic Web is a very of these concept types, the relations between nodes, the relation suitable platform for implementing an e-learning system, types that qualify those relations. because it provides all means for (e-learning): ontology Earlier studies on concept maps confirm their support for process development, ontology-based personalized and contextual oriented learning and their suitability for context management of course delivery, and rule based learner evaluation. The learning domain [3][7][8]. The reason is that the concept maps proposed e-learning system architecture for personalization reconstruct the knowledge context and span their applicability in of e-learning component is illustrated in fig. 1. representing the domain ontology. They also provide progressive exploration of knowledge and structured placement of documentary resources. Thus, concept maps support a systematic assessment which leads to improving teaching-learning strategies
Fig. 1 Personalized e-learning system architecture
Velammal College of Engineering and Technology, Madurai
Page 159
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
This architecture identifies a user profile database and ontology based courseware database. The implementation of Semantic web based e-learning necessitates the construction of ontology based design of the database. The Course delivery agent is responsible for providing learner interface and contextual course delivery. It is realized with the concept maps to represent the academic knowledge conceptually. Test Agent is responsible for launching intermediary assessment sequence and final testing of the learners. It updates the user profile with their performance grades. It also tests the correctness of the topic sequence in a course and reconstructs the courseware if necessary. The Learner evaluation agent with its rules to evaluate and classify the learners provides the competence level of the learner and tends to update the courseware design and the course delivery in tune with the interest of the learner. Courseware modeling agent is responsible for creating initial model of the course, analyzing the content and selecting the course structure. The modeling agent is accountable for structuring a course according to the current needs of learners. The management agent defines the ontology and constructs the knowledge base of the domain. A. Courseware Ontology construction The e-learning domain can be regarded as a knowledge management system since it deals with academic knowledge and requires its effective management [8]. Ontology construction is always perceived as knowledge acquisition in any knowledge management system. An important aspect of knowledge acquisition is the use of knowledge modeling, as a way of structuring of the acquired knowledge and storing knowledge for future use. A model is often a formalized visual representation of the acquired knowledge. Knowledge is perceived as a concept, and it is strongly argued and agreed that a knowledge unit could be represented in the notion of concepts, because knowledge itself is a conceptual understanding of the human brain. Knowledge models are structured representations of the identified concepts and relationship between them. Recent research works in ontology construction have proved the concept model of domain ontology could be efficiently built up by OWL (Web Ontology Language)[5][6][10][11]. OWL is the recent update of the semantic web world. The representation elements of OWL such as “class, instance, datatype property, object property, value and type of values” provide suitable classes and relationship for the identified domain concepts. Data property is the relation between instances of classes and the data types of the individual members of the classes while the object property is the specification of relations between instances of two classes. The rich syntactic structures of OWL represent a domain clearly and enable the semantic web agents to explore the domain in all its aspects. Protégé, an ontology editor developed by Standford University is chosen as a solution to enter the domain ontology due to its user friendliness and free access. In this regard, an initial attempt of ontology construction has been made for a small topic in the computer science subject, Data Structures: Stack Operation – PUSH. The following concepts and relationships are identified and listed out (Concepts are represented in bold font and the relationships are represented in italics style). The assumed ontology is realized with Protégé and
Velammal College of Engineering and Technology, Madurai
the following OWL classes are created with RDF support. Table I shows the preliminary Data Structures ontology of stack push operation and the relevant OWL-RDF format of the designed ontology. B. Personalized Course delivery using Concept Maps The very important feature that any learning environment demands is that the coherence between the learners’ needs and the learning content. An increase of coherence is achieved through the use of concept maps as educational aids that aim at interpreting, integrating and relating new ideas to existing ones – that are presented by the system upon the request of the learner [7]. Concept maps are proved to be the best implementation model of the semantic web based e-learning systems [4][7][8][9]. Earlier studies on concept maps confirm their support for process oriented learning [8]. The teachers will be constructing concept maps from the domain ontology and establishing relationship among concept maps. Fig. 2 shows the initial mapping of the ontology description to the concept maps. The class, instance, data type, value attributes of the OWL representation get mapped into concepts and the object property gets mapped into links of the concept maps. For example, “Stack has top pointer with initial value 0” identifies the concepts “Stack” as class, “Top pointer” as Datatype property, “Initial value” as Value property, “has toppointer” and “withinitialvalue” as object properties.. So, in the concept maps, “Stack, Toppointer, initial value of 0” are mapped as concepts and “has, with” is mapped as a link.
Fig. 2 OWL - Concept map Mapping
Page 160
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE I A PARTIAL LOOK OF THE PRELIMINARY STACK ONTOLOGY AND OWL REPRESENTATION
GG. HH.
JJ. Sample Ontology representation
Ontology Specification
KK.
(OWL – RDF)
II. LL.
WWW.
MM.
XXX.
NN. Stack ->has: Definition -> is: Data structure in which items are inserted and deleted at the same end(Definition type)
YYY. <owl:versionInfo>$Id: datastructures.owl $ </owl:versionInfo>
OO.
Stack ->Characteristic : Homogenouselements
ZZZ.
PP.
->Characteristic: Last-in-First-out
AAAA.
QQ.
BBBB.
RR. Homogenouselements -> characteristics : Elements ->has same : Datatype
CCCC. DDDD.
SS. Elements->are : operational data TT. Datatype -> types: integer, character, float,
EEEE.
<owl:Ontology rdf:about="">
<rdfs:comment>OWL ontology for Data Structures </rdfs:comment> <rdfs:seeAlso rdf:resource="resource location" />
<owl:Class rdf:ID="Stack"> <rdfs:label>Stack</rdfs:label>
string, records
FFFF.
UU. VV.
Stack -> has : MAXSIZE
GGGG. <rdfs:comment> Data structure in which items are inserted and deleted at the same end
WW.
Stack -> has : Top pointer
HHHH.
XX.
Toppointer -> denotes: ArrayLocation ->of: Insertion
IIII.
YY.
->of: Deletion
</rdfs:comment> </owl:Class>
JJJJ.
ZZ. Toppointer->initialvalue ->zero
KKKK.
AAA.
LLLL.
<owl:ObjecProperrty rdf:ID="hasMAXSIZE"> <rdfs:domain rdf:resource="#Stack" />
BBB.
Stack -> has: Operations
MMMM. </owl:ObjectProperty>
CCC.
Operations -> consists of : Push
NNNN.
DDD.
Operations -> consists of : Pop
OOOO.
EEE. FFF. GGG.
Stack: Push -> does : Insertion Push->has : PushArgument
Velammal College of Engineering and Technology, Madurai
<owl:DatatypeProperty rdf:ID="MAXSIZE">
PPPP.
<rdfs:domain rdf:resource="#Stack" />
QQQQ.
</owl:DatatypeProperty>
RRRR.
Page 161
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
HHH.
PushArgument -> consists of : Elements
III. JJJ.
Stack: Insertion -> adds : Elements
SSSS.
<owl:DatatypeProperty rdf:ID="toppointer">
TTTT.
<rdfs:domain rdf:resource="#Stack" />
UUUU.
</owl:DatatypeProperty>
KKK.
Insertion -> has: PushSteps
VVVV.
LLL.
PushSteps -> startwith : CheckOverflow
WWWW.<owl:FunctionalProperty rdf:ID="Operations">
MMM.
CheckOverflow ->check: overflow -> is : toppointer
XXXX.
<rdfs:domain rdf:resource="#Stack" />
YYYY.
<rdfs:range rdf:resource="#Operations" />
ZZZZ.
</owl:FunctionalProperty>
>= MAXSIZE
NNN.
CheckOverflow -> true : StackOverflowReport
OOO.
CheckOverflow->false: PushStep2
PPP.
PushStep2 -> increments: Top pointer
QQQ.
PushStep2 ->nextstep: PushStep3
RRR.
PushStep3 -> assigns : Pushargument ->
SSS.
AAAAA. BBBBB. <owl:Class rdf:about="#Operations"> CCCCC. arraylocationof :Toppointer
TTT. UUU.
<rdfs:comment>Push operation inserts element into
stack
DDDDD.
</rdfs:comment>
EEEEE. Stack : Push -> implemented by : Reference to C Program
FFFFF.
<rdfs:subClassOf>
code
VVV.
GGGGG.
<owl:type rdf:resource="#Push" />
HHHHH.
</rdfs:subClassOf>
IIIII.
</owl:Class>
JJJJJ. KKKKK.<owl:DatatypeProperty rdf:ID="push argument"> LLLLL.
<rdfs:domain rdf:resource="#Push" />
MMMMM.
</owl:DatatypeProperty>
NNNNN.<owl:Class rdf:about="#Operations"> OOOOO.
<rdfs:comment>Stack has two operations
PPPPP. </rdfs:comment> QQQQQ.
Velammal College of Engineering and Technology, Madurai
<rdfs:subClassOf>
RRRRR.
<owl:type rdf:resource="#Push" />
SSSSS.
</rdfs:subClassOf>
TTTTT.
</owl:Class>
Page 162
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig 3. Partial Concept map for Stack Push operation
Fig. 3 shows a partial concept map for the Push operation in the stack. The learners are listed with the concepts and their relationships to other concepts by means of concept maps and they are shown the parts of the concept map in which they are interested in. They are made to extend their understandability by showing indirectly related concepts on demand. For e.g. a concept map which deals with the basic sorting algorithms may also have a reference to the current applications of sorting along with their efficiency considerations and also may have a reference to an URL or a textbook reference related to Sorting. If the learner is interested in moving to the advanced level of learning, he/she can move through the advanced links of concept maps. Those advanced links of concept maps may be programmed to display only upon the demands of the learner i.e. based on learners’ interest. Similarly, the students are given option to specify the keywords of their concern to extract specific parts of the concept maps. They are also encouraged to build their own concept maps for any topic of their interest and explore the relationships between unrelated concepts. The developed concept maps are verified and added into the content knowledge base thus
transforming the learner capability into a reusable knowledge entry.. The web-enabled courseware design also demands interactive presentations and intermediary assessments of the learner on completion of each topic. C. Rule based learner evaluation and dynamic course delivery using k means clustering The learner assessment is based on his/her response for intermediate test questions in each topic of the learning content. The Test Agent of the proposed method prepares the multi choice objective type test questionnaire with a knowledge evaluation pattern which constitutes three levels of knowledge evaluation: 1. Knowledge (Knowing and defining the concepts (i.e.) what is what?), 2. Understanding (Identifying, classifying, interpreting concepts) and 3. Application (Analyzing and applying existing concepts, arriving and justifying at new concepts).
TABLE II THREE LEVEL CLUSTERING FOR A LEARNER PERFORMANCE
UUUUU.
BBBBBB.
Level
Knowledge
VVVVV. Clusters WWWWW. ry High
Ve
XXXXX. Hig h
YYYYY. Avera ge
ZZZZZ. Lo w
AAAAAA. ry low
CCCCCC. pic numbers
To
DDDDDD.
EEEEEE.
FFFFFF.
GGGGGG.
Velammal College of Engineering and Technology, Madurai
Ve
Page 163
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
HHHHHH. ing
Understand
IIIIII.
JJJJJJ.
KKKKKK.
LLLLLL.
MMMMMM.
NNNNNN.
Application
OOOOOO.
PPPPPP.
QQQQQQ.
RRRRRR.
SSSSSS.
At the end of each topic, the learner is assessed with such a set of questionnaire comprising a minimum of 20 multi choice objective questions supporting all the levels of the given knowledge evaluation pattern. If he/she answers correctly one mark is given, otherwise zero mark is given for each question in a topic. The learner’s response of correct answers is recorded as fraction of 1 in each of these three levels. Based on the fraction of the correct answers scored in each of these 3 levels, the performance of the learner in all the topics is clustered into five major clusters – Very high, High, Average, Low, Very low. The clustering is applied to all the three levels of knowledge evaluation pattern. Table II shows the initial clustering of the learner performance. The learners will definitely fall into any one of these categories for each topic and so the k means clustering algorithm is used to determine the performance rate of a learner where k = 5 indicating the five clusters. Given an initial set of k means m1(1),…,m5(1), which may be specified randomly or by some heuristic, and in this case, the means chosen initially are 0.9, 0.7, 0.5, 0.3, 0.1 with an heuristic approach based on the nature of the clusters respectively. The algorithm proceeds by alternating between two steps: 1. 2.
Assignment step: Assign each observation to the cluster with the closest mean Si(t) = { xj : || xj – mi(t) || ≤ ||xj – mi*(t) || for all i* = 1,…,k } (1) Update step: Calculate the new means to be the centroid of the observations in the cluster.
(2) The algorithm is deemed to have converged when the assignments no longer change. The clusters thus formed are then able to represent a single learner’s behavior in each topic. The three level clusters make the personalization of e-learning more powerful by automatically classifying the learning capabilities of learner in each topic. A set of rules are framed with these cluster information and the rule based inference enables perfect assessment of learner capabilities. Table III shows some of the rules and the inferences made. The entries in the table denote the following: K – Knowledge, U- Understanding, A- Application, VH – Very High, H- High, M – Average, L – Low, VL – Very Low.
Assignment Update
Velammal College of Engineering and Technology, Madurai
Page 164
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE III FEW RULE BASED INFERENCE ON CLUSTERS AND THE ASSESSMENT OF LEARNER BEHAVIOR
TTTTTT.Rule
WWWWWW. ∩ U x L)
Ti in{ K x VH
ZZZZZZ.Ti in{ U x VH ∩ A x L)
UUUUUU.
Inference
XXXXXX. Good in knowledge but low in Understanding level AAAAAAA. very high
Understanding–
BBBBBBB.
Application - Low
VVVVVV.
Assessment
YYYYYY.Needs more presentations and classification works
DDDDDDD. Needs more realtime examples and problem solving exercises
CCCCCCC.
EEEEEEE. ∩ U x VL)
IIIIIII.
Ti in{ K x VL
Ti in{ K x M ∩ U x M)
FFFFFFF. low
Knowledge – Very
GGGGGGG. Very low
Understanding–
LLLLLLL.
i => j
MMMMMMM. performance
Ti affects Tj
RRRRRRR.
i => j
JJJJJJJ. and KKKKKKK. U x L)
Tj in{ K x M ∩
OOOOOOO. U x L)
Ti in{ K x M ∩
PPPPPPP.
and
QQQQQQQ. U x H)
Tj in{ K x H ∩
UUUUUUU.
T1..n in (K x H)
SSSSSSS. Inspite of the average performance in Tj , good performance in Tj
VVVVVVV. High performance in all the topics for the knowledge level
Velammal College of Engineering and Technology, Madurai
HHHHHHH. A very dull learner. Needs more attention. Need More visualized presentations in Ti. Need more worked examples.
NNNNNNN. Learner understanding of Ti must be improved to improve his performance in Tj
TTTTTTT. Swapping of the two related topics suggested for the learner
WWWWWWW.
A consistent learner
Page 165
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
These rules are to be well defined with the collaborative effort of teachers and experts of the domain. Knowledge sharing is needed between them to predict a learner behavior based on the rules. An experimental study was made with 50 students of computer science undergraduate discipline by conducting intermediary tests in 10 basic topics of the data structure subject. Their test results are analyzed and clustered with the k means clustering technique and grouped into different categories. Fig. 4 shows a sample cluster distribution of a student performance in the subject. The cluster distribution statistics of the entire set of 50 students for the knowledge level is shown in fig. 5. From the graph shown, it is made clear that the students could easily make up with definition and introduction of new concepts but it is not easy to show the same response when it comes to the next higher level of applications based on the concepts.
Cluster Distribution 4 No. of 3 topics in 2 the cluster 1 0
VH
H
M
L
VL
Knowledge
4
2
3
1
0
Understanding
3
3
2
1
1
Application
1
2
2
2
3
Identified level of learner Fig. 3 Cluster distributions for a student test data
N o . o f s tu d e n ts o f th e s a m e c l u s te r
Distribution of topics in the student clusters 20 Very high
15
High Medium
10
Low 5
Very low
0 1
2
3
4
5
6
7
8
9 10
Topic No. Fig. 5 Students performance organized in clusters
After finding out the clusters, the rule based approach of assessing them was applied and the results are appreciable fulfilling the intended purpose of this proposed method. For
Velammal College of Engineering and Technology, Madurai
example, topics 4 and 7 which are designed to be higher level of application based on the defined concepts have low performance of students. More detailed presentations and solved problems are needed on such topics. This intended method is sure to open a way to find out the learner capabilities based on the test results and improve their performance. IV. IMPLICATIONS & FUTURE ISSUES The present e-learning scenario demands an extensive research to derive innovative practices in analyzing, constructing and restructuring the e-learning management systems. The system becomes more reliable when it is designed to be intelligent and so the proposed solution emphasizes the conceptual representation and access of the domain knowledge. The learner evaluation, being an inevitable and inseparable part of any courseware design, has to be performed deliberately. Before the implementation of the proposed solution, the rule based inference must be carefully designed by the experts of learning domain. The current study of the e-learning courseware design has marked the courseware content sequencing to play an important role in affecting learner behavior. A case study of content sequencing [6] has shown the incorrect content sequencing to be a reason for the decrease in performance of learners during the intermediary assessment of the misplaced topics. Recent research works in learner evaluation demands the inclusion of the time spent by the learner in each topic, the participation of the learner in discussion forums, the number of suggestions or clarifications raised by the learner and the assignment performance. The proposed solution may be extended to study the impact of lesson sequencing on the learners’ performance and the learner evaluation including the identified factors and the future issues pertain to the specified findings. V. CONCLUSION Effective technology is very important for the sustainable development of education. This paper finds a machine intelligent approach for constructing e-learning courseware. It identifies three phases related to the courseware design and discusses the need of conceptual representation to simulate the human brain. It also explains the technical implementation of the three phases along with a case study on the third phase of learner evaluation. The knowledge gained during learning is the most important factor for measuring the quality of education and it can be measured with the intermediary tests. This paper may be an opening for the realization of web based intelligent educational tools for effective utilization of academic knowledge. REFERENCES [1] Alsultanny Y, E-Learning System Overview based on Semantic Web , The Electronic Journal of e-Learning, Volume 4 Issue 2, 2006, pp 111 - 118, Available online: www.ejel.org [2] Fayed Ghaleb, Sameh Daoud, Ahmad Hasna, Jihad M. ALJa’am, Samir A. El-Seoud, and Hosam El-Sofany, E-Learning Model Based On Semantic Web Technology, International Journal of Computing & Information Sciences Vol. 4, No. 2, Pages 63-71, August 2006 . [3] Hsinchun Chen, Ann Lally, Byron Marshall, Yiwen Zhang, Convergence of Knowledge Management and E-Learning: the GetSmart
Page 166
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Experience, Paper presented in the 3rd ACM/IEEE-CS joint conference on Digital libraries, 2003, ISBN:0-7695-1939-3. [4] F. P. Rokou et al.,Modeling web-based educational systems: process design teaching model, Educational Technology and Society, Vol. 7, pp. 4250, 2004 [5] Ungkyu Park, Rafael A. Calvo, Automatic Concept Map Scoring Framework Using the Semantic WebTechnologies, Proceedings of Eighth IEEE International Conference on Advanced Learning Technologies, DOI 10.1109/ICALT.2008.125, July 1-- July 5, 2008,, Spain. [6] Chih-Ming Chen1 and Chi-Jui Peng2, Personalized E-learning System based on Ontology-based Concept MapGeneration Scheme, Seventh IEEE International Conference on Advanced Learning Technologies (ICALT 2007), 0-7695-2916-X/07, July 18-20, 2007, Japan. [7] Thanasis Giouvanakis,Garyfallos Fragidis,Eyaggelos Kehris,Haido Samaras, Exploiting Concept Mapping in a Semantic Web Environment, Proceedings of the Fifth IEEE International Conference on Advanced Learning Technologies (ICALT’05), 0-7695-2338-2/0, July 5-8, 2005, Taiwan. [8] Marco Pedroni, E-learning and Knowledge Management: Context Structuration, Paper Presented at the Informing Science and IT Education Joint Conference, 2007, 22-25 June 2007, Ljubljana, Slovenia. [9] Juha Puustjärvi, Leena Puustjärvi, Using Semantic Web Technologies in Visualizing Medicinal Vocabularies, IEEE 8th International Conference on Computer and Information Technology Workshops, 978-0-7695-3242-4/08, DOI 10.1109/CIT.2008. [10] Srimathi H, Knowledge Representation in Personalized ELearning, Academic open Internet Journal, Volume 23, ISSN 1311-4360 , 2008. [11] W3C Recommendation 10 , OWL Web Ontology Language Guide”, February 2004.
Velammal College of Engineering and Technology, Madurai
Page 167
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Adaptive visible watermarking and copy protection of reverted multimedia data S.T.Veena#1, Dr.K.Muneeswaran*2 # PG Graduate, * Professor & Head Computer Science and Engineering Department, Mepco Schlenk Engineering College Sivakasi, Tamilnadu, India 1 [email protected] [email protected]
2
Abstract —Visible watermark is traditionally used to authenticate the ownership of the media unambiguously and is irreversible. However it is desirable in some applications like medical imagery, remote sensing, to have a visible watermark, to authenticate the media, at the same time hide the important details from being accessible to outsiders. It is also known that even small alternation to the media in such applications may be disaster. So a reversible visible watermarking scheme which combats copyright piracy, at the same time losslessly recovers the originals back by authorized person is proposed. It is not possible in such case to have original watermark at the retrieval end. And thus it has to be blind. In case of applications like artwork preserving and press, the same can be used against piracy of the retrieved media by embedding a buyer authentication code in the retrieved media with negligible changes. In case of copy protection violation, it can be used to track to the buyer involved in piracy. The digital media may be gray scale image, color image or a video. The watermark is a binary image. The watermarking application uses visible embedding method for embedding watermark and reversible data embedding (invisible) method for recovering the digital media (original and watermark) losslessly and blindly. The visible watermark is embedded in the spatial domain such that it is not so obtrusive of the beneath host signal, while clearly authenticates the signal. So an adaptive (host dependent) scaling factor found by exploiting the HVS properties is used for embedding the watermark in user-specified region of the host media. In order to achieve reversibility, a reconstruction/ recovery packet, is constructed, compressed and embedded in the nonwatermarked region of the host media using a reversible data embedding technique. Buyer authentication is done by generating the hash value and embedding it reversibly in the watermark area. Retrieval and verifying against buyer authentication code helps to track to that buyer. Keywords — Blind watermarking, Human visual system HVS, reversible data embedding, visible watermarking, data compression, Authentication code, Copyright piracy.
I. INTRODUCTION Conventional visible image watermarking schemes [1],[2] impose strict irreversibility. However
Velammal College of Engineering and Technology, Madurai
certain applications like medical imagery, prepress industry, image archival systems, precious artworks, remotely sensed images require them to be removable and lossless. Various reversible
schemes have been developed using the techniques of modulo arithmetic, the circular interpretation of the bijective transform, wavelets and sorting or difference expansion [3]-[5].They are unsuitable to visible watermarking, because the embedding distortion inflicted by a visible watermark is often far greater. Achieving lossless recovery of the original host signal from a visibly watermarked signal is an acute challenge. The prior works in this area [6]-[8], need the original watermark for original image recovery. Moreover, all the existing methods do not consider Human Visual System (HVS) characteristics in the visible watermark embedding process. As a result, they are less visually satisfactory and more intrusive. In paper [9], the binary watermark is embedded spatially in the host image in user specified region using adaptive scaling factor. This factor utilizes HVS characteristics of the host image and varies with region. Reversibility is achieved by hiding a recovery packet in the non-watermarked region of the host using LSB hiding of pixels after applying simple integer transform (Reversible contrast mapping). The recovery packet is the difference between the watermarked image and the approximated image in the watermarked region. To reduce the payload data to be hidden, the packet is encoded and compressed. In this paper we extend the ideology of the prior work to color images, video and authenticate it to buyer in case of piracy. The rest of the paper is organized as follows. In Section II, the embedding process of the proposed reversible visible watermarking algorithm for color image is presented in detail. Section III briefly introduces the watermark removal and lossless media recovery. Section IV discuss the process of identifying the buyer to piracy. Section V provides the experimental results for evaluating
Page 168
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the performance of the algorithm. The conclusion is given in Section VI.
II. WATERMARKING PROCESS Here the embedding process involves separation of the luminance part of the media and data embedding. This is done because humans are more sensitive to the black-and-white information than to other color components (more cones than rods in our vision system) A. Separating Luminance Part To separate the luminance (luma) part of the image which is a grayscale version of the image, it is converted from RGB color space to YCbCr color space. The Red Green and Blue components are converted to its respective Luminance/luma (Y/Y’), blue difference chroma and red difference chroma components. Once luminance part is separated, the watermark is translucently overlaid in the spatial positions defined by the user in the ROI of it. The watermarked luma component is again combined with Cb and Cr components. Then the image from this color space is converted back to RGB color space. B. Data Embedding The data embedding process involves two main procedures: visible watermarking and invisible reversible data hiding. The details of the processes are given in the following section. 1) Visible Watermarking: Watermarking is either additive or multiplicative. This is a multiplicative watermarking scheme where the change is done in the host where the corresponding overlaying watermark has a pixel value zero. This is chosen because the binary mark is chosen such that white represents 1 and black zero and the watermark is in white backdrop. Let N = no of 8*8 blocks of Y component of image I S ⊂ {1, 2, . . . , N} be the set of block ID numbers corresponding to the 8×8 blocks in ROI. ⎧ ⎣α n × Y n (i , j )⎦, if W n ' = 0 ⎫ Y nw (i , j ) = ⎨ ⎬ if W n ' = 1 ⎭ ⎩ Y n (i , j ), 1 ≤ i, j ≤ 8 and n ∈ S (1)
…
and
Ynw(i, j) = Yn (i, j) , 1 ≤ i , j ≤ 8 and n∈{1,2 ,K, N }− S …(2) where Ynw and Yn is the (i,j) th spatial pixel value in the nth 8 × 8 block of Yth component of Image I and watermarked Y component of Iw. ⎣•⎦ is the represents the mathematical floor function, In of the watermarked and αn is the adaptive scaling factor for the nth block of I. Wn′ is the n′th position in
Velammal College of Engineering and Technology, Madurai
the watermark. It is given n′ for watermark since the absolute spatial positions of both watermark and image are different. 2) Finding Adaptive Scaling Factor: According to HVS, the human eye is sensitive to mid luminance area while less sensitive to textured region. Exploiting this fact, instead of having a constant scaling factor for embedding, it can be chosen in such a way that the scaling factor depends on the underlying host image. Higher the texture, lower the scaling factor and more of luminance, high the scaling factor is selected. It is that discrete cosine transform actively captures this information through its ac and dc components. The dc coefficients of the DCT domain are best represented by Gaussian distribution [11] and ac coefficients by [12]. The adaptive scaling factor is therefore determined as per equation 3 and values are adjusted to be in the range [0.84 0.89]. This is done because values >0.89 tend to move towards invisibility and < 0.84 tend to be obtrusive.
⎧− [Y (1 , 1) − μ ]2 ⎫ ˆ exp⎨ n 2 ⎬ +ν n 1 ≤ n ≤ N 2 σ 2 ⎩ ⎭ 2πσ 1
αn = μ=
1 N
σ2 =
vˆn =
L(3)
N
∑Y (1 , 1) n
is the mean of
dc coef
and
n =1
1 N
2
N
∑[Yn (1, 1) − μ] is
the variance
n =1
vn − min n (vn ) max n (vn ) − min n (vn )
is the normalised log arithm of vn
vn = ln(vn ), 1 ≤ n ≤ N vn =
1 [Yn (i , j ) − η n ]2 × 63 (i , j )≠(1,1)
ηn =
1 × Yn (i, j ), 63 (i , j )≠(1,1)
∑
∑
1≤ i, j ≤ 8 and 1 ≤ n ≤ N
2) .
3) Facilitating Recovery:The watermark and the Image must be recovered blindly at the detection/extraction end. To facilitate this, the difference information must be provided along with the watermarked image. This will have a great payload. Hence an approximate image is generated of the original. This is done by first finding the approximate adaptive scaling factor. The factor is determined by generating an approximate original image using prediction of pixels 4) Pixel Prediction : This prediction method uses neighboring non-watermarked pixels to predict the original
Page 169
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
value of watermark pixel. It is an iterative process. It involves choosing overlapping windows of size 3×3. The window is chosen such that the pixel to be recovered (watermarked pixel) is at the center. The new pixel intensity of that center pixel of the window is the average of non-watermarked pixels and recovered watermarked pixels. The window is moved across the image in raster scan and the recovered pixels are used in subsequent windows for further recovery in that window. 5) Approximate Scaling factor and Image Generation : With this approximated version of the image P, find the adaptive scaling factor αˆ using equation 3 as described above with the original image Y. Then use this to get an approximate version of the original Ya by reversing the embedding process. ⎫ ⎪⎪ ⎬ ⎪ ⎪⎭
⎧ ⎢ P w (i, j ) ⎥ ⎪⎪ ⎢ n ⎥, if Wn (i, j ) = 0 Yna (i, j ) = ⎨ ⎢⎣ αˆ n ⎥⎦ ⎪ w ⎪⎩Pn (i, j ), if Wn (i, j ) = 1 1 ≤ i, j ≤ 8 and n ∈ S
L(4)
L (5 )
6) Construction of Recovery Packet : A recovery packet is constructed as a difference between the watermarked image and the approximated image of the watermarked area as in (4)
D = (Y − Ya)ROI
C. Invisible reversible data embedding The invisible hiding process uses reverse contrast mapping [10], an Integer transform, to embed the data in the LSB using the transforms. It inverts exactly back even when LSB is lost/ changed, except in case of odd pixels, which is also taken care. This algorithm even though has low embedding capacity (yet satisfies requirement) than other reversible data embedding mechanisms has less arithmetic complexity and is fast. The transform processes are given as Forward transform is … x′ = 2x − y; y′ = 2 y − x (8) Reverse transform is
Yna (i , j ) = Pnw (i , j ),
1 ≤ i , j ≤ 8 and n ∈ {1, 2, K , N } − S
After encoding, the packet is compressed using simple runlength encoding. For the purpose of security, the encoded and compressed (De) is exclusively ored (XOR) with a pseudo random numbers of a secret seed before inserted into the non-ROI region.
… (6)
7) Compression and encoding: Once the recovery packet D is constructed, the values are compressed to reduce the payload. The payload is then invisibly hidden in the nonwatermarked region. To facilitate compression, the values of D are encoded in L bit 2’s complement binary form where L-1 is maximum number of bits needed to represent D as per equation (7). For example we have the difference in the range -4 to 2 pixel intensity. The maximum of the absolute value of pixel intensity is 4. Therefore maximum bits needed for coding this is 3 bits (100). Assign 4 bits(L) one for sign (ie) MSB or Lth bit is 1 if the number is negative else 0. Thus in example -4 is assigned as 1100 and +4 as 0100. This helps in recovering signed number after decompression. Similarly zeroes are specially coded as those belonging to watermarked pixel (i , j) ∈ Γ and those that do not (i , j) ∈ Γ ′. The former zero is represented with MSB 1 while the latter with 0. ⎧⎪ ( L−1) − D(i, j ) where D (i, j )≤0 and (i, j ) ∈ Γ⎫⎪ = 2 ⎬ … (7) De(i, j ) ⎨⎪D (i, j ) ⎪⎭ otherwise ⎩ (i, j ) ∈ Γ∪Γ′
Velammal College of Engineering and Technology, Madurai
2 ⎤ ⎡ 2 (9) 1 ⎤ ⎡1 x =… y = ⎢ x′ + y′⎥ ⎢ 3 x′ + 3 y′⎥ 3 ⎥ ⎢ ⎥ ⎢3 where (x,y), (x′,y′) are the chosen pixel values in the range [0-255] to prevent overflow and underflow conditions. III.
RECOVERY AND DECODING PROCESS The process of recovery of original and watermark extraction to prove ownership is the reverse of the operations involved in embedding. The process is blind and the keys used in reconstruction are the top-left position of ROI, size of watermark, size of payload, auxiliary information of compression, size of encoding length L, secret key – seed of pseudo-random generator. First, transform the RGB to YCbCr color space. Second, extract the invisibly embedded bits from LSB by using inverse transform. Third, using secret key (seed) reproduce the compressed payload back (recovery packet). Next, decompress the payload to recover the encoded reconstruction packet. The packet is decoded by taking the 2’s complement of L-bit number. ⎧2L−1 − D(i, j) if MSB is 1⎫ D′(i, j) = ⎨ ⎬ where (i, j) ∈ Γ ∪ Γ′ otherwise ⎩D(i, j) ⎭ … (10)
⎧1 if MSB is 0 and D′ (i, j)=0⎫ W(i, j) = ⎨ ⎬ where (i, j) ∈Γ∪Γ′ ⎩0 otherwise ⎭ …
(11) Apply pixel prediction method to the watermarked image after payload is removed, as done previously to estimate a scaling factor and use it to get an approximated version of original image Y. Now to recover original pixel values in
Page 170
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
watermarked region add recovery packet to the approximated version to get back the originals. Transform back to RGB color space from YCbCr color space. In case of video, the above process is repeated for each frame with its correspondent keys to recover back the original video. IV. BUYER AUTHENTICATION The copy protection does not stop with recovery of original and selling the media to an authorized buyer. But it is in some applications necessary to keep track of piracy even after sales. The following ideology helps to track down to the buyer in case of piracy. A. Embedding hash code To accomplish this, first a hash value is generated for say identifier code. Save the length of the digest, it serves as the secret key. Secondly, the hash value is reversibly hidden into the region of the media (decided by the seller) using the RCM integer transform, a LSB technique, with the modification that instead of using all pixels in the selected region, embedding is done only on the odd pixels belonging to the transform domain (Dc).This helps in reverting both the pixel and the digest. The reverted digest can be used to track to the buyer responsible for the piracy. Let the identifier code be concatenation of sales id, date and some other details. Apply hash can be SHA-1 or md5 as application demands and this can be kept as a secret. The hash bits are then embedded into a region of the recovered media by utilizing modified RCM technique. B. Hash checking Regenerate the hash code with the identifier code stored in database with its key. Check it with the one retrieved. If it matches then the corresponding buyer is involved in copyright infringement else proceed with other identifier code until it can be tracked to the buyer. IMPLEMENTATION AND RESULTS The proposed method has been implemented and tested on various media (images (webber database) & video) different formats & sizes with different watermarks in different ROIs. The perceptually inspired metrics PSNR, WPSNR, Structural Similarity (SSIM) are used to access the scheme. Figures 3 and 4 shows the watermarking results for image and video sequences. Table I and II show the performance of watermarking of various images. PSNR-1 is watermarked with original except ROI. PSNR-2 is for watermarked with original. PSNR-3 is for original with retrieved.WPSNR-1 is weighted PSNR using CSF function for watermarked with original.WPSNR-2 is that with retrieved and original.WPSNR-3 is between original and retrieved in ROI.SSIM-1,-2 are structural similarity between original and watermarked & original and retrieved.SSIM-3 is between original and retrieved in ROI In embedding the visible watermark in different ROIs, it was noted that there is a difference in watermark visibility. The watermark is more visible in smooth image
areas. To access the visual quality of the visible watermark, perceptually insipired merics WPSNR and SSIM are used and results are tabulated in Table I. From PSNR-2 it can be seen that image with large smooth regions (LENA & F-16) has the lowest PSNR. This is because of the large the embedding distortion caused by visible watermarking. The same can be concluded from WPSNR-3 & SSIM-3. Taking into consideration the size of the hidden payload, which is invisibly embedded into the non-ROI region, it can be inferred from PSNR-1 that the image with smaller payload has high PSNR values. Difference (in F-16 & splash ) may occurs due to encoding length (L=4 or 3). The prediction is more correct for smooth regions. In such case the approximate image generated is very close to the original image. This can be inferred from PSNR-3,SSIM2,WPSNR-2. The value of SSIM that is closer to 1 infers that the image quality is retained due to HVS exploitation. Table-II shows the result of embedding watermark mark-3 of 4 sizes at upper-left of the image lena. It can be clearly seen that larger the size of watermark implies larger payload. As a result the PSNR of Non-ROI decreases (PSNR -1). Also relatively low PSNR-2 demonstrates the distortion in embedding in smooth region. In addition it was noted that watermark complexity also added to payload. Thus of all watermarks of same size considered mark-3 was the one with larger payload. In summary the watermarking process depends on embedding area in host, watermark size and complexity and encoding length. Table III shows the results of embedding buyer digest which suggests that the digest length contributes to PSNR value. Higher the length, lower the PSNR value. However it was authenticable in all cases. Images of test:
V.
Velammal College of Engineering and Technology, Madurai
Fig.1a Lena
Fig.1b Peppers
Fig.1c Splash
Fig.1d Airplane
Watermarks used:
Fig.2a mark-1
Fig.2b mark-2
Fig.2c mark-3
Fig.2d mark-4
Page 171
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Watermarked Images:
Fig.3a
Fig.3b
Fig.3c
Fig 4(a) 11thframeFig. 4(b) 61thframe Fig. 4(c) 200thframe Fig.4(d) 400thframe
Fig.3d
Watermarked Video Frames:
Sample Buyer id : abcd Its Length : 107 Its Digest : 10011111000100000101000010100000110110000111111011011101110 100001111000111100110001101111011000110110111001
Fig.5a. Watermarked lena with mark-3
Fig.5c.Digest embedded lena
Fig.5b.Retrieved lena
Fig.5d.Extracted Lena
TABLE I: PERFORMANCE EVALUATION: WATERMARKING VARIOUS IMAGES WITH MARK-3 OF SIZE 128 × 128 AT VARIOUS ROI.
Images
Lena F-16 Peppers Splash
Payload (in bits) 25372 25701 26143 25188 24515 25207 23473 25099 24832 24828 25246 24322
Position
PSNR-1
PSNR-2
PSNR-3
(0,0) (160,160) (352,272) (0,0) (160,160) (352,272) (0,0) (160,160) (352,272) (0,0) (160,160) (352,272)
55.3490 42.8205 42.5805 66.8455 45.8896 45.7781 85.7012 50.6652 53.2093 65.2598 44.0934 44.0734
37.8900 39.0298 38.0277 39.1492 39.2714 39.5457 44.1700 40.1900 44.3139 42.2465 40.3683 38.6686
59.8601 65.0116 65.0359 65.2507 65.1645 64.9491 62.8522 65.0808 65.0948 65.2507 64.5645 64.6513
WPSNR -1 30.2225 28.7887 28.7301 25.0714 24.9415 25.1664 37.6478 46.8588 50.4606 34.2956 33.6492 33.4906
WPSNR -2 49.4495 49.2789 52.6813 62.4066 58.8382 57.0555 Inf 54.9316 Inf Inf Inf 47.0625
WPSNR -3 33.2196 49.0391 32.6273 29.2657 46.0943 47.1568 53.9220 47.7934 52.3948 54.0074 50.4374 47.0983
SSIM-1
SSIM-2
0.9631 0.9735 0.9634 0.9681 0.9738 0.9799 0.9857 0.9807 0.9852 0.9770 0.9705 0.9643
0.9995 0.9998 0.9996 0.9996 0.9996 0.9998 0.9992 0.9997 0.9995 0.9996 0.9995 0.9995
SSIM3 0.9642 0.9873 0.9788 0.9686 0.9829 0.9895 0.9858 0.9850 0.9892 0.9777 0.9835 0.9773
TABLE II : PERFORMANCE EVALUATION : WATERMARKING LENA IMAGE WITH MARK-3 OF VARIOUS SIZE AT (0,0)
Watermark size 32 × 32 64 × 64 128 × 128 256 × 256
Payload (in bits) 2831 9319 25372 52700
PSNR-1
81.0282 81.0093 55.3490 40.2160
PSNR-2
PSNR-3
49.2997 43.5129 37.8900 33.2593
66.1034 61.2387 59.8601 61.1828
WPSNR1 58.0810 43.7084 30.2225 30.2871
WPSNR2 Inf 55.6605 49.4495 Inf
WPSNR3 58.9150 49.4605 33.2196 31.0454
SSIM-1
SSIM-2
SSIM-3
0.9973 0.9877 0.9631 0.9277
0.9998 0.9994 0.9995 0.9992
0.9973 0.9877 0.9642 0.9366
TABLE III : PERFORMANCE EVALUATION : BUYER AUTHENTICATION OF RETRIEVED IMAGE AT ROI STARTING FROM (0,0).
Velammal College of Engineering and Technology, Madurai
Page 172
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Image
Authenticable
Digest Length
Lena F-16 Peppers Splash
yes yes yes yes
110 107 117 120
VI. CONCLUSION A reversible visible watermarking technique is presented in this paper which can be applied to any visual media. The paper proposes a method which considers HVS of the host to watermark, to achieve the desired features of visible watermarking. And at same time the image quality is retained as shown by our results. Further the application being blind makes it suitable for extraction of original at any place & time As a key dependent method the scheme allows only authentic users with correct key to retrieve the original. The buyer authentication when embraced in the process, though a simple technique, helps a lot in not only preventing piracy but also to find the buyer involved in piracy.
PSNR of retrieved image 59.8601 65.2507 62.8522 65.2507
PSNR of authentic image 59.9861 65.2663 63.4003 66.0321
[12] F.Muller, “Distribution shape of two dimensional DCT coefficients of natural Images” Electronics letters vol 29, no 22, pp 1935-1936, Oct 1993.
REFERENCES [1] G. Braudaway, K. A. Magerlein, and F. Mintzer, “Protecting publicly available images with a visible image watermark,” Proc. SPIE, International Conference on Electronic Imaging, vol. 2659, pp. 126–133, Feb. 1–2, 1996. [2] M. S. Kankanhalli, Rajmohan, and K. R. Ramakrishnan, “Adaptive visible watermarking of images,” in Proc. IEEE Int. Conf. Multimedia Comput. Syst., vol. 1. Florence, SC, Jul. 1999, pp. 568–573. [3] A. M. Alattar, “A Novel Difference Expansion Transform for Reversible Data Embedding,” IEEE Trans . on Information Forensics and Security,, vol. 3, no. 3, pp. 456–465, Sep. 2008. [4] C. De Vleeschouwer, J.-F. Delaigle, and B. Macq, “Circular interpretation of bijective transformations in lossless watermarking for media asset management,” IEEE Trans. Multimedia, vol. 5, no. 1, pp. 97–105, Mar. 2003. [5] Kamstra, L., Heijmans, H.J.A.M. “Reversible data embedding into images using wavelet techniques and sorting,” IEEE Trans. Image Process., vol. 14, no. 12, pp. 2082–2090, Dec. 2005. [6] S. C. Pei and Y. C. Zeng, “A novel image recovery algorithm for visible watermarked images,” IEEE Trans. Inf. Forens. Security, vol. 1, no. 4,pp. 543–550, Dec. 2006. [7] Y. Yang, X. Sun, H. Yang, and C.-T. Li, “Removable visible image watermarking algorithm in the discrete cosine transform domain,” J.Electron. Imaging, vol. 17, no. 3, pp. 033008-1–033008-11 Jul.–Sep. 2008. [8] Y. J. Hu and B. Jeon, “Reversible visible watermarking and lossless recovery of original images,” IEEE Trans. Circuits Syst. Video Technol., vol. 16, no. 11, pp. 1423–1429, Nov. 2006. [9] Ying Yang, Xingming Sun, Hengfu Yang, Chang-Tsun Li, and Rong Xiao “A Contrast-Sensitive Reversible Visible Image Watermarking Technique,” IEEE Transactions On Circuits And Systems For Video Technology, vol. 19, no. 5, pp. 656-677, May 2009 [10] D. Coltuc and J. M. Chassery, “Very fast watermarking by reversible contrast mapping,” IEEE Signal Process. Lett., vol. 14, no. 4, pp. 255–258, Apr. 2007. [11] Randall C. Reiningek and Jerry D. Gibson “Distribution of two dimensional DCT coefficients for Images” IEEE transactions on communications, vol. com -31, no 6 Jun 1983.
Velammal College of Engineering and Technology, Madurai
Page 173
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Web Personalization System for evolving user profiles in Dynamic Web Sites based on Web Usage Mining Techniques and Agent Technology G.Karthik)#1 B.E.,M.E.,(Ph.D), R.Vivekanandam*2 M.Sc., M.S.,(Ph.D), P.Rupa Ezhil Arasi#3 M.Sc., M.Phil.,M.E 1 Department of CSE,Vinayaka Mission Kirupananda Variyar Engineering College Salem, Tamilnadu,India 2 Department of Applied Science, Muthayammal Engineering College Raipuram, Namakkal Dt, TamilNadu,India 3 Department of CSE, Vinayaka Mission Kirupananda Variyar Engineering College Salem, Tamilnadu,India 1
[email protected] 2 [email protected] 3 [email protected] Abstract - Customer Relationship Management use data from within and outside an organization to allow an understanding of its customers on an individual basis or on a group basis such as by forming customer profiles. These profiles can be discovered using web usage mining techniques and can be later personalized. Web personalization system captures and models behavior and profiles of users interacting with the web sites. Web personalization is the process of customizing a web site to the needs of specific users taking advantage of the knowledge acquired from the analysis of user’s navigational behavior in correlation with the information collected namely
I.
structure, content and user profile data. The output of the web usage mining process can be improved by using agent technology. Agent technology provides a dynamic personalized guidance to the visitors of the web. This paper describes the design of a web usage mining architecture for web personalization system implemented using a multi-agent platform. Keywords-Customer Relationship Management, Web usage Mining, Web personalization, Agent technology, Customer profiles
INTRODUCTION
The output of the web usage mining process can be improved by using agent technology. This creates a friendly relationship between the web and its users. In this paper, we are interested in the web usagemining domain, which is described as the process of customizing the content and the structure of the web sites in order to provide users with the information they are interested in. Various personalization schemes have been suggested in the literature. Lelizia [9] is perhaps the first system, which takes into account the user’s navigation through the first system. Yan et al. [2] propose a methodology for the automatic classification of web users according to their access patterns, using cluster analysis on the web logs. In [3], Joachims et al. describe Web Watcher, and similarly the Personal Web Watcher in [4], an intelligent agent system that provides navigation hints to the user, on the basis of a knowledge of the user’s interests, the location and relevance of the many items in the site, and the way in which other users interacted with the collection in the past. We can finally conclude that most of the existing works try to classify a user i) while she is browsing the web site or ii)
Velammal College of Engineering and Technology, Madurai
using registration information. The main criticism stands in the fact that in some applications it is not possible to perform an “on line” classification if the number of visited pages is not sufficiently great. By the way using the registration form alone may result inaccurate if the interests of user change over time. The novelty of our approach is that we perform clustering of the user sessions extracted from the web logs to partition the users into several homogeneous groups with similar activities and then extract user profiles from each cluster as a set of relevant URLs. This procedure is repeated in subsequent new periods of web logging then the Previously discovered user profiles are tracked, and their evolution pattern is categorized. In order to eliminate most sessions from further analysis and to focus the mining on truly new sessions FM model is used. The FM model is suitable for real time matching of session to pre-generated cluster and it offers scalable models of clusters. The rest of this paper is organized as follows: In Section 2, we present an overview of web usage mining. In Section 3, we describe our approach to profile discovery using web
Page 174
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
usage mining. In Section 4, we discuss our approach for tracking evolving user profiles. In Section 5, we present the overview of web personalization. In Section 6 we present the personalization solution. In Section 7, we present the main agents of our architecture along with their roles and interactions among them. In Section 8, we present our results. Finally in Section 9, we present our conclusions. II. AN OVERVIEW OF WEB USAGE MINING Web usage mining is a tool for web personalization, since it captures and models behaviors and profiles of users interacting with a web site. These models can be used by the personalization systems to better understand the behavioral characteristics of visitors, the content and structure of the web sites and provide dynamic recommendation to visitors. Web mining is performed in several stages [5], [6] to achieve its goals 1) Collection of Web data such as activities / click streams recorded in web server logs. 2) Preprocessing of Web data such as filtering, crawlers requests, requests to graphics and identifying user sessions. 3) Analysis of web data, also known as web usage mining [7], to discover interesting usage patterns of profiles and 4) Interpretation / evaluation of the discovered profiles 5) Tracking the evolution of discovered profiles.Web usage mining can use various data mining or machine learning techniques to model and understand web user activity. A. Handling Profile Evolution Since dynamic aspects of Web usage have recently become important it is desirable to study and discover Web usage patterns. News Dude [8] is an intelligent agent built to adapt to changing user’s interests by learning two separate user models that represent short-term and long-term interests. In [9], a user profiling system was developed based on monitoring, the user’s Web browsing and e-mail habits. In [9], as user-profiling system was developed based on monitoring the user’s web browsing and e-mail habits. This system used a clustering algorithm to group user interests into several interest themes, and the user profiles had to adapt to changing interests of the users over time. The above approaches are based on supervised learning framework and the present work is based on an unsupervised learning framework that tries to learn mass anonymous user profiles on the server side. In this work the web logs are fully mined for each period and then the subsequent results are compared. III. PROFILE DISCOVERY BASED ON WEB USAGE MINING
1) Preprocess Web log file to extract user sessions 2) Cluster the user sessions by using Hierarchical Unsupervised Niche Clustering 3) Summarize session clusters / categories into user profiles 4) Enrich the user profiles with additional facets by using additional web log data and external domain knowledge 5) Track current profiles against existing profiles. A. Preprocessing the Web Log File and Clustering sessions into an optimal number of categories Each log entry in the file consists of access time, IP address, URL viewed, REFERRER etc. The main step in preprocessing is that it maps the URLs on a web site to distinct indices. Sessions are implemented as lists instead of vectors thus saving memory and computational costs. For Clustering Unsupervised Niche Clustering algorithm is used. H-UNC uses Genetic Algorithm and handles noise in the data and automatically determines the number of clusters. After grouping the sessions into clusters, session categories are summarized in terms of user profile vectors [6] [7]. The vector captures the relevant URL in the profile. The profiles are then converted into binary vectors. Each profile is discovered along with a measure that represents the amount of variance or dispersion of user sessions in a given cluster around the cluster representative. This measure is useful in determining the boundary of each cluster and also determines whether two profiles are compatible or not.
Fig. 1. Web usage mining process and discovered profile facets
IV. TRACKING EVOLVING USER PROFILES Each profile is discovered along with an automatically determined measure that determines the amount of variance or dispersion of the user sessions in a given cluster around the cluster representative. This measure determines the boundary around each cluster and also determines whether two boundaries overlap. This comparison leads to events such as Persistence, Birth and Death. Aggregating these events helps in tracking profiles over many periods. Profiles retrieved should be as close as possible to the original session data.
The automatic identification of user profiles is a knowledge discovery task consisting of periodically mining new contents of the user access log files is summarized on the following steps:
Velammal College of Engineering and Technology, Madurai
Page 175
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
V. WEB PERSONALIZATION Web personalization is defined as any action that adapts the information or services provided by a web site to the needs of a particular user or set of users, taking advantage of the knowledge gained from the user’s navigational behavior and individual interests, in combination with the content and structure of the web site. The overall process of usage-based web personalization consists of five modules, which correspond to each step of the process. They are 1) User Profiling-it is the process of gathering information specific to each visitor either explicitly or implicitly. 2) Log analysis and Web usage mining- Information stored in Web server logs is processed by applying data mining techniques in order to a. Extract statistical information and discover interesting usage patterns b. Cluster the users groups according to their navigational behavior c. Discover potential correlations between the web pages and user groups. 3) Content Management- It is the process of classifying the content of a web site in semantic categories in order to make information retrieval and presentation easier for the users. 4) Web site publishing- It is used to present the content stored locally in a Web server and / or some information retrieved from other Web resources in a uniform way to the end-user. 5) Information acquisition and searching-Since the users are interested in information from various Web sources searching and relevance ranking techniques must be employed both in the process of acquisition of relevant information and in the publishing of the appropriate data to each group of users. VI. THE PERSONALIZATION SOLUTION Web usage mining techniques when combined with the multiagent architecture gives a personalization solution to web sites. Multi-agent architecture consists of a set of autonomous agents interacting together to fulfill the main goal of the system. Agent’s taps into the communication stream between a user’s web browser and the web itself. Agents observe the data flowing along the stream, observe the data flowing along the stream and alter the data as it flows past. These agents can learn about the user, influence what the user sees by making up pages before passing them on, and provide entirely new functions to the user through the web browser. Agents are divided into modules that have well defined tasks and that are further divided into two working groups such as data mining module and personalization module. The personalization agent uses the user model knowledge along with the previously discovered sequential patterns and applies a set of personalization rules in order to deliver
Velammal College of Engineering and Technology, Madurai
personalization tasks or functions. The functions based upon the personalization policy provide a complete personalization solution .The multi-user personalization policy used is static and are adjusted to the browsing context of the user. VII.
INTERACTIONS AMONG AGENTS
Agents co-operate and co-ordinate their work by sending messages among them. The interface agent captures the navigational behavior of the user and informs the user data agent and online classification agent about it. The user data agent identifies the current user, creates and records a new navigational session for him. The classification agent classifies the active session in one of the groups discovered by the clustering agent using the information embedded in his or her session. The classification agent informs the evaluation agent about the group. This is notified to the decision maker agent. The decision maker agent informs the personalization agent about the personalization functions to execute according to the adaptation rules associated to that group of user. The personalization agent sends the results to the interface agent, which displays them to the user as response to his or her request.
Page 176
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 2. Functional structure of PWUM Architecture
VIII. RESULTS H-UNC [10] was applied on a set of Web sessions preprocessed from Web log data for several months. After filtering, the data was segmented into sessions based on the client IP address. After filtering the irrelevant URLs unique sessions were obtained. H-UNC partitioned the web user sessions of each period into several clusters and each cluster was characterized by one of the profile vectors. Web Usage Mining multi-agent system for web personalization enhances the quality of discovered models and hence optimizes the personalization process. The software agents of PWUM have been implemented using multi agent platform called JADE [11]. The results obtained by using both of multi agents systems and WUM techniques were very encouraging. IX CONCLUSIONS We presented our system and described the mechanism necessary for Web Usage Mining and Personalization tasks. The combination of more than one technique of WUM enhances the quality of discovered models, so this optimizes the personalization process. The use of multi-agent paradigm reduces the time complexity. We are looking forward testing our approach in tourism web sites as part of national research projects. REFERENCES [1] H. Lieberman. Letizia. An agent that assists web browsing. In Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, pages 924-929, 1995. [2] T.W.yan, M.Jacobsen, H.Garcia-Molina, and U.Dayal. From user access patterns to dynamic hypertext linking. In Proceeding of the fifth International World Wide Web Conference, Paris, 1996. [3] T.Joachims, D.Freitag and T. Mitchell. Web watcher: a tutor guide for the World Wide Web. In proceedings of the fourteenth International Joint Conference on Artificial Intelligence, pages 924-929, 1995. [4] D.Madenic. Machine Learning used by personal web watcher. In proceddings of the workshop on Machine learning and Intelligent Agents (ACAI-99), Chania, Greece, July 1999. [5] R.Cooley, B.mobasher, and J.Srivastava, “ Web mining: Information and Pattern Discovery in the World Wide Web,” Proc. Ninth IEEE Int’l Conf. Tools with Ai (ICTAI ’97),pp.558-567,1997. [6] O.Nasraoui, R.Krishnapuram, H.Trigui and A.Joshi, “Extracting Web User profiles Using Relational Competitive Fuzzy Clustering” Int’l J. Artificial Intelligence Tools, vol 9,no. 4, pp.509-526,2000. [7] J.Srinivastava, R.Cooley, M.Deshpande, and P.N.Tan, “Web Usage Mining: Discovery and Applications of Usage Patterns from Web Data,”SIGKDD Explorations, vol. 1, no.2,pp1-12, Jan 2000. [8] D.Billus and M.J.Pazzani , “ A Hybrid user Model for News Classification, “ Proc . Seventh Int’l Conf. User Modeling (UM ’99), J.Kay,ed., pp.99-108,1999 [9] I.Grabtree and S.Soltysiak, “Identifying and Tracking Changind Interests”, Int’l J. Digital Libraries, vol.2, pp.38-53 10] O.Nasraoui, R.Krishnapuram, “A New Evolutionary Approach to Web Usage and Context Sensitive Associations Mining” Int’l J.Computational Intelligence and Applications, special issue on Internet intelligent systems, vol.20, no.3, pp.339-348, Sept 2002. [11] Bellifemine F.etal, 2004 JADE Basic Documentation: Programmer’s Guide.
Velammal College of Engineering and Technology, Madurai
Page 177
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Automated Test Case Generation and Performance Analysis for GUI Application Ms. A.Askarunisa#1, Ms. D. Thangamari*2 #
Assistant Proffessor,Department of Computer Science and Engineering, Affiliated to Anna University, Thirunelveli Thiagarajar College of Engineering, Madurai, Tamilnadu, India. 1
[email protected],
*
Department of Computer Science and Engineering, Affiliated to Anna University, Thirunelveli Thiagarajar College of Engineering, Madurai, Tamilnadu, India. 2
[email protected]
Abstract— A Common method for GUI testing is the Capture and Replay (CR) technique.GUIs is complex pieces of software. Testing their correctness is challenging for several reasons: 1. Tests must be automated, but GUIs are designed for humans to use.2. Conventional unit testing, involving tests of isolated classes, is unsuitable for GUI Components. 3. GUIs respond to user-generated events 4. Changes in the GUI’s layout shouldn’t affect robust tests.5.Conventional test coverage criteria, such as “90 percent coverage of lines of code”. This paper proposes GUI Automation testing framework to test GUI-Based java programs as an alternative to the CR technique. The framework develops GUI-event test specification language for GUI application written using java swing APIs, which initiates an automated test engine. Visual editor helps in viewing the test runs. The test engine generates GUI events and captures event responses to automatically verify the results of the test cases. The testing framework includes the test case generation, test case execution and test case verification modules. The testing efficiency is measured by determining coverage metric based on Code coverage, Event coverage and Event Interaction coverage, while may be useful during Regression Testing. The paper uses Abbot and JUnit tools for test case generation and execution and Clover tool for code coverage. We have performed testes on various GUI applications and the efficiency of the framework is provided. Keywords— Automated testing, Coverage, GUI Testing, Test Suite Reduction
I. INTRODUCTION Test automation of GUI means mechanizing the testing process where testers use software in controlling the implementation of the test on the new products and comparing the expected and the actual outcomes of the product application. Scheduled testing tasks on a daily basis and repeating the process without human supervision is one advantage of automated testing. With all the mass production of gadgets and electronic GUI devices, the testing period seems quite demanding. Electronic companies must ensure quality products to deliver excellent products and maintain customer preferences over their products. In running automatic tests for the GUI application, the tester saves much time, especially when he is in a huge production house and needs to be multi-tasking. There are actually four strategies to test a GUI.1. Window mapping assigns certain names to each element, so the test is more manageable and
Velammal College of Engineering and Technology, Madurai
understandable.2. Task libraries sort the step sequence of the user task when they appear in multiple tests.3. The Datadriven type of test automation separates the limitation of the test case and test script so the test script will be reusable.4. Keyword-driven test automation converts tests as spreadsheets or tables, as it creates parsers to decode and perform the description of the test. The reminder of this paper is organized as follows: Section 2 covers the Background material for this proposal, Section 3 describes the proposed approach for GUI testing framework. Section 4 briefly highlights the implementation details. Section 5 gives the Conclusion and future enhancement. II. BACKGROUND AND RELATED WORK Existing works on GUI testing are mainly concerned with test automation assisting tools supporting the capture/replay technique. Researches have considered various techniques for GUI application testing, and coverage metric for test case. White et al [2, 3] and Belli [4] developed model based testing for GUI application under test. Each responsibility is simply the desired response for the user and can be specified as a complete interaction sequence (CIS) between the user and the GUI application under test. Then a finite-state machine is developed for each CIS, which generates the required tests and materializes the CIS. In the work of Memon et al [5], the GUI under test is modelled as a finite-state machine with hierarchical structure. The test case generation problem of GUI testing then follows the goal-oriented philosophy and is treated as an AI(Artificial Intelligence) planning problem. The approach can be viewed as a global one in the sense that a single or global finite-state machine is constructed for all test cases of interest. In the work of Cai et al [6], a GUI test case is defined as a word defined over a finite alphabet that comprises primitive GUI actions or functions of concern as symbols. Meyer [7] defined Capture/Replay Testing technique, a test case as an input with its expected output. Binder [12] defined a test case to consist of a pretest state of the software under test (including its environment), a sequence of test inputs, and a statement of expected test results. Memon et al [1,5] defined a test case to consist of an initial state and a legal action sequence. The problem here is how to define the state of the
Page 178
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
GUI. A few approaches have been proposed for GUI testing. The most known one may be the so-called capture/replay technique [8]. In the work of Sun [21] propose a specification-driven approach to test automation for GUI based Java programs as an alternative to the use of capture/replay. However the rationale of existing test case definitions is rarely discussed. An exception is the recent work of Rothermel et al [15], which assesses the effects of test suite granularity and test input grouping on savings in test execution time and costs in defectdetection effectiveness in the context of regression testing. Test suite granularity pertains to the size of the test cases so grouped or the number of inputs per test case, whereas test suite grouping pertains to the content of test cases or the degree of diversity among the inputs that compose a test case. Memon, Pollack and Sofia [5, 11, 12] exploit planning techniques developed and used extensively in artificial intelligence to automate the generation of test cases for GUIs. Given a set of operators, an initial state and a goal state, the planner produces a sequence of the operators that will transform the initial state to the goal state.
for extending the basic tester classes for new components, and has a well defined method for naming those new actions so they are automatically available at the scripting level. It also provides extensibility for Converting strings from scripts into arbitrary classes, and introducing new Individual steps into scripts. Scripts can call directly into java code (the script is actually just a thin veneer over method calls). Abbot provides both a script environment and a JUnit [13] fixture, both of which handle setup and teardown of the complete GUI environment.
III. PROPOSED APPROACHES FOR TESTING GUI APPLICATION In the Previous Existing methods, none of the tool does not support both capture replay technique and programmatic technique. Graphical user interface or GUI is the interactive graphical representation of its underlying application. Front end users find it easy to use this interface to operate the software since it is not necessary for them to understand the programming language used in the software. Testing such software for its reliability becomes complex since the testers have to test the software as well as the GUI for its design functionality. The proposed testing is based on functionality and in order to perform functional testing, different kinds of automated tools are available in market which will be quite hard for the tester to choose for satisfying his requirements. In this paper represents Abbot tool that supports both capture replay techniques (black box Testing) and Glass box technique (White box Testing) is used for test case narration and execution. This process is also called as Gray Box Testing. A.TestCase Generation Framework Testing Framework [9] application Involves, (Fig1) Test Case Generation, Testing Oracle, Automatic Executions [21], Test Result and Performance Analysis. Abbot uses a dedicated class to represent GUI components .It stores a range of attributes to be used when looking for a matching component in the hierarchy, and does a fuzzy lookup to avoid spurious errors due to component movement or minor changes to the GUI. The lookup mechanism is very general due to the fact it is used by the scripting layer which has no a priori knowledge of the GUI hierarchy or what needs to be looked up. The framework also provides a number of utilities to facilitate inspecting the GUI hierarchy itself. Abbot provides
Velammal College of Engineering and Technology, Madurai
Fig 1 Testing Framework
1) Test Case Generation Test case Generation and Execution is done using Abbot Tool. Abbot is a framework for driving java UI components programmatically. Costello is a script editor and a launcher which accompanies Abbot. They are based on the same capture/ playback/script principle that marathon uses, but Abbot can also be used programmatically. The framework can be invoked differently from Java code or accessed without programming through the use of 20 scripts. It is suitable for use both by developers for unit tests and QA for functional testing. The framework easily integrates with the JUnit test harness and therefore, during application development, the functional GUI tests can become a part of the test suite. All these features of Abbot make it an effective framework for rapidly creating a comprehensive test framework. Collections of test cases are generated by using Abbot and JUnit Tool. The test cases contain the sequence of user Input Events. GUI Model
Test Case Generation
JAR Files
Page 179
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Measurement is the process by which numbers or symbols are assigned to attributes of entities in the real world in such a way as to characterize them according to clearly defined rules. Coverage of GUI applications Require various tasks: Performance analysis and Coverage Report. Coverage measurement also helps to avoid test entropy. As your code goes through multiple release cycles, there can be a tendency for unit tests to atrophy. As new code is added, it may not meet the same testing standards you put in place when the project was first released. Measuring code coverage can keep your testing up to the standards you require. You can be confident that when you go into production there will be minimal problems because you know the code not only passes its tests but that it is well tested.
1) Code Coverage Code coverage analysis is sometimes called test coverage analysis. The two terms are synonymous. The academic world more often uses the term "test coverage" while practitioners more often use "code coverage". Likewise, a coverage analyser is sometimes called a coverage monitor. Code coverage is not a panacea. Coverage generally follows an 8020 rule. Increasing coverage values becomes difficult, with new tests delivering less and less incrementally. If you follow defensive programming principles, where failure conditions are often checked at many levels in your software, some code can be very difficult to reach with practical levels of testing. Coverage measurement is not a replacement for good code review and good programming practices. In general you should adopt a sensible coverage target and aim for even coverage across all of the modules that make up your code. Relying on a single overall coverage figure can hide large gaps in coverage. 2) Code Coverage with Clover Clover [24] uses source code instrumentation, because although it requires developers to perform an instrumented build; source code instrumentation produces the most accurate coverage measurement for the least runtime performance overhead. As the code under test executes, code coverage systems collect information about which statements have been executed. This information is then used as the basis of reports. In addition to these basic mechanisms, coverage approaches vary on what forms of coverage information they collect. There are many forms of coverage beyond basic statement coverage including conditional coverage, method entry and path coverage. Clover is designed to measure code coverage in a way that fits seamlessly with your current development environment and practices, whatever they may be. Clover's IDE Plug-in provide developers with a way to quickly measure code coverage without having to leave the IDE. Clover's Ant and Maven integrations allow coverage measurement to be performed in Automated Build and Continuous Integration systems, and reports generated to be shared by the team. The Clover Coverage Explorer: The Coverage Explorer allows you to view and control Clover's instrumentation of your Java projects, and shows you the coverage statistics for each project based on recent test runs or application runs. The main tree shows coverage and metrics information for packages, files, class and methods of any Clover-enabled project in your workspace. Clover will auto-detect which classes are your tests and which are your application classes - by using the drop-down box above the tree you can then restrict the coverage tree shown so that you only see coverage for application classes, test classes or both. Summary metrics are displayed alongside the tree for the selected project, package, file, class or method in the tree. The Clover Coverage Measurement: Clover uses these measurements to produce a Total Coverage Percentage for each class, file, and package and for
Velammal College of Engineering and Technology, Madurai
Page 180
Fig 2 Test Case Generation using Abbot Tool
This framework shown in Fig 2 chooses a specific model of the GUI application. This model satisfies the GUI application under test which is also the input of the Test Case generation. Collections of test cases are generated by using Abbot and JUnit Tool. The Test case contains the sequence of user Input events. The test cases are run with test runner. A test designer interacts with the GUI and generates mouse and keyboard events. 2) Test Case Execution Test cases from the repository are executed one by one automatically using Abbot and JUnit Tools as shown in Fig 3. Source Program
Jar files
Repository
Collection of test cases
Test Execution Automatically
Fig 3 Test Execution by JUnit Tool
3) Test case Verification: The expected results of various test cases are manually determined and stored in the GUI model of Testing Oracle. The Testing Oracle contains the expected state of sequences for each application. When the test cases start running, the start timer is initialized. The events are generated automatically and the actual results from the tool are verified with the expected results as shown in Fig 4. Pass/Failure testing report is generated accordingly. Actual State
Expected State
Automated Verified
Verified
Fig 4 Test Case Verification
B. Performance Analysis
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the project as a whole. The Total Coverage Percentage allows entities to be ranked in reports. The Total Coverage Percentage (TPC) is calculated as follows: TPC = (BT + BF + SC +MC)/(2*B + S + M) where BT - branches that evaluated to "true" at least once BF - branches that evaluated to "false" at least once SC – statements covered MC - methods entered B - total number of branches S - total number of statements M - total number of methods 3) Event Coverage These coverage criteria use events and event sequences to specify a measure of test adequacy. Since the total number of permutations of event sequences in any non-trivial GUI is extremely large, the GUI's hierarchical structure is exploited to identify the important event sequences to be tested. A GUI is decomposed into GUI components, each of which is used as a basic unit of testing. A representation of a GUI component, called an event flow graph [16], identifies the interaction of events within a component and intra component criteria are used to evaluate the adequacy of tests on these events. The hierarchical relationship among components is represented by an integration tree, and inter-component coverage criteria are used to evaluate the adequacy of test sequences that cross components. 4) Event Interaction Coverage The sequence of possible Event Interacts with an other Event. This type of event coverage is called as Event Interaction Coverage [22]. The event interaction coverage is consisting of 2 ways and 3 ways Combination. C. Coverage Report 1) Coverage HTML Report The clover html report task generates a full HTML report with sensible default settings. It is also generated prior to generation of the full report. 2) Coverage XML Report The clover xml report task generates a full HTML report with sensible default settings. It is also generated prior to generation of the full report. 3) Coverage PDF Report The clover pdf report task generates a PDF report with sensible default settings. It is also generated prior to generation of the full report. D. Coverage Metric The coverage metric CONTeSSi (n) (CONtext Test Suite Similarity) [23] for each model value factor is calculated and compared with the original pair of model. CONTeSSi (n) (CONtext Test Suite Similarity) that explicitly considers the context of n preceding events in test cases to develop a new “context-aware” notion of test suite similarity. This metric is an extension of the cosine similarity metric used in Natural Language Processing and Information Retrieval for comparing an item to a body of knowledge, e.g., finding a query string in a collection of web pages or determining the likelihood of finding a sentence in a text corpus (collection of documents).We evaluate CONTeSSi (n) by comparing four test suites, including suites reduced using conventional criteria, for four open source applications. Our results show
that CONTeSSi (n) is a better indicator of the similarity of test suites than existing metrics.
Velammal College of Engineering and Technology, Madurai
Page 181
This paper considers different models with varying frequencies of events like all individual events, all two pair events, all three pair events, etc. Coverage metric is calculated for various factors like statement, branch, method, etc. IV. IMPLEMENTATION This paper uses the GUI model of Calculator Application, which is written using Java Swing. A scientific Calculator Application contains a Collection of Standard Buttons, Control Buttons and Scientific Buttons. It is used to calculate the Arithmetic and Scientific data values. It performs several Basic operations such as Add, Sub, Mul and Div and Also Scientific Operations Such as Sin, Cos, Tan, Log, sqrt and etc,The Program also Contains Radio Buttons like Hexadecimal, Octal, Decimal, and Binary, which enable the The automatic execution of user input sequence is shown in Fig 5. All the operations are performed by Click Events .For this Application, the test cases are written by using java Swing. The unit testing is done by Abbot and JUnit Tool.
Figure 5: Automatic Execute of User Input Sequence
A. Automation Testing with Abbot The Specific Calculator Program written using Java Swing is selected for testing. The Calculator program is run with abbot and junit tool and the each action is performed using the button click event. The Calculator Application, test case generation makes use of the core components Component Test Fixture () and Component Tester () in the Abbot and also write the test cases are using action click (), Selectmen Item () , action Delay ().By making use of the components test cases are created. The assert method compares data value obtained from calculator program with data value automatically execute one by one using the test case sequences. The click event for Hexa Decimal option Button and the corresponding enabled buttons are shown in Fig 6
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1) Code Coverage Code coverage and event based performance analysis are done in the module. Clover coverage tool is used for collecting the metrics for frames, methods and package. Code coverage tool is used for statements, branches and loop.
Figure 6: Click Event for Hexa Decimal Button
B. Coverage View Manual analysis is done for event interaction and events and the relevant data is collected. Coverage metrics are collected for different units of application.The different explorer view of coverage report for all completed test cases is show in fig 7,8,9 Fig 10: Coverage view of Code Cover tool for Calculator program
Figure 7: Coverage Explorer View for all Test cases
2) Event Coverage Event Coverage is determined by the total number of click events generated by the user. For Eg table3 shows that Single event is generated for Basic button, 11 events are generated for calculating the area of circle (pi*r*r). 3) Event Interaction Testing It represents all possible sequences of events that can be executed on the GUI. In that Calculator Application contains as the collection Buttons such as standard, Control, and Scientific. In Figure 11 represents that, • e1 represents the clicking in the file Menu • e2 represents the clicking in the Menu Selection event • e4 represents the clicking in the Basic button after clicking the e2 event. It is similar to e5.
Fig 8: Project Coverage Report in XML Format
Fig 11 Shows that Event Based Sequences
Fig 9: Coverage Report in PDF Format
Velammal College of Engineering and Technology, Madurai
Table1 displays the report title and the time of the coverage contained in the report. The header displays metrics for the package files or project overview which is currently selected. Depending on the current selection, the metrics include all or a subset of: Number of Lines of Code (LOC),Number of Non-commented Lines of Code (NCLOC),Number of Methods, Number of Classes, Number of Files, Number of Packages.
Page 182
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
In considering this context, the event pair coverage suite in Table 4 is expected to be more similar to the original suite than the event coverage suite, since the event pair coverage suite is created based on the existence of event pairs. Table 4(b) shows the count of each event pair for each suite. This is the basis of CONTeSSi (n), for n = 1, since we are looking at events in the context of one other (previous) event. Now if we extend this example to compute CONTeSSi (2), we obtain the frequencies shown in Table 4(C). In general, as n increases, the frequencies for the event sequences decrease, as they appear less frequently in the test suites. Intuitively, comparing test suites on longer sequences will make it harder for the test suites to be similar. Therefore, if two test suites have a high similarity score with a larger n, they are even more similar than two suites being compared with a small n. By treating each row in Table 4 (a), (b), or (c) as a vector, CONTeSSi is computed as
follows:CONTeSSi(A,B) =(A · B) /(|A| × |B|) (1)where A and B are the vectors corresponding to the two test suites, A · B is the dot product of the two vectors, i.e.,Pj i=1(Ai×Bi) where j is the number of terms in the vector; and |A| = qPj i=1(Ai)2. The value of CONTeSSi lies between 0 and 1, where a value closer to 1 indicates more similarity. Hence, CONTeSSi (n) is computed as shown in Equation 1, creating a vector for each suite, representing the frequencies of all possible groups of n + 1 events. The inclusion of n previous events will increase the number of terms in the vector, thereby increasing j. The values in Table 5 show the values of CONTeSSi(n) for all our test suites, for n = 0, 1, 2, 3. From these values, we observe that if we ignore context, i.e., use n = 0, most of the reduced suites are quite similar to the original, as indicated by the high (> 0.9) value of CONTeSSi (0). However, the similarity between the test suites decreases as more context
TABLE:1 VIEW OF ALL TEST CASES VALUE USING CLOVER COVERAGE TOOL
Area Test Case
State ment
Bran ches
Metho ds
Classes
LOC
Circle Rectangle Parallelogram Triangle Trapezoid Total
280 293 291 298 297 339
132 132 132 132 132 132
21 24 24 24 24 28
1 2 2 2 2 2
625 667 666 672 672 738
Rectangle Prims Cylinder Cone Sphere Total
312 304 303 300 294 373
132 132 132 132 132 132
25 25 25 25 24 28
2 2 2 2 2 2
690 681 676 671 668 775
Rectangle Prims Cylinder Cone Sphere Pyramid
291 293 296 297 299 295
132 132 132 132 132 132
24 24 24 24 24 24
2 2 2 2 2 2
666 669 671 672 677 671
Total
351
132
29
2
NCLOC
762
483 513 511 518 517 570 Surface 535 527 525 522 513 602 Volume 511 513 516 517 519 515 586
Velammal College of Engineering and Technology, Madurai
Total Cmp
Cmp Densit y
Avg method Cmp
Stmt / methods
Methods / classes
Total Coverage (in %)
138 141 141 141 141 145
0.49 0.48 0.48 0.47 0.47 0.43
6.57 5.88 5.88 5.88 5.88 5.18
13.33 12.21 12.12 12.42 12.38 12.11
21 12 12 12 12 14
54.3 53.6 53.6 54.1 54.1 55
142 142 142 142 141 145
0.46 0.47 0.47 0.47 0.48 0.39
5.68 5.68 5.68 5.68 5.88 5.18
12.48 12.16 12.12 12 12.25 13.32
12.5 12.5 12.5 12.5 12 14
41.6 42.7 53.3 53.3 51.7 53.6
141 141 141 141 141 141
0.48 0.48 0.48 0.47 0.47 0.48
5.88 5.88 5.88 5.88 5.88 5.88
12.12 12.21 12.33 12.38 12.46 12.29
12 12 12 12 12 12
55 55 55 55 55.5 55
146
0.42
5.03
12.1
14.5
55
Page 183
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE: 2 VIEW OF ALL TEST CASES VALUE USING CODE COVERAGE TOOL
Test Case Basic Minus Add Mul Div Mod Hex Dec Oct Bin
Stmt 59.4 63.9 63.9 63.9 63.9 67.1 61.6 60.7 61.6 61.6
Circle Rectangle Parallelogram Triangle Trapezoid
68.9 63.9 63.9 65.3 65.8
Rectangle Prims Cylinder Cone Sphere
64.4 65.3 68.9 68.9 67.6
Rectangle Prims Cylinder Cone Sphere Pyramid Total
63.9 64.8 68.5 68.5 68.5 64.4 97.7
Branch 2.3 10.5 9.3 11.6 12.8 15.1 11.6 9.3 11.6 14 Area 36.0 12.8 12.8 17.4 18.6 Surface 14.0 16.3 34.9 34.9 30.2 Volume 11.6 15.1 34.9 34.9 33.7 15.1 93
Loop 10.1 8.7 8.7 8.7 8.7 10.1 14.5 13.0 14.5 13
Strict Con 6.0 7.7 7.7 7.7 7.7 9.4 19.7 7.7 15.4 9.4
10.1 8.7 8.7 8.7 8.7
14.5 11.1 10.3 13.7 15.4
8.7 8.7 10.1 10.1 10.1
13.7 13.7 16.2 15.4 9.4
8.7 8.7 10.1 10.1 10.1 8.7 34.8
8.5 10.3 14.5 12.8 12.8 11.1 94.9
Velammal College of Engineering and Technology, Madurai
TABLE: 3 VIEW OF ALL TEST CASE EVENT SEQUENCE
Test Plan
Event
Execution (in sec)
Execution Delay (1000 sec)
Basic Scientific Hex Dec Octal Binary
1 1 1 1 1 1
View 1.684 1.763 1.342 1.295 2.262 1.342
2.699 2.714 2.371 2.309 2.356 2.324
Circle Rectangle Parallelogram Triangle Trapezoid
11 7 6 12 12
Area 3.588 2.434 2.262 3.447 3.401
4.555 3.463 3.26 4.446 4.415
Rectangle Prism Cylinder Pyramid Cones Sphere
6 8 11 10 11 14
Volume 2.278 2.036 3.525 2.995 3.588 4.119
3.291 3.65 4.633 4.009 4.556 5.085
Rectangle Prims Cylinder Cones Sphere
26 18 17 13 9
Surface 6.006 4.524 4.68 3.916 3.183
6.989 5.506 5.647 4.898 4.165
Page 184
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE: 4 EXAMPLE TEST CASES YIELDED FROM SEVERAL REDUCTION TECHNIQUES ORIGINAL PAIR
EVENT PAIR
EVENT
STATEMENT
METHOD
BRANCH
e2,e5 e2,e4,e6 e2,e4,e8 e2,e4,e7 e2,e4,e9,e10,e9 e2,e4,e9 e2,e5,e11,e9,e10,e9 e2,e5,e11,e9 e9,e10,e9 e6,e9,e10,e9 e8,e9 e7,e9,e10,e9 e11,e9,e10,e9 e11,e9,e10 e6,e12 e2,e5,e11,e10 e9,e10,e2,e5,e11,e9
e6,e9 e7,e9 e8,e9 e2,e5,e11,e9,e10,e9 e2,e4,e9,e10,e9 e9,e10,e2,e5,e11,e9
e2,e4,e6 e2,e4,e8 e2,e4,e7 e2,e5,e11,e10 e9,e12 e8,e9
e9,e10,e2,e5,e11,e9 e2,e5 e6,e9 e7,e9 e8,e9 e2,e5,e11,e9,e10,e9
e2,e5,e11,e9 e9,e10,e9
e6,e9 e7,e9 e8,e9 e2,e5,e11,e9 e2,e4,e9,e10,e9
Illustrative Tests
e2 e4 e5 e6 e7 e8 e9 e10 e11 e12
TABLE: 4(A) FREQUENCY OF UNIQUE EVENTS OCCURRING IN THE TEST SUITE (LENGTH=0) TEST SUITE
E2
E4
E5
E6
E7
E8
E9
E10
E11
E12
ORIGINAL
10
5
5
3
2
2
18
9
6
1
EVENT PAIR
3
1
2
1
1
1
9
3
2
0
EVENT
4
3
1
1
1
2
2
1
1
1
STMT
3
0
3
1
1
1
7
2
2
0
METHOD
1
0
1
0
0
0
3
1
1
0
BRANCH
2
1
1
1
1
1
6
1
1
0
Illus. Suite
1
1
1
1
1
1
1
1
1
1
TABLE: 4( B) FREQUENCY OF ALL EVENTS PAIR OCCURRING IN THE TEST SUITE (LENGTH=1) TEST SUITE
E2,E4
E2,E5
E4,E6
E4,E7
E4,E8
E4,E9
E5,E11
E6,E9
E6,E10
E6,E12
E7,E9
E8,E9
ORIGINAL
5
5
1
1
1
2
4
1
1
1
1
1
EVENT PAIR
1
2
0
0
0
1
2
1
0
1
1
1
EVENT
3
1
1
1
1
0
1
0
0
0
0
1
STMT
0
3
0
0
0
0
2
1
0
0
1
1
METHOD
0
1
0
0
0
0
1
0
0
0
0
0
BRANCH
0
1
0
0
0
1
1
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
TEST SUITE
E2,E4
E2,E5
E4,E6
E4,E7
E4,E8
E4,E9
E5,E11
E6,E9
E6,E10
E6,E12
E7,E9
E8,E9
ORIGINAL
5
5
1
1
1
2
4
1
1
1
1
1
EVENT PAIR
1
2
0
0
0
1
2
1
0
1
1
1
EVENT
3
1
1
1
1
0
1
0
0
0
0
1
STMT
0
3
0
0
0
0
2
1
0
0
1
1
METHOD
0
1
0
0
0
0
1
0
0
0
0
0
BRANCH
0
1
0
0
0
1
1
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
Illus. Suite
TABLE: 4(C) FREQUENCY OF ALL EVENTS PAIR OCCURRING IN THE TEST SUITE (LENGTH=2)
Illus. Suite
Velammal College of Engineering and Technology, Madurai
Page 185
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE: 5 CONTESSI (N) VALUES FOR SUITE COMPARED TO ORIGINAL FOR ALL BUTTONS IN CALCULATOR APPLICATION IN GUI EXAMPLE SUITES n
EVENT PAIR
EVENT
STMT
METHOD
BRANCH
ILLUS. SUITE
0
0.97308
0.78513
0.95434
0.94405
0.94571
0.7816
1
0.9274
0.60669
0.788116
0.82365
0.685188
0.0000
2
0.9106
0.39509
0.79697
0.79018
0.79018
0.0000
3
0.9428
0.73786
0.82495
0.7071
0.68041
0.0000
4
0.9999
0.0000
0.8660
0.0000
0.5000
0.0000
5
0.9999
0.0000
0.9999
0.0000
0.0000
0.0000
V. CONCLUSIONS AND FUTURE SCOPE GUIs may introduce new types of error, increase complexity and make testing more difficult. The Automation testing reduces the work intensity of the developer and the tester. The main advantage of test automation is that as a result software developers can run tests more often, find and fix bugs on the early stage of development, before end users will face these bugs. The framework suggested in this project simplifies the creation and maintenance of robust GUI tests. Abbot is easy to learn and use, and provides some unique features that can make GUI development more productive. This project developed the method for (i) automatic test case generation, execution and verification (ii) performance analysis by calculating the coverage based on statement, method, branch, event and event interactions for GUI applications. Our results showed that CONTeSSi (n) is a better indicator of the similarity of Event Interaction test suites than existing metrics. In that metric also describes Event pair Coverage is the best Coverage Compared with Statement, Method, and Branch and Loop coverage. As a future enhancement is to conduct more intensive testing with the candidate tools by creating additional test cases and modifying default test settings to improve test coverage and conducting Automation testing. We will also further introduce the relationship between the test case lengths and the value of n used in CONTeSSi(n) to draw some Conclusions on how to pick the best value of n, starting with a value of n which matches the length of the most test cases in the input suite.
[4] F.Belli, "Finite State Testing and Analysis of Graphical User Interfaces", Proc. the 12th International Symposium on Software Reliability Engineering, 2001, pp34-43. [5] A.M.Memon, M.E.Pollack, M.L.So_a, "Hierarchical GUI Test Case Generation Using Automated Planning", IEEE Transactions on Software Engineering, Vol.27, No.2, 2001, pp144-155. [6] K.Y.Cai, L.Zhao, H.Hu, C.H.Jiang, "On the Test Case De_nition of GUI Testing", Proc. the 5th International Conference on Quality Software, IEEE Computer Society Press, 2005, [7] Brad A. Myers.” User interface software tools". ACM Transactions on Computer-Human Inter-action, 2(1):64103, 1995. [8] Capture-Replay Tool, 2003. http://soft.com. [9] T.Wall. "Abbot Java gui test framework", 2004.http://abbot.sourceforge.net/. [10] Tamas Daboci,Istvan Kollar,Tamas Megyeri,"How to Graphical User Interfaces" IEEE Instrumentation and Measurement Magazine,Sep 2003. [11] A. M. Memon, M. E. Pollack, and M. L. Soffa."Using a goal-driven approach to generate test cases for guis". In 10 Proceedings of the 21st international conference on Software engineering, IEEE Computer Society Press, 1999. [12] A. M. Memon, M. E. Pollack, and M. L. So_a. "Hierarchical gui test case generation using automated planning". IEEE Trans. Softw. Eng.,27(2):144-155, 2001. [13] E. Gamma and K. Beck."Gui tools for junit", 2004. http://www.junit.org/ [14] Atif M Memon."Advances in GUI testing". In Advances in Computers, ed. by Marvin V.Zelkowitz, volume 57. Academic Press, 2003.
REFERENCES [1] A. M. Memon."Gui testing: pitfalls and process". IEEE Computer, 35(8):87-88, 2002. [2] L.White, H.Almezen, "Generating Test Cases for GUI Responsibilities Using Complete Interaction Sequences", International Symposium on Software Reliability Engineering, 2001, pp54-63.Proc. the 11th International Symposium on Software Reliability Engineering, 2000, pp110-121. [3] L.White, H.Almezen, N.Alzeidi, "User-Based Testing of GUI Sequences and Their Interactions", Proc. the 12thInternational Symposium on Software Reliability Engineering,
[15] Monty L. Hammontree , Rothermel,Je_rey J. Hendrickson , Billy W.Hensley,"Integrated data capture and analysis tools for research and testingon graphical user interfaces", Proceedings of the SIGCHI conference on Human factors in computing systems, p.431-432, May 03-07, 1992, Monterey,California, United States [16] Atif M. Memon,"An Event-ow Model of GUI-Based Application for Testing" STVR2007. [17] Richard K. Shehady , Daniel P. Siewiorek," A Method to Automate User Interface Testing Using Variable Finite State
Velammal College of Engineering and Technology, Madurai
Page 186
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Machines, Proceedings of the 27th International Symposium on Fault-Tolerant Computing (FTCS'97), June 25-27, 1997. [18] Lee White , Husain Almezen,"Generating Test Cases for GUI Responsibilities Using Complete Interaction Sequences", Proceedings of the 11th International Symposium on Software Reliability Engineering. [19] Atif M. Memon, Qing Xie."Studying the Fault Detection Effectiveness of GUI Test Cases for Rapidly Evolving Software",Manuscript received 15 Mar. 2005; revised 22 June 2005; accepted 28 June 2005; [20] B. Marick,"When Should a Test Be Automated?" Proc. 11th Intl Software/Internet Quality Week, May 1998. [21] A.Sun, M. Finsterwalder, "Automating Acceptance Tests for GUI Applications in an Extreme Programming
Environment", Proc Second Intl Conf. EXtreme Programming and Flexible Processes in Software Eng. May 2001. [22] L. White, H. AlMezen, and N. Alzeidi, "User-Based Testing of GUI Sequences and Their Interactions", Proc. 12th Intl Symp. Software Reliability Eng., pp. 54-63, 2001. [23] A. Brooks and Atif M.Memon,"Introducing a Test Suite Similarity Metric for Event Sequence-based Test cases", 2007. [24] Clover coverage tool www.atlassian.com/software/ clover [25] www.planetsourcecode.com/calculator source code
Velammal College of Engineering and Technology, Madurai
Page 187
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
THE NEED OF THE HOUR - NOSQL Technology for Next Generation Data Storage K.Chitra1, Sherin M John2 Dept of Computer Applications, Thiagarajar School of Management Madurai, TN, India 1
[email protected] [email protected]
2
Abstract— A relational database is a table-based data system where there is no scalability, minimal data duplication, computationally expensive table joins and difficulty in dealing with complex data. The problem with relations in relational database is that complex operations with large data sets quickly become prohibitively resource intense. Relational databases do not lend themselves well to the kind of horizontal scalability that's required for large-scale social networking or cloud applications. NOSQL has emerged as a result of the demand for relational database alternatives. The biggest motivation behind NOSQL is scalability. NOSQL is meant for the current growing breed of web applications that need to scale effectively. This paper analyzes the need of the next generation data storage which is the need of the current large-scale social networking or cloud applications. We also analyze the capabilities of various NOSQL models like BigTable, Cassandra, CouchDB, Dynamo and MongoDB. Keywords— NOSQL, scalability, relational database alternatives, Next Generation Data Storage.
VI. INTRODUCTION A relational database management system (RDBMS) is a database management system (DBMS) that is based on the relational model as introduced by E. F. Codd. It supports a tabular structure for the data, with enforced relationships between the tables. Most popular commercial and open source databases currently in use are based on the relational model. The problem with RDBMS is not that they do not scale, it’s that they are incredibly hard to scale. The most popular RDBMS are Microsoft SQL Server, DB2, Oracle, MYSQL etc. Many Web applications simply do not need to represent data as a set of related tables that means all applications need not be a traditional relational database management system (RDBMS) that uses SQL to perform operations on data. Rather, data can be stored in the form of objects, graphs, documents and retrieved using a key. For example, a user profile can be represented as an object graph (such as pojo) with a single key being the user id. Another example: documents or media files can be stored with a single key with indexing of metadata handling by a separate search engine.
Velammal College of Engineering and Technology, Madurai
These forms of data storage are not relational and lack SQL, but they may be faster than RDBMS because they do not have to maintain indexes, relationships, constraints and parse SQL. Technology like that has existed since the 1960s (consider, for example, IBM’s VSAM file system). Relational databases are able to handle millions of products and service very large sites. However, it is difficult to create redundancy and parallelism with relational databases, so they become a single point of failure. In particular, replication is not trivial. To understand why, consider the problem of having two database servers that need to have identical data. Having both servers for reading and writing data makes it difficult to synchronize changes. Having one master server and another slave is bad too, because the master has to take all the heat when users are writing information. So as a relational database grows, it becomes a bottleneck and the point of failure for the entire system. As mega e-commerce sites grew over the past decade they became aware of this issue - adding more web servers does not help because it is the database that ends up being a problem.
Fig. 2 Scalability Issues with Relational Database
VII.
NEXT GENERATION DATA STORAGE MODEL
Page 188
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Trends in computer architectures and modern web-scale databases are pressing databases in a direction that requires: • storage of non-relational data, which are distributed to be available even when the nodes are crashed • open-source • effective scalability (Horizontal & Vertical) • schema-free • replication support • easy API • eventual consistency
NOSQL-style data stores attempt to address these requirements. It is the future of database technology that is suitable particularly for web applications that need to scale effectively. Many NOSQL products store data as BLOBs. Actually, NOSQL is an offhand term for a collection of technologies, some of which are DBMS (DataBase Management Systems) and some of which are not. All they have in common is: • They have something to do with managing data • They are not relational DBMS, if by “Relational DBMS” you mean “DBMS that want you to talk to them in SQL.” • select fun, profit from real_world where relational=false; XXXXXXX.
NOSQL (Not Only SQL)
The NoSQL movement is a combination of an architectural approach for storing data and software products that can store data without using SQL. Hence, the term NoSQL. The NOSQL products that store data using keys are called Key-Value (KV) stores. Because these KV stores are not relational and lack SQL they may be faster than RDBMS. The downside of NOSQL is that you cannot easily perform queries against related data. Prominent closed-source examples are Google's BigTable and Amazon's Dynamo. Several opensource variants exist including Facebook's Cassandra, Apache HBase, CouchDB LinkedIn's Project Voldemort and many others. Apache's open source CouchDB offers a new method of storing data, in what is referred to as a schema-free document-oriented database model. 4) The technology There are three popular types of NoSQL databases.
Column-oriented databases: Rather than store sets of information in a heavily structured table of columns and rows with uniform sized fields for each record, as is the case with relational databases, column-oriented databases contain one extendable column of closely related data. Facebook created the high-performance Cassandra to help power its website. The Apache Software Foundation developed Hbase, a distributed, open source database that emulates Google’s Big Table. Document-based stores: These databases store and organize data as collections of documents, rather than as structured tables with uniform sized fields for each record. With these databases, users can add any number of fields of any length to a document. VIII. NOSQL MODELS FOR DATA STORAGE Organizations that collect large amounts of unstructured data are increasingly turning to non-relational databases, now frequently called NoSQL databases. YYYYYYY. Bigtable: A Distributed Storage System for Structured Data Bigtable is Google’s internal database system. Bigtable is a distributed storage system for managing structured data that is designed to scale to a very large size i.e. petabytes (1 petabyte = 1.12589991 × 1015 bytes) of data across thousands of commodity servers. Many projects at Google store data in Bigtable, including web indexing, Google Earth, and Google Finance. These applications place very different demands on Bigtable, both in terms of data size (from URLs to web pages to satellite images) and latency requirements (from backend bulk processing to real-time data serving). Despite these varied demands, Bigtable has successfully provided a flexible, high-performance solution for all of these Google products. A Bigtable is a sparse, distributed, persistent multidimensional sorted map. The map is indexed by a row key, column key, and a timestamp; each value in the map is an uninterpreted array of bytes. Each cell in a Bigtable (like field in DBMS) can contain multiple versions of the same data; these versions are indexed by timestamp (Microseconds). Bigtable timestamps are 64-bit integers. Different versions of a cell are stored in decreasing timestamp order, so that the most recent versions can be read first. Bigtable depends on a cluster management system for scheduling jobs, managing resources on shared machines, dealing with machine failures, and monitoring machine status. ZZZZZZZ. Cassandra: A Decentralized, highly scalable, eventually consistent DB
Key-value stores: As the name implies, a key-value store is a system that stores values indexed for retrieval by keys. These systems can hold structured or unstructured data. Amazon’s SimpleDB is a Web service that provides core database functions of information indexing and querying in the cloud. It provides a simple API for storage and access. Users pay only for the services they use.
Cassandra is highly scalable second-generation distributed database. Cassandra was open sourced by Facebook in 2008 and is currently being developed as an Apache Incubator project. The system offers a fault tolerant, high availability, decentralized store for data which can be scaled up by adding
Velammal College of Engineering and Technology, Madurai
Page 189
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
hardware nodes to the system. Cassandra implements an "eventually consistent" model which trades-off consistency of data stores in the system for availability. Data is automatically replicated to multiple nodes for faulttolerance. Replication across multiple data centers is supported. Failed nodes can be replaced with no downtime. Cassandra is in use at Rackspace, Digg, Facebook, Twitter, Cisco, Mahalo, Ooyala, and more companies that have large, active data sets. The largest production cluster has over 100 TB of data in over 150 machines. AAAAAAAA. CouchDB – Extremely scalable, highly available and reliable Storage System CouchDB, is a free and open source document-oriented database written in the Erlang programming language for its emphasis on fault tolerance, accessible using a RESTful JavaScript Object Notation (JSON) API. The term "Couch" is an acronym for "Cluster Of Unreliable Commodity Hardware", reflecting the goal of CouchDB being extremely scalable, offering high availability and reliability, even while running on hardware that is typically prone to failure. Hence CouchDB is: •
A document database server
•
Ad-hoc and schema-free with a flat address space
•
Distributed, featuring robust, incremental replication
with bi-directional conflict detection and management. •
Highly available even if hardware fails
•
Query-able and index-able, featuring a table oriented
reporting engine that uses Javascript as a query language CouchDB is Not: •
A relational database
•
A replacement for relational databases
•
An object-oriented database. Or more specifically,
meant to function as a seamless persistence layer for an OO programming language BBBBBBBB. Dynamo - A Distributed Storage System for Amazon website Dynamo is an internal technology developed at Amazon to address the need for an incrementally scalable, highlyavailable key-value storage system. The technology is designed to give its users the ability to trade-off cost, consistency, durability and performance, while maintaining high-availability.
Velammal College of Engineering and Technology, Madurai
Amazon runs a world-wide e-commerce platform that serves tens of millions customers at peak times using tens of thousands of servers located in many data centers around the world. Reliability is one of the most important requirements because even the slightest outage has significant financial consequences and impacts customer trust. In addition, to support continuous growth, the platform needs to be highly scalable. Dynamo has a simple key/value interface, is highly available with a clearly defined consistency window, is efficient in its resource usage, and has a simple scale out scheme to address growth in data set size or request rates. Each service that uses Dynamo runs its own Dynamo instances. CCCCCCCC. MongoDB MongoDB (“humongous”) is a document database designed to be easy to work with, fast, and very scalable. It was also designed to be ideal for website infrastructure. It is perfect for user profiles, sessions, product information, and all forms of Web content (blogs, wikis, comments, messages, and more). It’s not great for transactions or perfect durability, as required by something like a banking system. Good fits for MongoDB include applications with complex objects or real-time reporting requirements, and agile projects where the underlying database schema changes often. MongoDB does not suit software with complex (multiple objects) transactions. 1) Inside MongoDB MongoDB is an interesting combination of modern Web usage semantics and proven database techniques. In some ways MongoDB is closer to MySQL than to other so-called “NoSQL” databases: It has a query optimizer, ad-hoc queries, and a custom network layer. It also lets you organize document into collections, similar to sql tables, for speed, efficiency, and organization. To get great performance and horizontal scalability however, MongoDB gives something up: transactions. MongoDB does not support transactions that span multiple collections. You can do atomic operations on a single object, but you can’t modify objects from two collections atomically. MongoDB stores BSON, essentially a JSON document in an efficient binary representation with more data types. BSON documents readily persist many data structures, including maps, structs, associative arrays, and objects in any dynamic language. Using MongoDB and BSON, you can just store your class data directly in the database. MongoDB is also schema-free. You can add fields whenever you need to without performing an expensive change on your database. Adding fields is also quick, which is ideal for agile environments. You need not revise schemas from staging to production or have scripts ready roll changes back.
Page 190
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IX. ANALYSIS OF NOSQL In this section we compare and analyze the nonfunctional requirements of NOSQL databases named BigTable, Cassandra, CouchDB, Dynamo, MongoDB, HBase. Challenges of NOSQL NOSQL databases face several challenges. Overhead and complexity: NOSQL databases don’t work with SQL; they require manual query programming, which can be fast for simple tasks but time-consuming for others. In addition, complex query programming for the databases can be difficult. Reliability: NOSQL databases do not support ACID properties which are natively supported by relational databases. NOSQL databases thus do not natively offer the degree of reliability that ACID provides. If users want NOSQL databases to apply ACID restraints to a data set, they must perform additional programming. Consistency: Because NOSQL databases do not natively support ACID transactions, they also could compromise consistency, unless manual support is provided. Not providing consistency enables better performance and scalability but is a problem for certain types of applications and transactions, such as those involved in banking. Unfamiliarity with the technology: Most organizations are unfamiliar with NOSQL databases and thus may not feel knowledgeable enough to choose one or even to determine that the approach might be better for their purposes. Limited Eco structure: Unlike commercial relational databases, many open source NOSQL applications do not yet come with customer support or management tools. DB Type
Bigtable Cassandra CouchDB Dynamo
MongoDB
Scalability
Key-Value store DB Column Oriented DB
Highly Scalable
JSON Document Oriented DB Key-Value store DB
Easily scalable and readily extensible Incremental
BSON Document Oriented DB
Scalable
Highly Scalable
scalable. The original intention has been modern web-scale databases. Often more characteristics apply as: schema-free, replication support, easy API, eventually consistency, and more. Hence the misleading term "NoSQL" is now translated to "Not Only SQL (NOSQL)". NOSQL databases generally process data faster than relational databases. Developers usually do not have their NOSQL databases supporting ACID properties, in order to increase performance, but this can cause problems when used for applications that require great precision. NOSQL databases are also often faster because their data models are simpler. “There’s a bit of a trade-off between speed and model complexity, but it is frequently a tradeoff worth making,” Because they do not have all the technical requirements that relational databases have. Major NOSQL systems are flexible enough to better enable developers to use the applications in ways that meet their needs. REFERENCES [43] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E. Gruber, ‘Bigtable: A Distributed Storage System for Structured Data’, OSDI'06: Seventh Symposium on Operating System Design and Implementation, Seattle, WA, November, 2006. [44] Hamilton, James, "Perspectives: One Size Does Not Fit All", 13 November 2009. Lakshman, Avinash; Malik, Prashant, Cassandra - A Decentralized [45] Structured Storage System. Cornell University, 13 November 2009.
Availability
Performance
Highly Available High availability is achieved using replication Highly Available Highly Available High writeavailability
TABLE II NONFUNCTIONAL REQUIREMENTS ANALYSIS OF NOSQL DATABASES
X. CONCLUSION Next Generation Databases mostly address some of the points: being non-relational, distributed, open-source and horizontal
Velammal College of Engineering and Technology, Madurai
High performance Low Loading speeds are better than retrieval speeds Performance at massive scale is one of the biggest challenges Excellent solution for short read
Reliability
Provides reliability at a massive scale. At massive scale is a very big challenge Reliable and efficient system Reliability at massive scale is one of the biggest challenges Low
[46] Chang, Fay; Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E. Gruber. Bigtable: A Distributed Storage System for Structured Data. Google. 13 November 2009. [47] Kellerman, Jim. "HBase: structured storage of sparse data for Hadoop", 13 November 2009. [48] Ian Eure, Looking to the future with Cassandra. Digg Technology Blog, September 2009
Page 191
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[49] Zhou Wei, Guillaume Pierre and Chi-Hung Chi. CloudTPS: Scalable Transactions for Web Applications in the Cloud. Technical report IR-CS-53, VU University Amsterdam, February 2010. [50] http://www.allthingsdistributed.com/2007/12/eventually_consistent . [51] http://s3.amzonaws.com/AllThings/Distributed/sosp/amazondynamo [52] http://labs.google.com/papers/bigtable.html http://www.cs.brown.edu/~ugur/osfa.pdf [53] [54] http://www.readwriteweb.com/archives/amazon_dynamo.php [55] http://nosql-database.org/ [56] http://couchdb.apache.org/
Velammal College of Engineering and Technology, Madurai
Page 192
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
AN EFFICIENT TURBO CODED ofdm system Prof. Vikas Dhere Hirasugar Institute of technology Nidasoshi ( Belgam ) [email protected]
Abstract: Orthogonal frequency division multiplexing (OFDM) is a promising technique for high data rate wireless communications. In this paper, we propose ‘An Efficient Turbo Coded OFDM system’ that can improve the performance by making all the Turbo coded symbols having the same reliability for OFDM transmission. We perform computer simulations using the Log-maximum-aposteriori (Log-Map) algorithm in order to access the performance of the proposed system. The simulation results show that, the Turbo coded OFDM system gives the best performance compared to conventional coded systems. Keywords: Turbo Codes, Orthogonal Frequency Division Multiplexing (OFDM).
invoking a longer interleaver size. We perform computer simulation using the LOG_MAP algorithm for iterative decoding to access the performance of the proposed system because of its accuracy and simplicity. The rest of the paper is organized as follows: section 2. Describes the OFDM fundamental and OFDM system model used to obtain the performance result presented in this paper. Section 3. Follows with a description of the turbo encoding and decoding and
1. Introduction Error control systems concern themselves with practical ways of achieving very low bit error rates after a transmission over a noisy band limited channel. Several error correcting techniques are deployed and their performances can be measured by comparing them each other and to the theoretical best performance given by Shannon channel information capacity theory. Following the invention of Turbo Codes in 1993, a number of turbo coded systems exploiting the powerful error correction capability have been developed. Many factors affect the performance of Turbo codes. They include the interleaver size and structure, the generator polynomial and constraint length, the decoding algorithm, the number of iterations, and so on. The presence of multipath with different time delays causes inter-symbol interference (ISI) and degrades the performance of the transmission system. To improve performance of the system, complex adaptive equalization techniques are required. Recently, OFDM systems have been attracting much attention, because of its robustness against frequency-selective fading. Some examples of existing systems, where OFDM is used, are digital audio and video broadcasting, asymmetricdigital-subscriber line modems, and wireless local-areanetworks systems, such as IEEE 802.11 and HiperLan/2. In mobile environment, since the radio channel is frequency selective and time-varying, channel estimation is needed at the receiver before demodulation process. Orthogonal Frequency Division Multiplexing (OFDM) could be tracked to 1950’s but it had become very popular at these days, allowing high speeds at wireless communications. Orthogonal frequency division multiplexing (OFDM) is becoming the chosen modulation technique for wireless communications. In this paper, we propose a scheme that can improve greatly the performance of turbo-coded OFDM systems without
2.
OFDM System
2.1
OFDM Fundamentals
Velammal College of Engineering and Technology, Madurai
turbo coded OFDM system. The computer simulation results are shown and discussed in section4. While section 5. Concludes the paper.
The basic principle of OFDM is to split a high rate data stream into a number of lower rate streams that are transmitted simultaneously over a number of sub-carriers. Because the symbol duration increases for the lower rate parallel sub carriers, the relative amount of dispersion in time caused by multi-path delay spread is decreased. Inter symbol interference (ISI) is eliminated almost completely by introducing a guard time in every OFDM symbol. In the guard time, the OFDM symbol is cyclically extended to avoid inter carrier interference.
Fig. 1. OFDM Transmitter and Receiver channel
A basic OFDM system consists of a QAM or PSK modulator/demodulator, a serial to parallel / parallel to serial converter, and an IFFT/FFT module. OFDM is a special case of multi-carrier transmission, In Fig.1. a block diagram of the baseband model of OFDM system is shown. The binary information is first mapped using baseband modulation schemes such as QAM or PSK. Then the serial-to-parallel conversion is applied to baseband modulated signals. The serial-to parallel converted data is modulated using Inverse Fast Fourier Transform (IFFT). This is one of
Page 193
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the main principles of the OFDM systems. After IDFT, the time domain signal is given as following equation;
where N is the length of FFT, S(k) is baseband data sequence. After IFFT, the guard interval called as cyclic prefix is inserted to prevent inter-symbol interference (ISI). This interval should be chosen to be larger than the expected delay spread of the multipath channel. The guard time includes the cyclically extended part of the OFDM symbol in order to eliminate the inter-carrier interference (ICI). The symbol extended with the cyclic prefix is given as follows;
Otherwise
Where: T Symbol length; time between two consecutive OFDM symbols
Where Nc is the length of the cyclic prefix. The resultant signal st(n) will be transmitted over frequency selective time varying fading channel with additive white Gaussian noise (AWGN). The received signal is given by following equation;
Where h(n) is the impulse response of the frequency selective channel and w(n) is AWGN. At the receiver, the cyclic prefix is first removed. Then the signal y(n) without cyclic prefix is applied to FFT block in order to obtain following equation;
After FFT block, assuming there is no ISI, demodulated signal is given by following equation; Where H(k) is FFT[h(n)], I(k) is ICI and W(k) is FFT [w(n)]. 2.2. OFDM Modeling Mathematically, the OFDM signal is expressed as a sum of the prototype pulses shifted in time- and frequency and multiplied by the data symbols. In continuous-time notation, the kth OFDM symbol is written as
( 7) (6)
Velammal College of Engineering and Technology, Madurai
Finally, continuous sequence of transmitted OFDM symbols is
Page 194
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
expressed as
Fig. 3.
(8)
Turbo decoder block diagram
3.2 MAP A map decoder chooses the most likely path so as to maximize the a posteriori probabilities and that can be mathematically presented as:
3. Turbo Encoding and Decoding A ½ rate recursive systematic convolutional code with constraint length three is used as the component of a turbo coding scheme. The encoder employs parallel concatenation as illustrated in Fig. 2. The binary information to be encoded is represented as uk, with ck,1 and ck,2 signifying the parity bits of the first and second component encoders, and where ‘D’ stands for a bit delay. Symbol ‘π’ represents pseudo-random interleaver, which accepts blocks of 8000 information bits.
which is the probability of the codeword a particular sequence
conditioned on that
was received. The quantity
is known as a posteriori for a MAP decoder.
is the joint probability 3.3
Design Criteria
Binary information, uk, is fed to the component encoders in order to generate the associated parity bits, ck,1 and ck,2,which are selected alternately by the puncturing block. In other words, for a binary input sequence of u1, u2, u3, u4, the encoded sequence would be u1, , u2, C1 ,C2, U3,1, u4, c4,2. The fact that the input sequence uk is retained at the output, is due to the ‘systematic’ nature of the encoder.
1. First, the constituent encoders were selected yielding good performance at low signal-to noise ratios, with particular attention to decoding complexity. The final choice was the optimum 4-state recursive systematic encoder. 2. Next, the turbo-code interleaver was designed based on the codeword weight distribution and on the achievable performance on the additive white Gaussian noise (AWGN) channel, using a maximum likelihood approach. 3. Finally, the puncturing schemes were selected based again on both the weight distribution and the achievable performance on the AWGN channel .
3.1. Turbo Decoding
3.4 Turbo Coded OFDM System
The following Fig.3 shows the turbo decoder which contains two component decoders concatenated in parallel, performing the log-MAP algorithm. Let ki is the extrinsic information related to the kth information symbol, rk, provided by the ith decoder, and yk,i is the parity symbol associated with rk, and fed into the ith component decoder.
The block Diagram of proposed Turbo coded OFDM system is as shown in Fig. 4.
Fig. 2. Turbo encoder block diagram
Fig. 4 A block Diagram of Turbo coded OFDM System
Velammal College of Engineering and Technology, Madurai
Page 195
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A generic turbo encoder first encodes the input data sequence . We use the component code with the constraint length K=3 and the generator polynomilas ( 7,5). We chose a code rate of ½ by puncturing two redundancy bits alternatively. After serial-to-parallel conversion for OFDM symbol stacking, each coded symbol is assigned to a distinct orthogonal code with the same length as the number of effective subcarriers to be transmitted. For the single turbo code adaptive OFDM scheme, the information bits are firstly turbo coded by a single turbo code, then modulated separately for each subband. In the same way, at the receiver, the signal from the output of the FFT is demodulated separately then decoded as a single turbo code frame. In the separate turbo code scheme , the OFDM frame is divided into several subbands , and the turbo code frame combines the same subband of several OFDM symbol. In the single turbo code scheme, the OFDM frame is also divided into several subbands, each subband using a different modulation scheme. 4.
4.1.2 Effect of varying number of iterations on performance Frame size = 700 Number of Iterations = 1 Eb/N0(dB)
0
0.2
0.4
0.6
BER
0.1316
0.1519
0.1098
0.1081
Eb/N0(dB)
0.8
1
1.2
1.4
BER
0.0926
0.0719
0.0663
0.0503
Tab. 2. Performance with number of iteration 1 Number of iterations = 4 Eb/N0(dB)
0
0.2
0.4
0.6
BER
0.1616
0.1298
0.0602
0.0321
Eb/N0(dB) BER
0.8 0.0106
1 0.0011
1.2 1.527e-4
1.4 2.6551e4
SIMULATION RESULTS
In this paper, we propose an efficient wireless transmission method that can improve greatly the performance of turbocoded OFDM systems. 4.1 SIMULATION SETUP 4.1.1 Turbo codes We are selecting the AWGN channel. The turbo encoder used in our simulation is of rate ½ by puncturing two redundancy bits alternatively and constraint length K =3 and generator polynomial GR ( D) = ( 7,5). The Log-MAP map algorithm is used for decoding with 08 iterations. The recursive least square algorithm is used for channel estimation. Simulation results for a turbo code are based on bit error rate (BER) performance over a range of Eb/No. The following table shows the rate 1/2 component RSC codes used in the simulation results. The following Tab. 1 shows the configuration used for simulations. Channel Component Encoders RSC Parameters Component Decoder Iterations
Tab.3. Performance with number of iteration 4 Number of Iterations 8 Eb/N0(dB) BER Eb/N0(dB) BER
0 0.1225 0.8 0.0034
0.2 0.0942 1 0.0014
0.4 0.0729 1.2 1.713e-4
0.6 0.0208 1.4 2.7042e4
Tab.4. Number of iterations = 8
Additive White Gaussian( AWGN) 2 Identical Recursive Convolutional codes ((RSC) n=2,k=1,k=2, G 0 = 7, G1 = 5 Log MAP 8
Tab. 1. Simulation parameters Turbo codes
Velammal College of Engineering and Technology, Madurai
Fig. 5 Performance comparison of different iteration times
Page 196
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The results show that as we increase the number of iterations for the same channel, the bit error rate decreases by increasing the signal to noise ratio. 4.2 . OFDM The following Tab.5 shows the configuration used for the simulations performed on the OFDM signal. QPSK 8-PSK 16-PSK FFT size Number of carrier used Guard time Guard Period Type
SNR = 10 dB SNR = 15 dB SNR = 25 dB 2048 800 512 samples Half zero signal , half a cyclic extension of the symbol
= 20 µs . the sample rate is 16 MHz which corresponds to a sample time of Ts= 0.0625 µs. The channel model is AWGN. The recurceive systematic encoder in the Turbo encoder is rate ½ with constraint length K = 3 and a generator polynomial GR ( D) = ( 7,5). In the simulation for both systems, we set the BER target to 10-2. Fig.7 illustrates the BER performance of a Turbo coded OFDM system with the optimal adaptation algorithm. The red solid line with circle in figure shows the BER performance for the separate turbo code system while the blue solid line with square in Fig.6 is the BER performance for the single code system. Both of the systems fulfill the BER target, but the single turbo code system is closer to it.
Tab.5. Simulation Parameters OFDM The following Fig. 6 shows the simulation results and the performance of different modulation scheme in an AWGN channel.
Fig. 7 BER of separate turbo code system and single turbo code system
5. CONCLUSION Red
QPSK
Green Blue
8-PSK 16-PSK
As shown in the numerical results the single turbo code adaptive system provides better performance for longer turbo codes.
Fig. 6 BER curves of different modulation schemes for OFDM system in AWGN channel
6.Refernces.
The results show that the lower order modulation schemes has low bit error rate as compared to higher order modulation at a given SNR.
1..Irving Reed, Xuemin Chen, Error-control coding for Data Networks,Kluwer Academic Publishers, Norwell, Massachusetts, USA, 1999. 2. Hosein Nikopour, Amir K. Khandani and S. Hamidreza Jamali, “Turbo coded OFDM transmission over a nonlinear channel,” IEEE transactions on vehicular technology, vol.54, no.4, (July 2005) 1361-1371.
4.3. TURBO CODED OFDM The simulation system uses Turbo Code –OFDM with N = 256 subcarriers out of which only 16 of them are pilot subcarriers. In the simulated system we use a Cyclic Prefix length of 65 samples , and OFDM symbol duration is TOFDM
Velammal College of Engineering and Technology, Madurai
Page 197
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3. B. Bougard, L.van der Perre, B. Gyselinckx, M. Engels,“Comparison of several FEC schemes for wireless OFDM transmission on frequency selective channels,”. 4. E.R Berlekamap, Algebraic coding Theory, New York. McGraw-Hill 1968. 5. C. Berrou and A. Flavieux, “Near-optimum error correcting coding and decoding. Turbo codes”, IEEE Trans. Common, Vol. 44, PP 1261-1271, Oct. 1996. 6. C. Kikkert, “Digital Communication Systems and their Modulation Techniques”,James Cook University, October 1995. 7. J. G. Proakis, Digital Communications, McGraw-Hill, New York, USA, 4th edition 2000. 8. Pyndiah, R.; Picart, A.; Glavieux, A., "Performance of block turbo coded 16-QAM and 64-QAM modulations," GLOBECOM '95., IEEE , vol.2, no.pp.1039-1043 vol.2, 14-16 Nov 1995 9. S. Le Goff, A. Glavieux and C. Berrou, “Turbo codes and high spectral efficiency modulation,” Proc. of the IEEE International Conference on Communications, vol. 2, pp. 645-649, May 1994.
Velammal College of Engineering and Technology, Madurai
Page 198
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Advancement in mobile technology Using BADA V.AISHWARYA* UG STUDENT DEPT OF ECE KALASALINGAM UNIVERSITY KRISHNAN KOIL [email protected]
J.MANIBHARATHI UG STUDENT DEPT OF ECE KALASALINGAM UNIVERSITY KRISHNAN KOIL [email protected]
Abstract : Bada is a complete smart phone platform. Bada runs high-performance native apps and services. With this use of BADA inour mobile phone the compatibility ,accessibility and the interface of the mobile phone version onescan be considerably increased. The use of Badaallows us to implement mobile with no control technology and speech with gesture recognitionfacility in the mobile devices. Also further the development of BADA allows us to install hand tracking detection in our mobile phones.in this paper ,feasibilities of using Bada in mobile is discussed indetail
Keywords: Software development kit(SDK),Application program interface(API),WII.
1.0 Introduction There were nearly 2000 million cell phones sold each year compared with fewer than 200 million PCs — and the gap was widening. Increasingly, phones were the way people wanted to connect with each other and with everything else. Phones were going to replace PCs as the main gateway to the Internet, and they were going to do it soon. The cell phones ran on different software, had less memory, and operated under the constraints of pay-per-byte wireless networks, the mobile Web was a stripped-down, mimeographed version of the real thing. To avoid this, Bada had the solution. Bada is a free, open source mobile platform that any coder could write for and any handset maker could install It would be a global, open operating system for the wireless future. Bada, (bɑdɑ, the Korean word for “ocean”,) is a new smartphone platform that allows developers to create featurerich applications that elevate the user experience in mobile spaces. The bada platform is kernel-configurable and it can run either on the Linux kernel or real-time OS kernels and it was first developed by Samsung.
1.1Building a better mobile phone
Velammal College of Engineering and Technology, Madurai
Dr.S.DURAI RAJ PROFESSOR & HOD DEPT OF ECE KALASALINGAM UNIVERSITY KRISHNAN KOIL [email protected]
To build a better mobile phone, a mobile must have the ability to record and watch videos with the camcordermode.Uploading videos to YouTube and pictures to Picasa directly from the phone .A new soft keyboard with an "Auto complete" feature. Ability to automatically connect to a Bluetooth headset within a certain distance. New widgets and folders that can populate the desktop. Animations between screens and moving sensors to play WII games .Expanded ability of Copy and paste to include web pages. Ability to embed the ADOBE flash player and webkit internet browser directly. Bada enables developers to take full advantage of mobile device capabilities to easily create compelling applications. Different service applications can share information such as personal profiles, social relations, schedules, and contents with simple user confirmation, all in order to provide services with greater personal relevance and cooperative service extensions. For example, social networking applications can share user information with commerce or location-aware applications, and share photos from other content publishing applications. The bada platform is kernel-configurable so that it can run either on the Linux kernel or real-time OS kernels, which makes bada applicable to a wider range of devices than any other mobile operating system.
2. Bada Runtime Bada run on any platform . Java run timeand dalvik virtual machine and other computing platform. 2.1 Dalvik Virtual Machine Every bada application runs in its own process, with its own instance of the Dalvik virtual machine. Dalvik has been written so that a device can run multiple VMs efficiently. The Dalvik VM executes files in the Dalvik Executable (.dex) format which is optimized for minimal memory footprint. The VM is register-based, and runs classes compiled by a Java language compiler that have been transformed (as ByteCodes, which is good for run fast) into the .dex format by the included "dx" tool.
Page 199
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.2 Linux Kernel Bada relies on Linux version 2.6 for core system services such as security, memory management, process management, network stack, and driver model. It also act as an abstraction layer between the hardware and rest of the software stack.
3.0 Supported Operating Systems 1. Windows XP (32-bit) or Vista (32- or 64-bit)- 1.4 GB RAM or higher . 2. 1.8 GB free disk space or higher. Simulator screen size is 480 x800, Mac os and linux.
3.1 Supported Development Environments 1. Eclipse IDE , Eclipse 3.3 (Europa), 3.4 (Ganymede) . Recommended Eclipse IDE packages: Eclipse IDE for Java EE Developers, Eclipse IDE for Java Developers, Eclipse for RCP/Plug-in Developers. Eclipse JDT plugin. 2. Eclipse Classic IDE package is not supported. 3. 4. JDK 5 or JDK 6 (JRE alone is not sufficient). JAVA FX 1.1 and HELIX. 3.2 Supported web browser environment Internet Explorer 6.0 or higher, Firefox 3.0 or higher and Safari 3.2 or higher.
4.0
Bada Architecture
Bada has a four-layer architecture: kernel, device, service, and framework layers. The kernel layer can be the Linux kernel or a real-time OS kernel, depending on the hardware configuration. The device layer provides the core functions as a device platform, such as system and security management, graphics and windowing system, data protocols and telephony, as well as audio-visual and multimedia management. The service layer provides service-centric functions that are provided by application engines and webservice components interconnecting with bada Server. Only the framework layer can export the C++ open API. The framework consists of an application framework and functions exported by the underlying layers.
5.0 Software Development Kit (SDK) It is a typically a set of development tools that allows a engineer to create applications for a certain software package, software framework, hardware platform, computer system, video game console, operating system or similar platform.
5.2
Basic data types
Object, String, Date Time, Byte Buffer, uid, and other base types. It has Wrapper classes for C++ primitive types, such as Integer, Short, Double, and Long. Its has certain condition for run time. It is interms of timer thread and its Synchronization with Mutex, Semaphore, and Monitor. The other basic data types are Collection and its utility. The SDKfeatures in data types are, ArrayList, HashMap, Stack, and other collection types. Object-based collections and template-based collections.. Math, StringUtil, StringTokenizer, and Uri with Standard library support. C++ STL (Standard Template Library)
5.3 Osp::Io Namespace for input/output data handling. It contains file and Directory. The database involves transaction such as begin, commit and rollback. Its also Support of both SQL queries and enumerations with DB statements. It has registry data IO which involves system-provided data store in which applications store and retrieve settings.
5.4 Osp::Text and Osp::Locales It has namespaces for internationalization and localization. Conversion between the major character encoding schemes, such as UTF8, UCS2, Latin1, GSM, and ASCII and it has separate locale by Identifying a specific language code, country code, and variant code and its Formatting localesensitive information, such as date, number, and currency and it has the ability to Convert a date, time, and number into a string Unicode.
5.5 Osp::System Namespace for controlling devices and getting system information. It has the separate device control with register an alarm. 1. Screen power management: 2. It has Keep LCD on with or without dimming . 3. Vibrator activation with level, on/off period, and count 4. System information: 5. Getting uptime and current time in UTC, Standard, and Wall time mode. The features of sdk in Bada relates with all mobile phone applications like connection of networking which is in the form of Osp::Net, and for multimedia purpose in Osp::Media . It has separate security systems and it is different from all other mobile phones. 6.
5.6 Osp::Security 5.1 Bada SDK features The sdk features is represented in osp: base model. It has several osp for each and every application. There are separate Namespaces for basic types, execution environment, and utilities.
Velammal College of Engineering and Technology, Madurai
It has separate namespace for cryptography keys, algorithms, and certificate management. It has Cryptography keys and PRNG (Pseudo Random Number Generator).The Cryptographic algorithms, such as HMAC, Hash, AES, DES, and RSA occur in basic bada. The X.509 certificate information retrieval and certificate path validation.
Page 200
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6.0 Future SDK Development In Bada The bada system is going to deliver the SDK, which gives compatible to bada powered handsets. The system has fully comfort and include the map external library. The bada system delivers an updated version of the frame work API after the bada s8500 devices. As with previous sensors, the bada has sensor bar with softwares of WII games and also sensor software. The updated API is assigned an integer identifier that is stored in the system itself. It is used to determine whether the application is compatible with the system, prior to installing the applications.
7.0
InBuilt Applications
In the developer application the inbuilt of sensors with software provide the best playment of WIIgames and other 3D games. Some spare parts are software of wii games which is installed through API and sensors to capture from the wii remote. Other applications are as same as the previous bada SDK. The spare parts are the includement of map external library and other external libraries. The software of wii and sensors are transferred through Bluetooth so it brings autopairing effect. The wii sensor is also effective in password security, secret fixing of cameras, effective tool to speech and hearing impaired people . Hacking can be avoided by these sensors. It is applicable to WII 2 games also.
a.
Wii Remote
The main feature of wii remote is the gesture recognition and pointing devices. The gesture recognition detects motion and rotation in three dimension through the use of accelerometer technology. By separating the controller from the gaming console, the accelerometer data can be used as an input for the gesture recognition. The accelerating sensors represented the gestures by characteristic patterns of incoming signals data ie. Vectors, representing the current accelerations of controller in all three dimensions. By filtering and quantizer technique the unwanted vector data can be identified and cleared. This can be done by k-mean algorithm. In this recognition all different size objects, and other images in 3D can be recognized.
b.
Vector Quantization
Like other acceleration-sensors the one integrated into the Wiimote delivers too much vector data to be put into a single HMM. In order to cluster and abstract this data the common kmean algorithm is applied with k being the number of clusters or codes in the so-called codebook. A codebook size delivering satisfying results and Its empirically identified k = 8 for gestures in a two-dimensional plane. However, adopt the idea of arranging the 8 cluster centres on a circle by extending it to the 3D case. Instead of distributing the centres uniformly
Velammal College of Engineering and Technology, Madurai
on a two-dimensional circle its possible to put them on a threedimensional sphere, intersecting two circles orthogonal to each other. Consequently this leads to k = 8 + 6 = 14 centres. The radius of each circle/sphere dynamically adapts itself to the incoming signal data.
9.0 Implementation In Bada The implementation of the gesture recognition in Java using the standardization of Java APIs for Bluetooth Wireless Technology (JABWT) defined by the JSR-82 specification. Using Java ensures platform independency, for developing and testing purposes the use of GNU/Linux platform with the Avetana Bluetooth implementation2 is needed. The recognition process is realized as a reusable and extensible gesture recognition library based on an event-driven design pattern. The library provides an interface for basic functions, e.g. acceleration readout with the Wii Listener interface, as well as recognition functions using a Gesture-Listener interface. Through its modularity it is easy to adapt the prototype to other acceleration-based controllers.
9.1 Sensing The wii remote has the ability to sense acceleration along with three axes through the use of an ADXL 330 accelerometer. The wii remote also has the feature of pix art optical sensors, allowing it to determine where the wii remote is pointing. Unlike a light gun that senses light from a mobile screen, the wii remote senses light from the sensor which is inbuilted in the form of chip in the bada phones allows consistent usage regardless of the mobile sizes. The infrared LED’s in the sensor gives an accurate pointer. The sensor allows the wii remote to be used as an accurate pointing device 30 meters away from the mobile. The wii remote’s image sensor is used to locate the sensor’s points of light in the wii remote’s field of view. The inbuilted wii software in the mobile devices calculates the distance between the wii remote and sensor using triangulation. Games can be programmed to sense whether the image sensor is covered, which is demonstrated in a microgame of smooth moves, where if the player does not uncover the sensor, the champagne bottle that the remote represents will not open.
9.2 Map External Library The map external library is made to add powerful mapping capabilities to this sdk and the good application programme interface add on includes a map external library. The classes of the map library offer built in downloading, rendering, caching of map files as well as the variety of display options and controls. The key class in the maps library is map view and a subclass of view group is in the bada standard library.capture key presses and touch gestures to pan and zoom the map automatically, including handling network requests for additional map files. It also provides all of the UI elements necessary for users to control the map.
Page 201
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
10.0 Summary The implementation of Bada in the mobile phones leads to lot of facilties in each and every applications. The introduction of WII in Bada ensures gesture recognition and its sensing is used as the basic technique for no control technology. Bada has the special feature to have Bluetooth in the standardized Java API, in which no other mobile phones have that features. This analysis lead to the highest cost reduction in WII games if it is implemented.
11.0 References 1. Wii: The Total Story IGN. http://wii.ign.com/launchguide/hardware1.html Retrieved 2006-11-202 2. Consolidated Financial Highlights(PDF). Nintendo. 200907-30. p. 9. http://www.nintendo.co.jp/ir/pdf/2009/090730e.pdf#page=9, Retrieved 2009-10-29. 3 Electronic Arts (2008-01-31). Supplemental Segment Information http://media.corporateir.net/media_files/IROL/88/88189/Q3FY08SupSeg.pdf#page= 4 Retrieved 2008 (PDF). Thomson Financial. p. 4. -02-09.. 4. Wii Remote Colors. news.com. http://news.cnet.com/i/ne/p/2005/0916nintendopic4_500x375.j pg. Retrieved 2006-07-15. 5. http://badawave.com, what is bada, Retrieved 7 April 2010.
Velammal College of Engineering and Technology, Madurai
Page 202
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Mixed-Radix 4-2 Butterfly FFT/IFFT For Wireless communication A.Umasankar, S.Vinayagakarthikeyan EEET- Department, ECE-Department Yanbu Industrial College, Yanbu, Kingdom Of Saudi Arabia, Vel’s Srinivasa College of Engineering and Technology, Chennai, India [email protected] [email protected]
Abstract The technique of orthogonal frequency division multiplexing (OFDM) is famous for its robustness against frequency-selective fading channel. In general, the fast Fourier transform (FFT) and inverse FFT (IFFT) operations are used as the modulation/demodulation kernel in the OFDM systems, and the sizes of FFT/IFFT operations are varied in different applications of OFDM systems. The modified Mixed Radix 4-2 Butterfly FFT with bit reversal for the output sequence derived by index decomposition technique is our suggested VLSI system architecture to design the prototype FFT/IFFT processor for OFDM systems. In this paper the analysis of several FFT algorithms such as radix-2, radix-4 and split radix were designed using VHDL. The results show that the proposed processor architecture can greatly save the area cost while keeping a high-speed processing speed, which may be attractive for many real-time systems. Key words: OFDM, FFT/IFFT, VLSI, VHDL, Mixed Radix with bit reversal.
I.
Introduction
Fast Fourier transform (FFT) are widely used in different areas of applications such as communications, radars, imaging, etc. One of the major concerns for researchers is the enhancement of processing speed. However according to use of portable systems working with limited power supplies, low-power techniques are of great interest in implementation of this block. FFT and FFT blocks are used in OFDM links such as very high speed digital subscribe line (VDSL), Digital Audio Broadcasting (DAB) systems and microwave portable links. Of course there are many ways to measure the complexity and efficiency of an algorithm, and the final assessment depends on both the available technology and the intended application. Normally the number of arithmetic multiplications and additions are used as a measure of computational complexity. Several methods of for computing FFT (and IFFT) are discussed in [1],[2],[3]. These are basic algorithms for implementation of FFT and IFFT blocks. ORTHOGONAL FREQUENCY DIVISION MULTIPLIXING
Velammal College of Engineering and Technology, Madurai
The FDM system above had been able to use a set of subcarriers that were orthogonal to each other; a higher level of spectral efficiency could have been achieved. The guard bands that were necessary to allow individual demodulation of sub carriers in an FDM system would no longer be necessary. The use of orthogonal subcarriers would allow the subcarriers’ spectra to overlap, thus increasing the spectral efficiency. As long as orthogonality is maintained, it is still possible to recover the individual subcarriers’ signals despite their overlapping spectrums. If the dot product of two deterministic signals is equal to zero, these signals are said to be orthogonal to each other. Orthogonality can also be viewed from the standpoint of stochastic processes. If two random processes are uncorrelated, then they are orthogonal. Given the random nature of signals in a communications system, this probabilistic view of orthogonality provides an intuitive understanding of the implications of orthogonality in OFDM. Later in this article, we will discuss how OFDM is implemented in practice using the Fast Fourier transform (FFT) show fig 1. Recall from signals and systems theory that the sinusoids of the FFT form an orthogonal basis set, and a signal in the vector space of the FFT can be represented as a linear combination of the orthogonal sinusoids. One view of the FFT is that the transform essentially correlates its input signal with each of the sinusoidal basis functions. If the input signal has some energy at a certain frequency, there will be a peak in the correlation of the input signal and the basis sinusoid that is at that corresponding frequency. This transform is used at the OFDM transmitter to map an input signal onto a set of orthogonal subcarriers, i.e., the orthogonal basis functions of the DFT. Similarly, the transform is used again at the OFDM receiver to process the received subcarriers show fig 2. The signals from the subcarriers are then combined to form an estimate of the source signal from the transmitter. The orthogonal and uncorrelated nature of the subcarriers is exploited in OFDM with powerful results. Since the basis functions of the FFT are uncorrelated, the correlation performed in the FFT for a given subcarrier only sees energy for that corresponding
Page 203
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
subcarrier. The energy from other subcarriers does not contribute because it is uncorrelated. This separation of signal energy is the reason that the OFDM subcarriers’ spectrums can overlap without causing interference.
have different radices. For instance, a 64-point long FFT can be computed in two stages using one stage with radix-8 PEs, followed by a stage of radix-2 PEs. This adds a bit of complexity to the algorithm compared to radix-r, but in return it gives more options in choosing the transform length. The Mixed-Radix FFT algorithm is based on sub-transform modules with highly optimized small length FFT which are combined to create large FFT. However, this algorithm does not offer the simple bit reversing for ordering the output sequences. Mixed-Radix FFT Algorithms with Bit Reversing The mixed-radix 4/2 [9] butterfly unit is shown in Figure 2. It uses both the radix-2^2 and the radix-2 algorithms can perform fast FFT computations and can process FFTs that are not power of four. The mixed-radix 4/2, which calculates four butterfly outputs based on X (0) ~X (3). The proposed butterfly unit [10], [11] has three complex multipliers and eight complex adders. Four multiplexers represented by the solid box are used to select either the radix-4 calculation or the radix-2 calculation.
Fig.1 OFDM Transmitter and Receiver
II. FAST FOURIER TRANSFORM The Discrete Fourier Transfer (DFT) plays an important role in many applications of digital signal processing including linear filtering, correlation analysis and spectrum analysis etc. The DFT is defined as:
(1) Fig 2: The basic butterfly for mixed-radix 4/2 DIF FFT algorithm.
is the DFT coefficient. Where Evaluating the Equation (1) directly requires N complex multiplications and (N-1) complex additions for each value of the DFT. To compute all N values therefore requires a total of N^2 complex multiplications and N(N-1) complex additions. Since the amount of computation, and thus the computation time, is approximately proportional to N2, it will cost a long computation time for large values of N. For this reason, it is very important to reduce the number of multiplications and additions. This algorithm is an efficient algorithm to compute the DFT [4],[5], which is called Fast Fourier Transform (FFT) algorithm or radix-2 FFT algorithm, and it reduce the computational complexity from O(N²) to O(N log₂(N ). Mixed Radix 4-2 A mixed radix algorithm is a combination of different radix-r algorithms. That is, different stages in the FFT computation
Velammal College of Engineering and Technology, Madurai
In order to verify the proposed scheme, 64-points FFT based on the proposed Mixed-Radix 4-2 butterfly with simple bit reversing for ordering the output sequences is exampled. As shown in the Figure 2, the block diagram for 64-points FFT is composed of total six-teen Mixed-Radix 4-2 Butterflies. In the first stage, the 64 point input sequences are divided by the 8 groups which correspond to n3=0, n3=1, n3=2, n3=3, n3=4, n3=5, n3=6, n3=7 respectively. Each group is input sequence for each Mixed-Radix 4-2 Butterfly. After the input sequences pass the first Mixed-Radix 4-2 Butterfly stage, the order of output value is expressed with small number below each butterfly output line in the figure 3. The proposed Mixed-Radix 4-2 is composed of two radix-4 butterflies and four radix-2 butterflies. In the first stage, the input data of two radix-4 butterflies which are expressed with the equation B4 (o, n3, kj) B4 (i, n3, k1), are grouped with the x(n3),
Page 204
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
x(N/4±n3), x(N/2±n3), x(3N/4±n3) and x(N/ 8±n3), x(3N/8±n3), x(5N/8±n3), x(7N/8±n3) respectively. After the each input group data passes the first radix-4 butterflies, the outputted data is multiplied by the special twiddle factors. Then, these outputted sequences are inputted into the second stage which is composed of the radix-2 butterflies. After passing the second radix-2 butterflies, the outputted data are multiplied by the twiddle factors. These twiddle factors WQ (1+k) is the unique multiplier unit in the proposed MixedRadix 4-2 Butterfly with simple bit reversing the output sequences. Finally, we can also show order of the output sequences Fig above. The order of the output sequence is 0,4,2,6,1,5,3 and 7 which are exactly same at the simple binary bit reversing of the pure radix butterfly structure. Consequently, proposed mixed radix 4-2 butterfly with simple bit reversing output sequence include two radix 4 butterflies, four radix 2 butterflies, one multiplier unit and additional shift unit for special twiddle factors.
Fig.3 Proposed Mixed-Radix 4-2 Butterfly for 64 point FFT
III.
RESULT
Employing the parametric nature of this core, the OFDM block is synthesized on one of Xilinx’s Virtex-II Pro FPGAs with different configurations. The results of logic synthesis for 64 point FFT based orthogonal frequency division multiplexing (OFDM) using Radix-2, Radix-4, split Radix and mixed radix 4-2 are presented in Table 1. We analysis the 64-point FFT based OFDM is chosen to compare the number of CLB slices and Flip Flop for different FFT architectures .
Velammal College of Engineering and Technology, Madurai
Table 1: Comparison of FFT Algorithm based on CLB Slices, Function Generators.
IV.
DFF, and
Conclusion
In this paper, we design an OFDM for different algorithms implemented in OFDM modem are identified. It was found during the algorithm design that many blocks need complex multipliers and adders and therefore special attention needs to be given to optimize these circuits and maximize reusability. In particular, the models have been applied to analyze the performance of mixed-radix FFT architectures used in OFDM. Actual hardware resource requirements were also presented and simulation results were given for the synthesized design. The 64-point Mixed Radix FFT based OFDM architecture was found to have a good balance between its performance and its hardware requirements and is therefore suitable for use in OFDM systems. V. References [1] Shousheng. He and Mats Torkelson, "Design and Implementation of a 1024-point Pipeline FFT Processor", IEEE Custom Integrated Circuits Conference, May. 1998, pp. 131-134. [2] Shousheng He and Mats Torkelson, "Designing Pipeline FFT Processor for OFDM (de)Modulation", IEEE Signals, Systems, and Electronics, Sep. 1998, pp. 257-262. [3] Shousheng He and Mats Torkelson, "A New Approach to Pipeline FFT Processor", IEEE Parallel Processing Symposium, April. 1996, pp.776-780. [4] C. Sidney Burrus, "Index Mapping for Multidimensional Formulation of the DFT and Convolution", IEEE Trans. Acoust., Speech, and Signal Processing, Vol. ASSP-25, June. 1977, pp. 239-242. [5] Lihong Jia, Yonghong GAO, Jouni Isoaho, and Hannu Tenhunen, "A New VLSI-Oriented FFT Algorithm and Implementation", IEEE ASIC Conf., Sep. 1998, pp. 337-341.
Page 205
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[6] Martin Vetterli and Pierre Duhamel, "Split-Radix Algorithms for Length-ptmDFT's", IEEE Trans. Acoust, Speech, and Signal Processing, Vol. 37, No. 1, Jan. 1989, pp. 57-64. [7] Daisuke Takahashi, "An Extended Split-Radix FFT Algorithm", IEEE Signal Processing Letters, Vol. 8, No. 5, May. 2001, pp. 145-147.
Velammal College of Engineering and Technology, Madurai
[8]Y.T. Lin, P.Y. Tsai, and T.D. Chiueh, "Low-power variable-length fast Fourier transform processor", IEE Proc Comput. Digit. Tech, Vol. 152, No. 4, July. 2005, pp. 499506..
Page 206
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Towards Energy Efficient Protocols For Wireless Body Area Networks Shajahan Kutty#1, Laxminarayana J.A*2 #
Departmen of E&TC Engineering, Goa College of Engineering Farmagudi, Ponda, Goa , India 1
[email protected]
Department of Computer Engineering, Goa College of Engineering Farmagudi, Ponda, Goa , India [email protected]
Abstract The major medium access control (MAC) layer protocol design challenges for wireless body area network (WBAN) involve Quality of Service assurance for high reliability and guaranteed latency requirement for real time data, especially vital signs that needs a deterministic structure with special care for emergency reaction alarm; flexibility since it must support various types (periodic, non-periodic, medical, entertainment) of traffic/data rate and importantly, energy efficiency, since energy consumption especially for implanted device must be limited. These requirements necessitate design of efficient active/inactive scheduling. This paper evaluates the needs and performance of energy efficient MAC protocols proposed for wireless body area networks. Keywords- body sensor nodes; BAN co-ordinator node; beacon; preamble; superframe.
Introduction The major sources of energy wastage in a typical sensor network are due to consumptions occurring in three domains, viz; sensing, data processing and communication [1]. Of these, the energy communication wastage is a significant contributor and occurs due to idle listening as dominant factor in most applications, collision of packets, overhearing and control packet overhead. Mainly design of energy efficient MAC protocols target a reduction or elimination of the energy communication waste in particular idle listening. The central approach to reduce idle listening is through duty cycling. Suitable wake-up mechanisms can save significant amount of energy in Wireless Body Area Networks (WBAN) and increase the network lifetime. In these schemes, the device wake-ups only when necessary, otherwise it sleeps thereby saving energy. Coordinated and controlled data transmission can therefore reduce energy consumption. Data traffic in a WBAN is classified into [2]:
Velammal College of Engineering and Technology, Madurai
Normal traffic: Based on normal operation between device and coordinator using a defined pattern. On-demand traffic: Initiated by Coordinator to know certain information. Emergency traffic: In case of critical condition.
Fig (1) Data traffic in a WBAN The various MAC issues that WBAN has to contend with are [3]: Heterogeneous traffic: due to normal, on-demand and emergency traffic, Interoperability due to multiple frequency bands and correspondingly multiple PHY techniques and connecting different devices working on different bands/PHYs Scalability: due to variable data rate, kbps to Mbps and variable number of devices The options for implementing energy saving mechanisms involve choosing between synchronous or asynchronous transmission, beacon or preamble and different periods of wake up and sleep. The three main approaches adopted are: 1. Low power listening (LPL) 2. Scheduled contention 3. Time Division Multiple Access (TDMA)
Page 207
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Low power listening (LPL) involves channel polling where nodes wake up very briefly to check channel activity without receiving data. If the channel is idle, the node goes back to sleep, otherwise it stays awake to receive data. This action is performed regularly but not synchronised among nodes. To rendezvous with receivers, senders send a long preamble before each message to intersect with polling . Since the technique is asynchronous it is sensitive to tuning for neighbourhood size and traffic rate. The result is poor performance when traffic rates vary greatly and thus it is effectively optimised for known periodic traffic. As can be seen, the receiver and polling efficiency is gained at the much greater cost of senders. Scheduled contention involves schedule coordinated transmission and listen periods. Here the nodes adopt common schedule, synchronised with periodic control messages. The receiver listens to brief contention periods while senders contend. Only nodes participating in data transfer remain awake after contention periods while others can sleep. Since the method is synchronous it is sensitive to clock drift. However, improved performance with traffic increase results in similar cost incurred by sender and receiver. As seen the technique is scalable, adaptive and flexible. TDMA is a contention-free/cluster based method where the base station (BS)/cluster head (CH) allocates time slots to every node. Only one node is allowed to transmit in a slot. Timing and synchronisation provided by BS/CH. Normally this requires nodes to form clusters with one node as CH/BS Communication is only between nodes and cluster head; no peer to peer communication exists. The advantage is low duty cycle and fine grained time synchronisation. However the system is very sensitive to clock drift and results in limited throughput. Moreover, it requires clustering which means the cost incurred more on cluster head. Limited scalability and adaptivity to changes on number of nodes are the other issues with this scheme. A WBAN consists of low-power invasive and non-invasive Body Sensor Nodes (BSN). BSNs can be full functional, eg when dealing with multiple PHYs or reduced functional that are mostly common in the in-body networks [4]. One WBAN device is selected as a BAN Network Co-ordinator (BNC).
Traffic between BSNs and the BNC and On-Demand Traffic between the BNC with BSNs and BSNs with BSNs. The BNC can wake-up all the time and can support normal, emergency, and on-demand traffics. On the battery power supply side it has certain limitations and should adopt lowpower scavenging techniques. It should also calculate its own wakeup pattern based on the BSN’s wake-up patterns. This necessitates maintenance of the pattern-based wake-up table. The BSN devices operate on limited power and support a default normal wake-up state. They wake-up and sleep according to a pattern-based wake-up table. The BSN wakesup upon receiving an ‘on demand’ request from a BNC. The BSN also wakes-ups by itself to handle emergency events. .
Fig (2) Traffic characteristic in a WBAN MAC LAYER FRAME FORMAT A typical frame structure includes the advertisement which specifies the synch, interval, address, the CAP specifying reservation and non-periodic data, Beacon with synch, length of slot, reservation status announcement, DTP (Data Transmit Period), DTS (Data Transmit Slot) Continuous, Periodic data, ETS (Emergency Data Transmit Slot) Emergency data & Periodic data Advertisement
Velammal College of Engineering and Technology, Madurai
DTS
ETS ( CAP )
Batch Ack
( Optional ) Inactive Period Or CAP Extension (Optional )
CAP
The traffic characteristic in a WBAN varies from low to high with periodic or non-periodic data and vice versa. The dynamic natures of BSNs does not urge synchronized periodic wake-up periods. Communication flows include Normal Traffic between BSNs and the BNC, Emergency
Beacon
TDMA Data Transmit Period
Fig (3) A typical BAN superframe (BSF)
Page 208
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
BANs may be coordinated most of the time. The coordinator can allocate a negotiated bandwidth to QoS demanding nodes (data rate, BER, latency) with a beaconenabled mode including TDMA and Slotted-ALOHA (for UWB) with relaying. The inherent advantage of TDMA architecture for power saving over non-TDMA structure is that the coordinator goes into inactive mode if no activity is scheduled after CAP or DTP Sensor (continuous data). Once DTS reservation is done, the node stays inactive period except for beacon and its reserved transmit time. Sensors (routine data) wakeup only for report and transmit during CAP or CAP extension whichever available first. Sensors (Alarm) wakeup only for report and transmit during CAP, CAP extension or ETS whichever is available first. Some applications imply a lower channel load. A beacon-free mode is more efficient. NETWORK ARCHITECTURE THE PRIMARY REQUIREMENT FOR WBAN IS TO COPE WITH LOW-POWER EMISSION AND SEVERE NON LINE OF SIGHT (NLOS) CONDITIONS MAINTAINING HIGH RELIABILITY WITH HIGH LEVEL OF QOS FOR CRITICAL / VITAL TRAFFIC FLOWS [5]. A SUITABLE ARCHITECTURE IS THE MESH NETWORK CENTRALIZED ON THE GATEWAY. A FULL MESH
The Beacon-based MAC mode Superframe structure is based on 802.15.4 with an inclusion of a control portion for management messages. The control portion is large enough to allow dynamic changes of topology. The CAP is minimized, mostly used for association using slotted ALOHA Control portion
Beacon period
Request period
Topo mgmt period
Data portion
CAP
CFP
Inactive
Fig (5) Beacon-based MAC mode Superframe structure In the Fine structure the superframe is divided equally into slots. Use of minislots in the control portion provides flexibility and adaptation to different frame durations. Guaranteed Time MiniSlot (GTMS) is introduced in CFP
TOPOLOGY BASED ON A SCHEDULING TREE WITH GUARANTEED ACCESS FOR MANAGEMENT AND DATA MESSAGES (REAL TDMA) IS PROPOSED.
111111111 000000000 000000000 111111111 Preamble
Packet source
DATA
111111 0 1 0 1 000000 1111111111111111111111111111 0000000000000000000000000000 111111 0 1 0 1 000000 TCI
Node #1
Node #i
1 0 0 1
11 00 00 11
1111111 0000000 0000000 1111111
0 1 0 1 000 111 0 1 0 1 000 111 0000000000000000000000000000 1111111111111111111111111111 0 1 0 1 000 111 0 1 0 1 1 0
Node #N
Receive mode
Transmit mode
Fig (4) Mesh network and scheduling tree structures
Velammal College of Engineering and Technology, Madurai
Fig (6) Fine superframe structure In the Control portion structure beacon period, the beacons are relayed along the scheduling tree. The beacon-frame length is minimized and the beacon alignment procedure is used During the request period, a set of GTS dedicated to allocation demands. In the topology management period, Hello frames, for advanced link state procedure are introduced with a scheduling tree based update Beacon-free MAC mode is used in situations that might not require a beacon-enabled MAC protocol. Examples are a symmetric or asymmetric network, low communication rate and small packets, network set-up, coordinator disappearance, etc. In such conditions, the downlink is an issue : nodes must be powered-up to receive data from the coordinator (e.g. polling/by-invitation MAC scheme) A Preamble Sampling MAC protocol for Ultrawideband (UWB) involves nodes that periodically listen to the channel. If it is clear, they go back to sleep; conversely, they keep on listening until data. Nodes are not synchronized but share the wake up period (TCI) and a packet is transmitted with a preamble as long as TCI
Page 209
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
70 60
Fig (7) The Preamble Sampling Scheme Preamble sampling protocols are known to be the most energy efficient uncoordinated MAC protocols since TCI depends on the traffic load and the TRX consumption (TX/RX/listen modes). This can be of the order of 100 ms. The preamble is a network specific sequence, either a typical wake-up signal or a preamble with modulated fields (e.g. @, time left to data)
Rate (%)
50 40 30 20 10
Prioritized Backoff with Slotted Aloha Slotted Aloha
0 1
2
3
4
5
Number of Node per Class
PERFORMANCE EVALUATION A.. Beacon-free MAC mode : The simulation is carried out for three setups viz ; Mode when only one burst per device is emitted, Mode 1 when burst is re-emitted by a device, under the LDC limit, until a neighbor relays and Mode 2 which is same as Mode 1 plus former relays, with emission credits left, can relay again.
Fig(9) Average number of acknowledged packets Simulation Parameters: Simulation time: 60 sec, Inter-arrival time: 1pkt/sec (Poisson), Packet size: 16byte, BO=6, SO=3, CAP size = 8 slot, 1 to 5 node per each class DISCUSSION AND CONCLUSION The responsibilities of the upper layers are to enable/disable beacons switching procedure as follows: (i) from the beacon-free to beacon-based mode, BAN formation (coordinator election), new coordinator election if former coordinator leaves that, for example, can be triggered by a user requesting a link with high QoS (ii) from the beacon-based to beacon-free mode which is the fallback mode if the coordinator leaves or if the required BW (rate of GTS requests) is below a given threshold and requested QoS is adequate. A combination of beacon-based true TDMA mode including fast relaying and mesh support and a beacon-free mode using preamble sampling and extra features adapted to UWB is suitable for WBAN systems
Fig (8) PDP vs Wake up period plots for the three modes B. Access Policy in CAP for UWB PHY: Since CCA or carrier sense is not feasible with UWB PHY Aloha type of access is the only possible solution. While Aloha is a non-beacon mode, Slotted Aloha (Beacon mode) is better than Aloha but still the probability of collision is still high. Slotted Aloha with Prioritized Back-off is a new scheme.
Velammal College of Engineering and Technology, Madurai
References [1] J.S Yoon, Gahng S. Ahn, Myung J Lee, Seong-soon Joo “Versatile MAC for BAN proposal responding to TG6 Call for Proposals (15-08-0811-03-0006-tg6 call for proposals” [2] A. El-Hoiydi and J. D. Decotignie, “WiseMAC: An Ultra Low-Power MAC Protocol for Multi-hop Wireless Sensor Networks”, Proceeding of First International Workshop on Algorithmic Aspects of Wireless Sensor Networks (ALGOSENSOR), July 2004. [3] E. A. Lin, J. M. Rabaey, and A. Wolisz, “PowerEfficient Rendezvous Schemes for Dense Wireless Sensor Networks”, Proceedings of ICC, June 2004. TICER/RICER [4] The French ANR "BANET" project [5] M. J. Miller and N. H. Vaidya, “A MAC Protocol to Reduce Sensor Network Energy Consumption Using a Wakeup Radio”, IEEE Transactions on Mobile Computing, 4, 3, May/June 2005. [6] 15-08-0644-09-0006-tg6-technical-requirementsdocument.doc [7] Timmons, N.F. Scanlon, W.G. "Analysis of the performance of IEEE 802.15.4 for medical sensor body area
Page 210
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
networking", IEEE SECON 2004, Santa Clara, Ca, October 2004 [8] European ICT SMART-Net project deliverables [9] Jean Schwoerer , Laurent Ouvry , Arnaud Tonnerre, “PHY and MAC proposal based on UWB impulse radio for the IEEE 802.15.6 CFP-Response to IEEE 802.15.6 call for proposals” [10] J. Polastre, J. Hill and D. Culler, “Versatile Low-Power Media Access for Wireless Sensor Networks”, in Proc. of ACM Sensys, November 2004. B-MAC [11] E. Shih, P. Bahl, and M. J. Sinclair, “Wake On Wireless: An Event Driven Energy Saving Strategy for Battery Operated Devices”, Proceedings of ACM MobiCom, September 2002. [12] HyungSoo Lee, Jaehwan Kim, et al; “ Merged Baseline Proposal for TG6 -response to IEEE 802.15.6 call for proposals”
Velammal College of Engineering and Technology, Madurai
Page 211
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The Medical Image Segmentation R.Kalaivani M.phil (final year) Department of Computer Science Dr.N.G.P Arts and Science College Coimbatore ABSTRACT The Medical Image Segmentation is developed to extract the suspicious region from the patient image. The proposed system design aim is to give second opinion for the radiologist in tumor deduction. Image Acquisition is used to collect the image from a specified location or database, then the image data is converted to digital data or digital matrix. Image Preprocessing removes the labels and unwanted areas from the image using tracking algorithm. Image Enhancement covers the study related the removal of film artifacts and noises from the image using filters. Image Registration correlates the source and target image at same position and place. It also crop unwanted image from the original image by using the point mapping technique. Image Blocking blocks existing registered source and target images. Image Comparison compares the sources and average intensity of each and every block of the source and target image. Image segmentation the suspicious region is segmented from the target image if comparing the difference from source and target image. KEYWORDS Image acquisition, Image preprocessing Image enhancement, Image registration, Image blocking, Image comparison, Image segmentation, Performance evolution.
1. INTRODUCTION In the clinical context, medical image processing is generally equated to radiology or "clinical imaging" and the medical practitioner responsible for interpreting (and sometimes acquiring) the images are a radiologist. Diagnostic radiography designates the technical aspects of medical imaging and in particular the acquisition of medical images. The radiographer is usually responsible for acquiring medical images of diagnostic quality, although some radiological Medical imaging is often perceived to designate the set of techniques that no invasively produce images of the internal aspect of the body. In this restricted sense, medical imaging can be seen as the solution of mathematical inverse problems. This means that cause (the properties of living tissue) is inferred from effect (the observed signal). In the case of ultra sonography the probe consists of ultrasonic pressure waves and echoes inside the tissue show the internal structure. In the case of projection radiography, the probe is X-ray radiation
Velammal College of Engineering and Technology, Madurai
Guided by, Hemalatha, Lecturer Department of Computer Science Dr N.G.P Arts and Science College Coimbatore which is absorbed at different rates in different tissue types such as bone, muscle and fat. Medical Image diagnosis is considered as one of the fields taking advantage of high-technology and modern instrumentation. Ultra sound, CT, MRI, PET etc., have played an important role in diagnosis, research, and treatment. Image registration is a process of overlaying two or more images of the same scene taken at different times, from different viewpoints, and/or by different sensors. The present differences between images are introduced due to different imaging conditions. 2. METHOD Assuming that we have two images of the same Object, a structural image and a functional image. The Process of registration is composed of following steps: • Collect the image from a specified location or database, then the image data is converted to digital data or digital matrix. • Removes the labels and unwanted areas from the image using tracking algorithm. • Removal of film artifacts and noises from the image using filters. • Image Registration correlates the source and target image at same position and place. • Crop unwanted image from the original image by using the point mapping technique. • Blocks existing registered source and target images. • Compares the sources and average intensity of each and every block of the source and target image • The suspicious region is segmented from the target image if comparing the difference from source and target image. 2.1 Image Acquisition Image Capturing is a process to acquire the digital image into the Mat lab. In this module medical image was given as input and it is the base input to the whole project. All types of medical images can be acquired in this module. Images of a patient obtained
Page 212
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
are displayed as an array of pixels (a 2D unit based on the matrix size and the field of view) and stored in memory. In mat lab there are various formats of image encoding Binary Images Large matrix whose entries are one of the two values will specify a black and white image.0 corresponds to black, 1 to white. Grey Scale/Intensity Images A grey scale image can be specified by giving a large matrix whose entries are numbers between 0 and 1.A black and white image can also be specified by giving a large matrix with integer entries. The lowest entry corresponds to black, the highest to white. Fig shows typical Image Captured in Mat lab from patient database
For removing the unwanted portions of the image we have developed an algorithm named Tracking Algorithm. Fig shows a typical preprocessed image after the removal of film artifacts
2.3 Image Enhancement Image enhancement methods inquire about how to improve the visual appearance of images from Magnetic Resonance Image (MRI), Computed Tomography (CT) scan; Positron Emission Tomography (PET) and the contrast enhancing brain volumes were linearly aligned. Fig : Shows the Median Filter output images
2.2 Image Pre-processing Preprocessing indicates that the same tissue type may have a different scale of signal intensities for different images. Processing functions involve those operations that are normally required prior to the main data analysis and Extraction of information, and are generally grouped as geometric corrections. Radiometric corrections include correcting the data for sensor irregularities and unwanted sensor or atmospheric noise, removal of non-brain voxels and converting the data so they accurately represent the reflected or emitted radiation measured by the sensor. Tracking Algorithm In preprocessing module image acquired will be processed for correct output. Medical images surely will have some Film Artifacts like labels and marks which are detected and removed for better result.Pre-processing was done by using tracking algorithm. If these images are too noisy or blurred, they should be filtered and sharpened.
Velammal College of Engineering and Technology, Madurai
(a).3*3filtered image (b).5*5 filtered image
(c). 11*11 filtered image Fig : shows the enhanced output
Page 213
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.4 Image Registration Image registration is the process of transforming the different sets of data into one coordinate system. Registration is necessary in order to be able to compare or integrate the data obtained from different measurements. In medical imaging (e.g. for data of the same patient taken at different points in time or by different angles) registration often additionally involves elastic (or non rigid) registration to cope with elastic deformations of the body parts imaged. Non rigid registration of medical images can also be used to register a patient's data to an anatomical atlas, such as the Talairach atlas for neuro-imaging. Algorithm Classifications •
Area-based vs. Feature-based.
•
Transformation model.
•
Search-based vs. direct methods.
•
Image nature.
•
Other classifications.
2.5.1 Block Based Method In Block based method, both the original image and the sensed image has been divided into several blocks. Then block wise subtraction has been done between the two images. This subtracted value is then checked with the threshold value, Block Based method Image. Fig shows the original image in blocks
Fig : shows the tumor image in blocks
Fig shows the image before Correlation 2.5 Image Blocking(Correlation) Here, we have taken two reference points first, in front view and second in the top view of the image. The enhanced image has to be resized to the original image size by fixing the same reference points as in the original image. Since in our technique, the size of the original image is 256*256, the enhanced image has been resized to (256-x)*(256-y) by removing the extra portions in the image. Fig shows the Correlated image
Velammal College of Engineering and Technology, Madurai
3. And
Conclusions future work
Relevance of these techniques is the direct clinical application for registration. This provides a diagnosis and preoperative planning surgery. Alternatively, it is a useful tool in image processing training. Algorithms of finding a transformation between two images is of importance in the methodology, spatial and density resolution have to be achieved to reduce geometric errors In this survey paper various automatic detection methods of brain tumor through MRI has been studied and compared for the period of one year. This is used to focus on the future of developments of medical image processing in medicine and health care. We have described several methods in medical
Page 214
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
image processing and discussed requirements and properties of techniques in brain tumor detection. This paper are used to give more information about brain tumor detection and registration. In this paper various step in detection of automatic system (a) Image capturing (b) Image preprocessing (c) Image Enhancement (d) Image Registration. The program needs to be compiled to run in an independent environment and algorithms need to be optimized inorder to reduce the time consuming and performing and increase in accuracy. The program could be developed in ability of access to host computer or hospital via intranet to get patients information and images. 4. References 1. Amini L, Soltanian-Zadeh H, Lucas’s,:“Automated Segmentation of Brain Structure from MRI”, Proc. Intl.Soc.Mag.Reson.Med.11 (2003). 2. BrainTumors: Gd-DTPA-Enhanced, Emodynamic, and Diffusion MR Imaging Compared with MRSpectroscopic Imaging”, Neuroradiol 23, February 2002. 3. Azadeh yazdan-shahmorad, Hesamoddin Jahanian, Suresh Patel, Hammed Soltanian- Zadeh,:”Automatic Brain Tumor Segmentation .Using Diffusivity Characteristics”, IEEE, ISBI,2007. 4. Boada F.E, Davis.D, Walter.K, Torres- Trejo.A, Kondziolka.D, Bartynski.W, Lieberman. “Triple Quantum Filtered Sodium MRI of Primary Brain Tumors, “IEEE, USA, 2004. 5. Dimitris N Metaxas, Zhen Qian, Xiaolei Huang and Rui Huang,Ting Chameleon Axal.:”Hybrid Deformable Models for Medical Segmentation and Registration, “IEEE Transactions on Medical Image processing, ICARCV ,2006. 6 Gilbert Vezina:”MR Imaging of Brain Tumors- Recent Developments”, Childhood Brain Tumor Foundation, Germantown, Washington. Evaluating Image Segmentation Algorithm”, IEEE, Medical Image Processing. 7 Hideki yamamoto and Katsuhiko Sugita, Noriki Kanzaki, Ikuo Johan and Yoshio Hiraki,Michiyoshi Kuwahara , : “Magnetic Resonance Image Enhancement Using VFilter”, IEEE AES Magazine, June 1990.
Velammal College of Engineering and Technology, Madurai
Page 215
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Selection of a Checkpoint Interval in Coordinated Checkpointing Protocol for Fault Tolerant Open MPI #
Mallikarjuna Shastry P.M.#1, K. Venkatesh #2
Research Center, Department of Computer Science & Engg., M. S. Ramaiah Institute of Technology, M.S.R Nagar, Bangalore-54, Karnataka, India. 1
[email protected] 2 [email protected]
Abstract — The goal of this paper is to address the selection of efficient checkpoint interval which reduces the total overhead cost due to the checkpointing and restarting of the applications in a distributed system environment. Coordinated checkpointing rollback recovery protocol is used for making the application programs fault tolerant on a stand-alone system under no load conditions using BLCR and OPEN MPI at system level. We have presented an experimental study in which we have determined the optimum checkpoint interval and we have used it to compare the performance of coordinated checkpointing protocol using two types of checkpointing intervals namely fixed and incremental checkpoint intervals. We measured the checkpoint cost, rollback cost and total cost of overheads caused by the above two methods of checkpointing intervals Failures are simulated using the Poisson distribution with one failure per hour and the inter arrival time between the failures follow exponential distribution. We have observed from the results that, rollback overhead and total cost of overheads due to checkpointing the application are very high in incremental checkpoint interval method than in fixed checkpoint interval method. Hence, we conclude that fixed checkpointing interval method is more efficient as it reduces the rollback overhead and also total cost of overheads considerably.
Keywords: Checkpoint, Checkpoint Interval, Fault tolerance, Marker, Checkpoint Overheads.
I. INTRODUCTION Since, the recent trends in HPC and even stand alone systems employ a very large number of processors to execute the large size application programs in a distributed system environment, it is required to provide the fault tolerance to such applications.
Velammal College of Engineering and Technology, Madurai
As the complexity of the program increases, the number of processors to be added to the cluster / HPC / Super Computer also increases which in turn decreases the MTBF (mean time between failures) of the processors or the machines. It means that the probability of failure of one or more processors will be very high before the completion of the execution of the long running application being executed parallely on several processors. When a processor fails, we need to restart the entire application on all the processors from the beginning. Hence, it is required to address the issues like the scalability and fault tolerance. Fault tolerance provides the reliability and availability to the large size applications programs executed in a distributed system environment. Fault tolerance is achieved using coordinated checkpointing rollback recovery protocol in which an initiator takes a checkpoint by synchronizing with all other processes of MPI application [1]. For MPI applications, a cluster consisting of a group of processes interacting with each other is formed and each individual process in the cluster is checkpointed and a global state is formed out of it. The global state contains the “set of checkpoints exactly one from each processor”. The global state is consistent if and only if for each message received by a processor (receiver), there is a corresponding sender. The latest consistent global state is known as the recovery line [2]. The checkpoint / restart scheme has been widely used in [3]-[9] to address the failure of execution of an application. Checkpoints can be taken using either fixed checkpoint interval or variable checkpoint interval [10]. In case of
Page 216
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
fixed checkpoint interval, checkpoint interval size remains same between any two successive checkpoints. But, in case of the incremental checkpoint interval method discussed in this paper, the second checkpoint interval size is 2 times the 1st one and third checkpoint interval is 3 times the 1st one and so on in each cycle A cycle is the execution time interval of the application between two successive failures. Since, these two methods of checkpoint intervals are not compared in the literature as we understand; we have carried out an experiment to determine the behaviour of the coordinated checkpointing protocol using the fixed and incremental checkpoint interval methods. The rest of the paper is organized as follows. Section 2 introduces the related works carried out in checkpoint and restart schemes using different checkpoint intervals. Section 3 presents the different notations used in this paper. The implementation of the coordinated checkpointing protocol, determination of optimum checkpoint interval, complete description of fixed and incremental checkpoint intervals are discussed in section 4. Computation of cost of overheads is discussed in section 5. Section 6 presents the experimental setup and the results. Section 7 presents the conclusion.
3. R - restart time required to resume the execution of an application from the most recent checkpoint. 4. F - the number of failures during the execution of the application. 5. TS - time required to save each checkpoint on to a local disk. 6. Ni - the number of checkpoints taken in ith cycle. 7. tc - checkpoint interval without restart cost. 8. TC - optimum checkpoint interval size and is used as fixed checkpoint interval. 9. TCi– ith checkpoint interval which is incremental. 10. CCi - the cost of checkpoints in ith cycle. 11. CC – total cost of checkpoints 12. P - the number of processes / processors used for parallelism. 13. λ – Number of failures per hour. 14. TF – time to failure. 15. Ti - the time at which the ith failure occurs. IV. IMPLEMENTATION OF COORDINATED CHECKPOINTING PROTOCOL A. Protocol
II. RELATED WORKS
Master MPI process with rank i=0 takes the tentative checkpoint and then sends the marker to MPI process with rank (i+1) % N. When MPI process i > 0 receives the marker from (i + N-1) % N, takes its tentative checkpoint and sends the marker to MPI process with rank (i + 1) % N.
Young [5] has presented an optimum checkpoint and restart model and has shown that the total waste time due to checkpointing can be reduced using fixed checkpoint interval. But, this model [5] does not consider the restart time required to resume the application from the most recent checkpoint after a failure.
When the MPI process with rank 0 receives the marker from MPI process N-1, a global consistent checkpoint is formed out of all the local checkpoints and then sends the checkpoint file to the local disk and then initiates the next checkpoint cycle after the checkpoint interval as specified by the user.
An optimal checkpoint / restart model presented by Yudun liu [11] uses varying checkpoint interval with different failure distributions. But, varying checkpoint interval does not yield optimal rollback and checkpoint cost. R. Geist et. Al [12] discusses the selection of checkpoint interval in a critical task environment, but it does not present any optimal solution for selecting the checkpoint interval.
The checkpoint period can be either fixed or incremental which will be discussed in the following subsections.
J.T. Daly [6], [9] presents a method for determining the optimum checkpoint interval but they do not discuss the comparison of the performance of the coordinated checkpointing protocol with respect to fixed and incremental checkpointing interval methods. III. NOTATIONS USED 1. 2.
Rbi - the cost of rollback in ith cycle. Rb - total rollback cost.
Velammal College of Engineering and Technology, Madurai
B. Computing Optimal Checkpoint Interval Knowing the number of processors (P) used for computation and the failure rate (λ) of the processors, we can compute the time to failure [14] of the application during run time as follows.
TF = 1 / (P λ )
(1 )
Once, we determine the time to failure TF and the checkpoint overhead TS (time required to save each checkpoint onto local disk), the checkpoint interval tC can be computed [ 5] as follows.
Page 217
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
tC =
2 TS TF
(2)
i
Ci = ∑ TCK + (i − 1) TS
(5)
k =1
But, the above equation does not consider the restart time of the application after a failure. Hence, considering the restart time (R) also, the above equation can be modified by adding the restart time (R) multiplied by the number of failures (F) to tC as follows to get the optimal checkpoint interval TC.
TC =
2 TS T F + R F
(3)
And the ith checkpoint interval is computed as follows.
TC i = i TC
( 6)
C1
C2 TC1
TS
CNi TC2
TS
TS
Fig. 2 Incremental Checkpoint Interval
C. Fixed Checkpoint Interval The total execution time of the application program is divided into n checkpoint intervals of length TC as computed in the previous section. The first checkpoint is initiated by the master MPI process after completion of Tc minutes of execution of the application program and second checkpoint is initiated after completion of 2 TC + TS minutes and so on as shown in figure 1. In general, the starting time of computed as follows.
ith checkpoint Ci is
Ci = i TC +(i −1) TS
(4)
V. COST OF OVERHEADS DUE TO CHECKPOINTING AND RESTARTING During recovery from failure, the master MPI process coordinates with all the other MPI processes and restarts by rolling back the application to the most recent consistent global checkpoint. The cost of rollback, cost of checkpointing and the restart cost are the 3 components which are used to determine the waste time in each cycle of the application. If the application undergoes F failures, the execution of the application will have F cycles. We present in the following sections the determination of these costs using two different methods of checkpoint intervals such as fixed and incremental checkpoint intervals.
But, the length of each checkpoint interval is fixed.
A. Cost of Overheads in Fixed Checkpoint Interval
i.e., TC1 = TC, TC2 = Tc, TC3 = Tc and so on.
Failure of a fault tolerant application using fixed checkpoint interval is shown in figure 3. Since, all the checkpoint intervals have same length; the number of checkpoints to be taken in ith cycle (before ith failure occurs) is computed as follows.
In general, TCN = TC C1 TC
TS
TC
C2
CNi
TS
TS
TC
Fig. 1 Fixed Checkpoint Interval
D. Incremental Checkpoint Interval Figure 2 shows the checkpointing of the application with an incremental checkpoint interval. In this case, in each cycle, the first checkpoint is initiated after (Tc1) or TC minutes of execution of the application and second checkpoint is initiated after TC1 + TC2 + TS minutes and the third checkpoint is initiated after TC1 + TC2 + TC3 + 2TS and so on as shown in figure 2. In general the starting time of ith checkpoint Ci is computed as follows.
Velammal College of Engineering and Technology, Madurai
N i = ⎣Ti / (TC + TS )⎦
(7 )
Then, the cost of checkpoint in ith cycle is computed as follows.
CC i = N i TS
(8)
The cost of rollback in ith cycle is then computed as follows
Rbi = (Ti − N i (TC + TS ))
(9)
The time lost in ith cycle TLi due to a failure can be obtained by adding checkpoint cost, rollback cost, restart cost together as follows.
Page 218
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TLi = CC i + Rbi + R
(10)
The total waste time due to F failures is then computed as follows. F
TL = ∑ TLi
(11)
i =1
Suppose, if the first failure occurs at 40 minutes of execution and checkpoint interval size TC is 4 minutes, time to save a checkpoint is 35 seconds and number of checkpoints taken before failure occurs is 8, The different overhead costs are determined as follows. i) Rollback cost: By applying equation (8), we can determine the rollback cost as follows.
ii) Checkpoint cost = 8 checkpoints * cost of each checkpoint = 6 * 35 = 280 seconds iii) It was found from the experimental setup that the time required to restart an application after a failure is just about 24 seconds. So, the total time lost due to a failure of application in fixed checkpoint interval case is TL1=cost of rollback (200 seconds) + cost of checkpoints (280 seconds) + restart cost (24 seconds) = 504 seconds (about 21 % of execution time of 1st cycle is wasted due to checkpointing, rollback and restart). It was observed that, the cost of rollback is dependent on the amount of time elapsed since the last checkpoint. Time required to restart
Restart
Application Failure C2 TC
TS
CNi TC
TS
a. .if T1 < (TC1 + TS ), b . .if (
n
∑K k =1
(12)
T C + n T S ) < T i and
Ti < ( for
N1 = 0
n +1
∑K k =1
n = 1, 2 , 3 , ..
T C + ( n + 1) T s ) and
for
(13 )
i = 1, 2 , 3 , .. F
The number of checkpoints to be taken in ith cycle can be computed as follows.
Rbi = (2400 – 8 * (240 +35)) = 200 seconds
C1
Failure of a fault tolerant application using incremental checkpoint interval is shown in figure 4. In this, method, size of first checkpoint interval is TC minutes and size of other checkpoint intervals is TC minutes more than its previous checkpoint interval in each cycle. Hence, the number of checkpoints (Ni) to be taken in ith cycle will vary in this method and it is computed as follows.
TS
TC
Execution time till a failure (Ti ) Cycle i Fig. 3 Cost of Overheads in Fixed Checkpoint Interval
B. Cost of Overheads in Incremental Checkpoint Interval
Velammal College of Engineering and Technology, Madurai
Ni = n
(14)
Though, the checkpoint interval length keeps increasing, from one checkpoint to another checkpoint, the per checkpoint cost remains same as per the experimental results that we have obtained. So, the total checkpoint cost in ith cycle can be computed as follows.
CCi = Ni TS
(15)
The cost of rollback in ith cycle is then computed as follows. n
Rbi = (Ti − ( N i TS + ∑ K TC ))
(16)
k =1
The time lost in ith cycle TLi due to a failure can be obtained by adding checkpoint cost, rollback cost, and restart cost together using equation (9). Total waste time due to F failures is computed using equation (10). It was observed that, the cost of rollback depends on two factors like the time of failure and the checkpoint interval size. This is because, in this method, the checkpoint interval size varies from one checkpoint to another checkpoint. Suppose, if the first failure occurs at 40 minutes of execution and initial checkpoint interval TC is 4 minutes,2nd checkpoint interval is 8 minutes and 3rd checkpoint interval is 12 minutes and the 4th checkpoint interval is 16 minutes (during which the failure occurs), time to save a checkpoint is 35 seconds and number of checkpoints taken before failure occurs is 3 after applying the equation (12) and (13). Then, the different overhead costs are determined as follows.
Page 219
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
i ) rollback cost : after applying the equation (15), we get Rbi =(2400 – (105 + 1440 )) = 855 seconds ii) checkpoint cost is = 3 checkpoints * cost of each checkpoint=3*35 = 105 seconds iii) and it was found from the experimental setup that the time required to restart an application after a failure is just about 24 seconds. So, the total time lost due to a failure of application in incremental checkpoint interval case is TL1 =cost of rollback + cost of checkpoints + restart cost = 984 seconds (41 % of execution time of 1st cycle is wasted due to checkpointing, rollback and restart). Time required to restart Restart Application Failure C1 TC1
TS
C2 TC2
During the restart or recovery state, the MPI application rolls back to the most recent checkpoint as discussed and resumes the execution of the application from that point. BLCR checkpoint and restart library [13] is used to implement the blocking coordinated checkpointing protocol to checkpoint the application. The application was run 10 times for different number of processors varying from 1 to 10. We observed that the checkpoint cost and the restart cost increase linearly with the increase in the number of processors. The results obtained for 10 processors are presented in this paper. The number of arrivals N (t) in a finite interval of length t obeys the Poisson (גt) distribution.
P { N (t ) = n} = (λ t ) n e − λt / n ! The inter arrival times of the failures are independent and obey the exponential distribution.
f ( x )
= λ e = 0,
− λ x
for
x >=
0
otherwise
CNi
TS
TS
TCNi
Execution time till a failure (Ti) Cycle i Fig. 4 Cost of Overheads in Incremental Checkpoint Interval
VI. EXPERIMENTAL SET-UP AND RESULTS We have taken the application program that multiplies 2 integer matrices of size 7000 * 7000. The above application is written in C language and run on a standalone system using scattered method of MPI under no load conditions. In scattered method, one of the matrices, say the matrix B is broadcasted across all the processors using MPI_Bcast(). The matrix A is divided equally among the number of processors used for parallelism and each of the processors gets only a portion of matrix A allocated for it for the computation using MPI_Scatter(). The above application was run on a system with 6 GB of RAM, Intel ® Core ™ 2 Duo CPU,E7200 @ 2.53 GHz and 110 GB of HDD and the execution time of the application considered in our experimental setup on this system is 67 minutes without checkpointing. The monitor program is written in a shell script which runs at the background and keeps monitoring whether the MPI processes grouped under mpirun are running or not. Once, monitor program learns that an MPI process has failed, it calls the restart() routine of BLCR to restart the application.
Velammal College of Engineering and Technology, Madurai
We have used Poisson distribution with 1 failure per hour (λ = 1) for 10 processors and generated the probability distributions for the inter arrival times of failures. Failures are simulated using these probability distributions and the results are presented in the form of graphs. Figures 5a and 5b present the comparison of total cost of all the overheads due to checkpointing and restarting of the application with different checkpoint interval sizes when the application considered in our experimental setup fails at different timings. We have tested for 10 different failures (whose failure timings are generated using Poisson arrival process) and determined the total cost of overheads incurred due to checkpointing and restarting using 5 different checkpoint intervals of size 2,3,4,5 and 10 minutes as shown in figures 5a and 5b. In 7 different cases out of 10 failures at different timings, checkpoint interval size with 4 minutes was found to be optimal as it yields minimum total cost of all the overheads as shown in figures 5a and 5b when the application fails at 10,15,20,42,47,50 and 60 minutes of execution of the application. From our experiment, we determined that, per checkpoint cost TS is 35 seconds for 10 processors and it remains same for all the other checkpoints taken at different timings for the same number of processors. Per checkpoint cost is determined by taking the average value of checkpointing the application at 20 different timings on 10 processors. Restart cost is found to be only 24 seconds to resume the execution of the application on 10 processors after a failure occurs. This restart cost is determined by taking the average of 20 different restart costs measured when the application failed at different timings.
Page 220
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Restart cost R is added only once because, we have considered only one failure i.e F = 1 during the execution of our application. This checkpoint interval value is almost matching with the optimum checkpoint interval of 4 minutes obtained from the figures 5a and 5b based on our experimental results. The value of TC calculated should yield almost the exact result, if the value of TF is quite large in which case R << TF. Hence, the equations (1), (2) and (3) are validated based on our experimental results and discussion. As, we have obtained 4 minutes as the optimum checkpoint interval size from our experimental analysis, in our further analysis and discussion (figures 6 to 9), fixed checkpoint interval size taken is 4 minutes and in incremental checkpoint interval method, the first checkpoint interval size taken is 4 minutes, second checkpoint interval size taken is 8 minutes and third checkpoint interval size is 12 minutes and so on. We have presented the results in the form of graphs for one failure in an hour with λ = 1 using Poisson distribution for arrival of failures. Figure 6 presents the comparison of number of checkpoints taken in fixed and incremental checkpoint interval methods. Figures 7, 8, and 9 present the comparison of i ) cost of checkpoints, ii ) cost of rollback and iii) total cost of overheads caused by fixed and incremental checkpoint interval methods respectively.
2 Min.
34
4 Min.
5 Min.
Total Cost (Sec)
500 400 300 200
10 Min.
42 47 50 Failure Time (Min)
60
Incremental
Fixed 14 12 10 8 6 4 2 0
5 10 15 20 25 30 35 40 45 50 55 60 Failure Time (Min) Fig 6. Comparison of Number of Checkpoints of Fixed and Incremental Checkpoint Intervals
Comparison of Checkpoint Cost Fixed
Incremental
500 400 300 200 100 0 5
100
5 Min.
Comparison of Number of Checkpoints
10 Min.
600
4 Min.
Fig 5b. Comparison of Total Cost of Overheads with Different Checkpoint Interval Size.
Checkpoint Cost (Sec)
3 Min.
3 Min.
900 800 700 600 500 400 300 200 100 0
Comparison of Total Cost of Overheads with Different Checkpoint Interval Size 2 Min. 700
Different
Comparison of total cost of overheads with different Checkpoint Interval Size
Total Cost (Sec)
TF = 1 / (P * λ) = 1 / ( 10 * (1/60)) minutes = 6 minutes and TC = sqrt (2 TS TF ) + R F = sqrt( 2 * (35 /60) * 6) + (24 seconds * 1) = 3.2 minutes
Fig 5a. Comparison of Total Cost of Overheads with Checkpoint Interval Size.
Number of Checkpoints
We have used the equations (1), (2) and (3) to determine the optimum checkpoint interval size when the time to failure TF and checkpoint cost TS and the restart cost R are known. The value of TC obtained from equation (3) shows that the optimum checkpoint interval size is 3.2 minutes as shown below.
10 15 20
25 30 35 40 45 50 55 60
Failure Time ( Min )
0 10
12
15
20
30 Fig 7. Comparison of Checkpoint Cost of Fixed and Incremental Checkpoint Intervals
Velammal College of Engineering and Technology, Madurai
Page 221
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Comparison of Rollback Cost Fixed
Incremental
Rollback Cost ( Sec)
1200 1000 800 600 400 200 0 5
10 15 20 25 30 35 40 45 50 55 60 Failure time (Min)
Fig 8. Comparison of Rollback Cost of Fixed and Incremental Checkpoint Intervals
Comparison of Total Cost 1400
Fixed
Incremental
Total Cost (Sec)
1200 1000 800 600 400 200 0 5 10 15 20 25 30 35 40 45 50 55 60 Failure Time (Min)
Fig 9. Comparison of Total Cost of overheads caused by Fixed and Incremental Checkpoint Intervals
VII. CONCLUSIONS Figures 5a and 5b show the comparison of total cost of overheads with different checkpoint interval size. From figures 5a and 5b, it is clear that the total cost of overheads is quite minimum when checkpoint interval size is 4 minutes. We have even validated the model developed by Young [5] to determine the optimum checkpoint interval. But, the model [5] does not consider the restart cost required to restart the application when it fails. We have added the restart cost multiplied by the number of times the application undergoes failures to Young’s model [5] to get the optimum checkpoint interval which yields the optimum overheads cost irrespective of number of failures for any application. An approximate estimate of the checkpoint interval can be calculated from equation (3). From figure 6, we see that the fixed checkpoint interval method causes more number of checkpoints than
Velammal College of Engineering and Technology, Madurai
incremental checkpoint interval method. So, the checkpoint cost is also quite high in fixed checkpoint interval method as compared to the incremental checkpoint interval method when the application fails after first checkpoint as shown in figure 7. But, the rollback cost and the total cost of overheads produced by fixed checkpoint interval are quite low as compared to the incremental checkpoint interval when the application fails after first checkpoint as shown in figure 8 and figure 9 respectively. Fixed checkpoint interval reduces more than 50% of total overhead cost as compared to the incremental checkpoint interval. Hence, we conclude that using fixed checkpoint interval for checkpointing an application would be more advantageous than using incremental checkpoint interval because fixed checkpoint interval reduces both rollback cost and the total cost of overheads significantly. ACKNOWLEDGEMENT The authors thank the Head of Research Centre, CSE dept. and head of ISE dept M.S Ramaiah Institute of Technology, Bangalore for their constant encouragement. The Computer system on which the experimental analysis has been carried out is acquired under the project sanctioned by BRNS, INDIA, bearing sanction No. 2008/37/15/BRNS. REFERENCES [1] Luis Moura Silva and Joao Gabriel Silva, “The Performance of Coordinated and Independent Checkpointing”, IEEE Trans, 1999. [2] G.E. Fagg, A. Bukovsky and J.J. Dongarra, “Harness and Fault Tolerant MPI”, Parallel Computing, 27(11):1479-1495, 2001. [3] K.M. Chandy, “A survey of analytic models of roll-back and recovery strategies,” Computer 8, 5 (May 1975), 40-47. [4] K.M. Chandy, J.C. Browne, C. W. Dissly, and W. R. Uhrig, “Analytic models for rollback and recovery stratagems in data base systems,” IEEE Trans Software Engg. SE-1, ( March 1975), 100-110 [5] J.W. Young, “A first order approximation to the optimum checkpoint interval,” Communications of ACM 17, 9(Sept 1974), 530531. [6] J.T.Daly, “A Higher Order Estimate of the Optimum Checkpoint Interval for Restart Dumps,” Future Generation Computer Systems [Elsevier ], Amsterdam, 2004. [7] E. Elnozahy, J. Plank, “Checkpointing for Peta Scale Systems: A Look into the Future of Practical Rollback-Recovery,” IEEE Trans. Dependable Sec. Comput.1 (2):97-108(2004). [8] M.Treaster, “A survey of fault-tolerance and fault-recovery techniques in parallel systems, “Technical Report cs.DC / 0501002, ACM computing Research Repository (CoRR), January 2005. [9] J. T. Daly, “A Model for Predicting the Optimum Checkpoint Interval for Restart Dumps,” ICCS 2003, LNCS 2660, Proceedings 4 (2003) 3-12. [10] Yudun Liu, Raja Nassar, Chokchai (box) Leangsuksum, Nichamon Naksinehaboon, Mihaels Paun, Stephen L. Scott, “An Optimal Checkpoint /Restart Model for a Large Scale High Performance Computing System,” IEEE Trans. 2008. [11] Y. Liu, “Reliability Aware Optimal Checkpoint / Restart Model in High Performance Computing, PhD Thesis,” Louisiana Tech university, Ruston, LA, USA, May-2007.
Page 222
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[12] R. Geist, R. Reynolds, and J. Westall,” Selection of a checkpoint interval in a critical-task environment,” IEEE Trans. Reliability, 37, (4), 395-400 (1988). [13] H. Paul Hargrove and C. Jason Duell, “Berkeley lab checkpoint / restart (BLCR) for Linux clusters”, Journal of Physics, Conference series 46 (2006), 494-499, SciDAC 2006. [14] James S. Plank and MichG.Thomason, “The Average Availability of Parallel Checkpointing Systems and Its Importance in th Selecting Runtime Parameters”,29 Internatioonal symposium on Fault Tolerant Computing , Madison WI, Jun-1999, pg 250-259.
Velammal College of Engineering and Technology, Madurai
Page 223
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Simplification Of Diagnosing Disease Through Microscopic Images Of Blood Cells T.Benazir Fathima#1 #1,2
K.V.Gayathiri Devi #2
M.Arunachalam*3
M.K.Hema*4
Final year Computer Science Engineering , K.L.N.College of Information Technology. [email protected],[email protected]
*3,4
Computer Science Engineering Department, K.L.N.College of Information Technology Pottapalayam,Sivagangai dt,TamilNadu,India [email protected],[email protected]
Abstract— This paper is the implementation of a simple, fast and reliable method for automatically diagnosing diseases through digitized images of blood. Hematopathology reveals that there is an intrusion of the disease cell in the blood having identical characteristics for each disease. Principal effectors in the blood that are Erythrocyte, Leukocyte and Platelets of the blood play a crucial role in supplying blood to the various parts of the body, resisting foreign particles and in clotting of blood respectively. The effectors are identified and skipped for our observation. In microscopic images of the diseased blood cells, the diagnosis is based on the evaluation of some general features of the blood cells such as color, shape, and border, and the presence and the aspect of characteristic structures. Perception of these structures depends both on magnification and image resolution. The introduction of vision system associated with image processing techniques enables the numerical and objective description of some pattern cells features. Matlab image processing software is used for the experimental work. A new algorithm is introduced which will efficiently reduce the overall running time of object recognition. Keywords - Medical imaging,BPN,Pattern matching
I.INTRODUCTION This project is the implementation of a simple, fast and reliable method for automatically diagnosing diseases through digitized images of blood. Hematopathology reveals that there is an intrusion of the disease cell in the blood having identical characteristics for each disease. Principal effectors in the blood that are Erythrocyte, Leukocyte and Platelets of the blood play a crucial role in supplying blood to the various parts of the body, resisting foreign particles and in clotting of blood respectively. The effectors are identified and skipped for our observation. In microscopic images of the diseased blood cells, the diagnosis is based on the evaluation of some general features of the blood cells such as color, shape, and border, and the presence and the aspect of characteristic structures. Perception of these structures depends both on magnification and image resolution. The introduction of
Velammal College of Engineering and Technology, Madurai
vision system associated with image processing techniques enables the numerical and objective description of some pattern cells features. Matlab image processing software is used for the experimental work. A new algorithm is introduced which will efficiently reduce the overall running time of object recognition. Concept of Back Propagation network is used to set appropriate error values during object recognition. Experimental results using proposed technique gives a way for new research area in Bio mechanics. Nevertheless, this would be an interesting future investigation, since our approach is sufficiently general to be applied to modeling tasks from other two dimensional or three dimensional application domains. The extension of this may open a new area of research in bio mechanics (modeling and simulation). II.IMPLEMENTATION A.EXPERIMENTAL SETUP The technology used in solid state imaging sensors is based principally on charged coupled devices (CCD), which can convert the images from analog to digital format. Simple strategies can be adapted to grab the images from the Electron Microscope. The experiment apparatus consists of structured light source, Electron Microscope, CCD camera and vision system. The relative cells of the blood are first measured. Before the actual diagnosis is made, the real coordinates and world coordinates must be determined in the camera coordinate system by a calibration process. The camera acquisition requires the knowledge of calibration analysis. The camera can be viewed simply as a device for mapping a 3D image into a 2D image. A detailed experimental study is carried out in diseased cell structure with the normal one. The real time experiment is to illustrate the actual implementation of the estimation method developed by spatial encoding technique. The innovative hardware and vision-processing algorithm with menu driven user interface was used to analyze the condition of the blood cells in the diseased cell structure. The experiment is planned taking into consideration of the
Page 224
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
above requirements and detection of diseases through microscopic images of the diseased cell structure. The software facilities capturing of high resolution images in color or in black and white, comparison of built up model (Template) with that of subjected model and the images can be sized and stored for future reference. The last stage of the experimental setup in which only one of the diseases would be activated from the output layer of the back propagation algorithm denoting the disease that is to be determined.
image is made to fix into the standard dimensions as shown below.For example INPUT SLIDE SIZE STANDARDIZATION
Î
Color/ Grey Image
Image capturing
(Reduced)
Î (Enlarged) Fig2.1.1inputslidesize Standardization
Segmentation
Binarized Image
Analysis (BP Algorithm)
• Color/Gray image The image is then passed to the above component from which the color image is converted to gray scale using the necessary formula. Weights of the Red, Blue and Green are mixed in proper proportions and then we calculate a single value as required in gray scale. • Segmentation The image is then segmented according to the required values of height and width. When there is some exceeding part of the block they are filled with appropriate values of padding. •
Disease Detection
Feed back to control (Vision System) Figure 2.1 Experimental Setup •
Image capturing
The blood strip taken from the patient’s body is placed in the slide. This is viewed through the electron microscope. A frame is retrieved using the frame grabber. From the frame the region of interest is extracted according to the height and width in number of pixels as required. Thus the
Velammal College of Engineering and Technology, Madurai
Binarized Image
The Image is then binarized into 0 and 1 using necessary approximations.For example terms with values 0.65 are approximated to 1 and terms with 0.35 are considered as 0. •
Back propagation algorithm
Back propagation algorithm is explained in detail later which possesses certain formula so that an image that almost resembles a particular image of pattern is considered to be that pattern. For example if the difference between the two images is negligible or fixed to a certain error rate is considered to have the same pattern. Comparisons are made repeatedly so that the inputted slide is found to either pattern matched or not pattern matched. •
Disease detection
Page 225
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The last stage of the experimental setup in which only one of the diseases would be activated from the output layer of the back propagation algorithm denoting the disease that is to be determined.
Chronic Myeloid Leukemia
Gaucher's Disease
Hypersplenism
Fig 2.2 Slides Containing Blood Cells B. BLOOD: THE VITAL FLUID Blood is a red fluid pumped by the heart to various parts of the body through the circulatory system. No other artificial fluid can replace the function of blood. Blood contains hemoglobin which makes the blood appear red in color. Blood contains 45% of solid content and the remaining liquid. Components of blood Blood contains plasma that is 12% consisting of the solid particles and the remaining liquid. The three major components of blood is • Erythrocytes • Leukocytes • Platelets as shown in the below figure.
organs of the human body. The size of red blood cell is 7.5 microns in diameter. It is spherical in shape. In a human body weighing 75 kg containing 5 litres of blood can have about 1000 red blood cells in it. Red blood cells are almost same in diameter. • Leukocytes Leukocytes are also termed as White blood cells. The function of Leukocytes is to fight against the foreign organisms or the cells getting into the blood due to the various diseases. When compared to Erythrocytes, Leukocytes are large in number. The main way in differencing Erythrocytes and Leukocytes is the nucleus that is only possessed by the Leukocytes. The size of Leukocytes is also bigger in diameter compared to that of Erythrocytes. • Platelets Platelets play a vital role in performing the clotting of blood. Platelets do not have a definite structure. Digital Image Processing Digital image processing is the processing of the image data for storage, transmission and representation for autonomous machine perception. A digital image is composed of a finite number of elements each of which has a particular location and value. These elements are referred to as picture elements, image elements and pixels. “Pixels” is the term most widely used to denote the elements of the digital image. Vision is the most advanced of our senses, so it is not surprising that images play the single most important role in human perception. Thus digital image processing encompasses a wide and varied field of applications. There are fields such as computer vision whose ultimate goal is to use computers to emulate human vision, including learning and being able to make inferences and take actions based on visual inputs. This area is itself a branch of artificial intelligence whose objective is to emulate human intelligence. The area of image analysis also called as “Image understanding” is in between image processing and computer vision. Lets us now discuss about Medical Imaging that is part of Digital image processing.
C.Medical Imaging
Fig 2.2.1
Components Of Blood
• Erythrocytes Erythrocytes are also termed as Red blood cells. The function of Erythrocytes is to carry oxygen to various
Velammal College of Engineering and Technology, Madurai
The parts of the digital image processing that are exclusively used in medical related fields define Medical Imaging. There are numerous applications of the Digital image processing concepts used in medical fields like that of Gamma ray and X-ray imaging. D.Neural network concepts involved The architecture of the human brain is significantly different from the architecture of a conventional computer. The conventional computers are not suited to perform pattern recognitions problems. We therefore borrow features from physiology of the brain as the basis for our new process models. Hence the technology has come to be known as Artificial Neural Systems technology or simple Neural Networks. Thus we
Page 226
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
are going to a concept involved in the Neural Networks termed as Back Propagation networks that plays a vital role in our project in the pattern recognition problem. Back propagation network Back propagation network can be used to solve complex pattern matching problems. The network learns a predefined sets of input/output example pairs by using a two phase propagate adapt cycle. After an input pattern has been applied as a stimulus to the first layer of the network units, it is propagated through each upper layer until an output is generated. This output pattern in then compared with the desired output and an error signal is computed for each output unit. In our project the input is shown in table containing the input slides and the output is the table containing the diseased cell pictures. The significance of the process is that as the network trains the nodes in the intermediate layers organize themselves such that different nodes learn to recognize different features of the total input space. After training when presented with the arbitrary input pattern that is noisy or incomplete, the units is the hidden layer of the network will respond with the active output with the new input contains the pattern that resembles the feature the individual units learn to recognize during training. The error term is calculated using the formula: M Ep = (1/2) ∑δ2pk k=1 where δpk is the error term for the output
activated indicated by 1. Else if the error difference is exceeding Ep then the output is not activated and is indicated by 0. As we know in the Back Propagation network we fix the error value. Using the error value or the threshold we could determine whether the inputted image matched the already trained image. If the difference is below the error or the threshold value then it is obvious that the inputted image is one among the trained sample of the blood. From the figures in the above table we could find the following • When the obtained image is heavily blurred exceeding the error value and thus the result is 0. • When the obtained image is less blurred exceeding the error value and thus the result is 0. E. PATTERN MATCHING Basic principle of pattern matching: Lets consider the following case in which the first matrix in the top to be the slide containing only the diseased cell. Let the second matrix be considered as the slide retrieved from the frame grabber containing the diseased cell. As shown below we can observe that the first matrix is present in the second containing the values 53, 44, 67, 55. In ordinary or classic methods of pattern matching we would obviously think of solving the problem by sliding the first matrix with the second.
units. The above error value Ep is set so that the difference between the obtained and the actual image is fixed to a certain quantity. Result analysis of Error differences by BPN Actual image
Obtained image
Result
0
1
Fig 2.4.1 Result Analysis Of Error Differences By BPN Thus only if the obtained image has an error difference within the allowed error term that is Ep the output is
Velammal College of Engineering and Technology, Madurai
53 67 12 22 54 22
44 55 11 33 76 44
13 53 67 88
14 44 55 55
So the step involved in the normal case is as below: For i= 1 to m2 For j= 1 to n2 For k=1 to m1 For l= 1 to n1 { Processing statements } Thus the running time is as below: Runtime= O(m2*n2*m1*n1) Where ‘O‘denotes the BIG O notation representing the lower bound of any algorithm.
Page 227
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Enhanced pattern matching We now find a method from which we could obviously reduce the running time of the pattern matching involved. The steps involved is described as below • No of pixels of a sample and the slide for height and width is collected for every sample • The first element of the sample is compared with each and every element of the slide • Comparison is done along with the error value of the BPN already prefixed • If at all the first element is equal (considering the error value) to a element then the remaining m2 elements in that row are found for match • If a row is matched a counter is set. • Now comparison is done simply by skipping m1 value thus saving runtime. COMPONENTS OF BLOOD • Pattern is matched if the counter value equals to n2. • Depending upon the matched values that exceeds the threshold value gives a true result. Thus the Enhanced pattern matching analysis is done as below. Running time=O(m1*m2) that is comparatively less to O(m2*n2*m1*n1) Thus time saved =O(m2*n2*m1*n1) - O(m1*m2) Pattern matching using selected points: Cell of Chronic Myeloid Leukemia
Fig 2.5.3 Bloc k Extr actio n
In orde r to still reduce the running time of the algorithm we can select some points may be in some fashion and find out appropriate match. This is represented in the following diagrams. This is represented as below: Considering Slide dimensions as below the total no of comparisons required are 100X100 =10000 TABLE 2.5 IMPROVED ALGORITHM ANALYSIS Block dimensions
NO of Percentage of Total no of blocks running time comparisons required in saved
Velammal College of Engineering and Technology, Madurai
minimum
20X20
8
3200
68
10X10
16
1600
84
5X5
32
800
92
The above is graphically represented below:
Graph Representation 100 90 80 70 60 % 50 40 30 20 10 0
Time saved
50X50 20X20 10X10
5X5
Dimensions
Inappropriate
Less accuracy
Fig2. 5 Graphical Representation Of Algorithm Analysis Thus improved running time considering dimensions 10X10 = 16% [O(m1*m2)] Thus total time saved [Improved Technique] O(m2*n2*m1*n1) - 16%[O(m1*m2)] that is considerably more to that of O(m2*n2*m1*n1) - O(m1*m2) III.RESULTS AND DISCUSSION Blood cell recognition, disease diagnosis and some of the morphological methods are briefly reviewed in this section. This includes experimental investigations and image processing algorithms that have been developed to diagnosis diseases. To overcome the problems associated with blood cells, disease constituents and effectors during the diagnosis process, the image processing technique have been formulated with vision system to provide closed loop control. Several experiments were carried out to characterize the disease of the main modeling steps, i.e., establishing corresponding cells from a set of learning shapes and deriving a statistical model. We applied the modeling procedure to distinct the incoming image with the normal one. We test our methods on 15 images of healthy and 30 images with disease. In our approach, we established the corresponding cells prior to pose and scaling estimation during statistical modeling. This assumes that the position of the cells does not change significantly with different pose and scaling, which is justified by the high accuracy of the model.
Page 228
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
To assess how well the statistical model generalizes to unseen objects, we measured how close a shape not part of the learning sample can be approximated with the statistical shape model. To compute the robustness of the method with respect to different parameter settings, the entire modeling was carried out for a range of parameter at a time while keeping the remaining parameters at their optimal values. The images used are presented here, which have exhibited a wide variety of lesions and confusing manifestations. On this difficult data set, our methods successfully detect the disease in most of the reliable cases. IV.CONCLUSION In this paper, we have presented a method to automatically construct statistical blood cells from segmented microscopical images. Standard models are used to diagnosis the disease that can also be used for haematological applications. Corresponding parameters on the surface of the shape are established automatically by the adaptation of back propagation algorithm to segmented volumetric images. To assess the robustness of the method with respect to the parameter settings, the entire statistical model was carried out for range of parameter values, varying one at a time while keeping the remaining parameters at their optimal values. MS Visual C++ and Matrox Inspector image processing software are used for the experimental work. Nevertheless, this would be an interesting future investigation, since our approach is sufficiently general to be applied to modeling tasks from other two dimensional or three dimensional application domains. The extension of this may open a new area of research in bio mechanics (modeling and simulation). REFERENCES
cells”, Science Direct, Pattern Recognition, Volume 16, Issue 6, Pages 571-577, 1983. 5)Landeweerd.G.H, Timmers.T and Gelsema.E.S,” Classification of normal and abnormal samples of peripheral blood by linear mapping of the feature space”, Science Direct, Pattern Recognition, Volume 16, Issue 3, Pages 319-326, 1983. 6)Lester.J.M, Williams.H.A , Weintraub.B.A and Brenner.J.F, " Two graph searching techniques for boundary finding in white blood cell images”, Science Direct, Computers in Biology and Medicine Volume 8, Issue 4, Pages 293-308, 1978. 7)Parthenis.K, Metaxaki-Kossionides.C and Dimitriadis.B, "An automatic computer vision system for blood analysis ",Science Direct, Microprocessing and Microprogramming, Volume 28, Issues 1-5, Pages 243-246, March 1990. 8)Philip E. Norgren Ashok V. Kulkarni and Marshall D, Graham,“Leukocyte image analysis in the diff3 system”, Science Direct, Pattern Recognition, Volume 13, Issue 4, Pages 299-314, 1981. 9)Schönfeld.M and Grebe.R,”Automatic shape quantification of freely suspended red blood cells by isodensity contour tracing and tangent counting”, Science Direct, Computer Methods and Programs in Biomedicine, Volume 28, Issue, Pages 217-224, April 1989. 10)Xubo B. Song, Yaser Abu-Mostafa, Joseph Sill, Harvey Kasdan and Misha Pavel, “Robust image recognition by fusion of contextual information Information Fusion”, Science Direct, Volume 3, Issue 4, Pages 277-287, December 2002.
1)Artmann.G, Schneider.G and Henning, "Automated image processing system for shape recognition of single red blood cells based on out-of-focus images", Science Direct, Biorheology, Volume 32, Issues 2-3, Pages 237238, March-June 1995. 2)Brenner.J.F, Selles.W.D and Zahniser.D.J, "Pattern recognition of nucleated cells from the peripheral blood”, Science Direct, Pattern Recognition, Volume 16, Issue 2, Pages 131-140, 1983. 3)Daniela Mayumi Ushizima Sabino, Luciano da Fontoura Costa, Edgar Gil Rizzatti and Marco Antonio Zago "A texture approach to leukocyte recognition”, Science Direct, Real imaging in Bioinformatics, Volume 10, Issue 4, Pages 205-216, August 2004. 4)Ldeweerd.G.H, Timmers.T and Gelsema.E.S, ” Binary tree versus single level tree classification of white blood
Velammal College of Engineering and Technology, Madurai
Page 229
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Cloud Computing And Virtualization Mr. Nilesh Madhukar Patil Lecturer (IT Dept.), MCT’s Rajiv Gandhi Institute of Technology, Mumbai, State: Maharashtra, Country: India Email: [email protected]
Mr. Shailesh Somnath Sangle Lecturer (IT Dept.), MCT’s Rajiv Gandhi Institute of Technology, Mumbai, State: Maharashtra, Country: India Email: [email protected] Abstract Cloud Computing is defined as a pool of virtualized computer resources. Based on this Virtualization the Cloud Computing paradigm allows workloads to be deployed and scaled-out quickly through the rapid provisioning of virtual machines or physical machines. A Cloud Computing platform supports redundant, self-recovering, highly scalable programming models that allow workloads to recover from many inevitable hardware/software failures. The paper throws light on how to use cloud computing as a technical design center; how to take advantage of the economics of cloud computing in building and operating cloud services; and how to plan for the technical and operational scale that cloud computing makes possible. Keywords: SaaS, PaaS, IaaS
1. Cloud Computing Fundamentals For an organization to execute its operations, it requires components or hardware and software abided with it to furnish them. For instance, consider that in a software company each and every service is dependent upon other. Starting from Data center, server requires un- interrupted power supply, coolant solutions for its components to deliver effectively, Redundant Servers, Wide Range of Bandwidth to access them, complicated software’s to manage or monitor them, Team of Expert Employees monitoring or refurbishing them, development, production, fail over network. Upon all that suppose a new version or update is introduces than the whole system crashes down or requires a reallocation of resources. Ultimately the investment is found to be high but the results doesn’t seem stable or ever requires technical expert’s involvement. The search was up for a service which is agile, could be spread wider as time grows and could contract with the same speed as it grew, automatic updates without any user intervention, no knowledge or little requirement of technical trouble shooting expertise.
Velammal College of Engineering and Technology, Madurai
Figure 1.1: Cloud Computation
Cloud computing is using virtualized resources as a service over the internet with provision of scaling up dynamically. The improvement is that the users need not have to have knowledge of, expertise in, or control over the technology infrastructure over the “Cloud” that supports them. It operates through browsers while the software and the data are stored in the servers. A technical definition is "a computing capability that provides an abstraction between the computing resource and its underlying technical architecture (e.g., servers, storage, networks), enabling convenient, on-demand network access to a shared pool of configurable computing resources that can be rapidly provisioned and released with minimal management effort or service provider interaction." There have been many well known participants (ZOHO, Microsoft,
Amazon,
Rackspace
Cloud,
SalesForce,
Sun)
to
the
Yahoo, Cloud
Google, Computing
technologies, blowing the trumpet at high peak by introducing customary packages about the services offered by the Cloud Computing.
Page 230
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Figure 2.1: Virtualization Bundle Figure 1.2: Cloud Hosters
2. Virtualization Whenever an application or a service, is being developed by a company, its reliability to deliver at the native environment requires all its dependent services or resource to be bundled together due to the reason that the Application and Infrastructure are found to be dependent. Here technologies advanced to find a solution like bundling application with all it required services like Database, subsequent immediate services as well as Operating System into one, thereafter making Application independent of the Infrastructure into which the application is deployed. Yes very close, fueling the application to work independent of the Infrastructure is the architecture terms as virtualization. It refers to the Abstraction of computer resource. With the emergence of Virtualization, the internet cloud gathered multiple resources to equip for the Virtualization. Most of the time, servers don't run at full capacity. That means there's unused processing power going to waste. It's possible to fool a physical server into thinking it's actually multiple servers, each running with its own independent operating system. The technique is called server virtualization. By maximizing the output of individual servers, server virtualization reduces the need for more physical machines. Maneuvering virtualization technique according to the demands of the vendors is the simplest definition for Cloud Computing.
Velammal College of Engineering and Technology, Madurai
GMAIL is a Cloud Computing services. Now creating GMAIL account does not takes much time rather than a username and password. That’s it. This is the benefit of cloud computing. No server restart, no software updated, Just a simple username, password authentication, and then the resource is ready to be accessed by the end user. Cloud computing doesn’t only help customer application like “Account”, “Web Sites” but also support Business Architecture As such. Thereby Entitling it self as Enterprise Cloud Computing. Business Associates could pay for the resources they use. No additional cost/No wastage of Cost/Less cost. Talking about scalability, the cloud computing is termed as Elastic, enables customers to increase or decrease capacity within minutes, not hours or days. You can commission one, hundreds or even thousands of server instances simultaneously. Of course, because this is all controlled with web service APIs, your application can automatically scale itself up and down depending on its needs. To the End user, it is approached as a Multi-tenancy architecture, where in each compartment is handled individually without sharing resource for other compartment, but the fact is it is a shared platform where in individual Resource is being shared with all the available vendors compartmented by policies and permissions. Users have started rapidly using Cloud Computing technologies, and its usage is being encouraged. Few of the hosting service providers has prompted usage of Cloud Computing to help its reach to the common people have provided multiple shared services for free like Google Document (Sharing / Storing Documents in the cloud) and ZOHO Document (Sharing / Storing Documents in the cloud) are few to be mentioned.
Page 231
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3.1 Virtualized Infrastructure Virtualized Infrastructure provides the abstraction necessary to ensure that an application or business service is not directly tied to the underlying hardware infrastructure such as servers, storage, or networks. This allows business services to move dynamically across virtualized infrastructure resources in a very efficient manner, based upon predefined policies that ensure specific service level objectives are met for these business services. 3.2 Virtualized Applications Virtualized applications decouple the application from the underlying hardware, operating system, storage, and network to enable flexibility in deployment. Virtualized Application servers that can take advantage of grid execution coupled with Service Oriented Architectures and enable the greatest degree of scalability to meet the business requirements. 3.3 Enterprise Management Figure 2.2: Multitenancy Conceptual
Architecure
3. Cloud Building Blocks The building blocks of cloud computing are rooted in hardware and software architectures that enable innovative infrastructure scaling and virtualization. Many data centers deploy these capabilities today. However, the next infrastructure innovations are around more dynamic provisioning and management in larger clusters both within and external to the conventional corporate data center. There are also implications for next generation application design to make optimum use of massively parallel processing and fault tolerance. The figure 3 below illustrates some common architectural components:
Enterprise management provides top-down, end-to-end management of the virtualized infrastructure and applications for business solutions. The enterprise management layer handles the full lifecycle of virtualized resources and provides additional common infrastructure elements for service level management, metered usage, policy management, license management, and disaster recovery. Mature cloud service management software allows dynamic provisioning and resource allocation to allow applications to scale on demand and minimize the waste associated with underutilized and static computing resources. 3.4 Security and Identity Management Clouds must leverage a unified identity and security infrastructure to enable flexible provisioning, yet enforce security policies throughout the cloud. As clouds provision resources outside the enterprise’s legal boundaries, it becomes essential to implement an Information Asset Management system to provide the necessary controls to ensure sensitive information is protected and meets compliance requirements. 3.5 Development tools
Figure 3: Cloud Components
Velammal College of Engineering and Technology, Madurai
Next generation development tools can leverage cloud’s distributed computing capabilities. These tools not only facilitate service orchestration that can leverage dynamic provisioning, but also enable business processes to be developed that can harness the parallel processing capabilities available to clouds. The development tools
Page 232
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
must support dynamic provisioning and not rely on hard coded dependencies such as servers and network resources. 4. Cloud Computing Infrastructure Models There are many considerations for cloud computing architects to make when moving from a standard enterprise application deployment model to one based on cloud computing. There are three basic service models to consider: Public, private, and hybrid clouds IT organizations can choose to deploy applications on public, private, or hybrid clouds, each of which has its trade-offs. The terms public, private, and hybrid do not dictate location. While public clouds are typically “out there” on the Internet and private clouds are typically located on premises, a private cloud might be hosted at a colocation facility as well. Companies may make a number of considerations with regard to which cloud computing model they choose to employ, and they might use more than one model to solve different problems. An application needed on a temporary basis might be best suited for deployment in a public cloud because it helps to avoid the need to purchase additional equipment to solve a temporary need. Likewise, a permanent application, or one that has specific requirements on quality of service or location of data, might best be deployed in a private or hybrid cloud. 4.1 Public clouds
virtual machine images, but also servers, storage systems, network devices, and network topology. Creating a virtual private datacenter with all components located in the same facility helps to lessen the issue of data locality because bandwidth is abundant and typically free when connecting resources within the same facility.
Figure 4.1: Public Cloud
4.2 Private clouds Private clouds are built for the exclusive use of one client, providing the utmost control over data, security, and quality of service (Figure 4.2). The company owns the infrastructure and has control over how applications are deployed on it. Private clouds may be deployed in an enterprise datacenter, and they also may be deployed at a colocation facility. Private clouds can be built and managed by a company’s own IT organization or by a cloud provider. In this “hosted private” model, a company such as Sun can install, configure, and operate the infrastructure to support a private cloud within a company’s enterprise datacenter. This model gives companies a high level of control over the use of cloud resources while bringing in the expertise needed to establish and operate the environment.
Public clouds are run by third parties, and applications from different customers are likely to be mixed together on the cloud’s servers, storage systems, and networks (Figure 4.1). Public clouds are most often hosted away from customer premises, and they provide a way to reduce customer risk and cost by providing a flexible, even temporary extension to enterprise infrastructure. If a public cloud is implemented with performance, security, and data locality in mind, the existence of other applications running in the cloud should be transparent to both cloud architects and end users. Indeed, one of the benefits of public clouds is that they can be much larger than a company’s private cloud might be, offering the ability to scale up and down on demand, and shifting infrastructure risks from the enterprise to the cloud provider, if even just temporarily. Portions of a public cloud can be carved out for the exclusive use of a single client, creating a virtual private datacenter. Rather than being limited to deploying virtual machine images in a public cloud, a virtual private datacenter gives customers greater visibility into its infrastructure. Now customers can manipulate not just
Velammal College of Engineering and Technology, Madurai
Figure 4.2: Private Cloud
4.3 Hybrid clouds Hybrid clouds combine both public and private cloud models (Figure 4.3). They can help to provide on-demand, externally provisioned scale. The ability to augment a private cloud with the resources of a public cloud can be used to maintain service levels in the face of rapid workload fluctuations. This is most often seen with the use
Page 233
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of storage clouds to support Web 2.0 applications. A hybrid cloud also can be used to handle planned workload spikes. Sometimes called “surge computing,” a public cloud can be used to perform periodic tasks that can be deployed easily on a public cloud.
market, including the Google Apps offering of basic business services including email and word processing. Although salesforce.com preceded the definition of cloud computing by a few years, it now operates by leveraging its companion force.com, which can be defined as a platform as a service.
Hybrid clouds introduce the complexity of determining how to distribute applications across both a public and private cloud. Among the issues that need to be considered is the relationship between data and processing resources. If the data is small, or the application is stateless, a hybrid cloud can be much more successful than if large amounts of data must be transferred into a public cloud for a small amount of processing.
Figure 5: Cloud computing means using IT infrastructure as a service
5.2 Platform as a service (PaaS) Platform as a service encapsulates a layer of software and provides it as a service that can be used to build higherlevel services. There are at least two perspectives on PaaS depending on the perspective of the producer or consumer of the services:
Figure 4.3: Hybrid Cloud
5. Architectural layers of cloud computing In practice, cloud service providers tend to offer services that can be grouped into three categories: software as a service, platform as a service, and infrastructure as a service. These categories group together the various layers illustrated in Figure 5, with some overlap. 5.1 Software as a service (SaaS) Software as a service features a complete application offered as a service on demand. A single instance of the software runs on the cloud and services multiple end users or client organizations. The most widely known example of SaaS is salesforce.com, though many other examples have come to
Velammal College of Engineering and Technology, Madurai
• Someone producing PaaS might produce a platform by integrating an OS, middleware, application software, and even a development environment that is then provided to a customer as a service. For example, someone developing a PaaS offering might base it on a set of Sun™ xVM hypervisor virtual machines that include a NetBeans™ integrated development environment, a Sun GlassFish™ Web stack and support for additional programming languages such as Perl or Ruby. • Someone using PaaS would see an encapsulated service that is presented to them through an API. The customer interacts with the platform through the API, and the platform does what is necessary to manage and scale itself to provide a given level of service. Virtual appliances can be classified as instances of PaaS. A content switch appliance, for example, would have all of its component software hidden from the customer, and only an API or GUI for configuring and deploying the service provided to them.
Page 234
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PaaS offerings can provide for every phase of software development and testing, or they can be specialized around a particular area such as content management. Commercial examples of PaaS include the Google Apps Engine, which serves applications on Google’s infrastructure. PaaS services such as these can provide a powerful basis on which to deploy applications, however they may be constrained by the capabilities that the cloud provider chooses to deliver. 5.3 Infrastructure as a service (IaaS) Infrastructure as a service delivers basic storage and compute capabilities as standardized services over the network. Servers, storage systems, switches, routers, and other systems are pooled and made available to handle workloads that range from application components to highperformance computing applications. Commercial examples of IaaS include Joyent, whose main product is a line of virtualized servers that provide a highly available on-demand infrastructure. 6. Benefits of Cloud Computing 1. Decoupling and separation of the business service from the infrastructure needed to run it (virtualization) 2. Flexibility to choose multiple vendors that provide reliable and scalable business services, development environments, and infrastructure that can be leveraged out of the box and billed on a metered basis—with no long term contracts. 3. Elastic nature of the infrastructure to rapidly allocate and de-allocate massively scalable resources to business services on a demand basis. 4. Cost allocation flexibility for customers wanting to move Capital Expenditures (CapEx) into Operational Expenditures (OpEx). 5. Reduced costs due to operational efficiencies, and more rapid deployment of new business services. 7.
Issues in Cloud Computing
The main concerns about cloud computing are security and privacy. The thought of handing your important data over to something called a “cloud” can be daunting. Nervous corporate executives might hesitate to take advantage of a cloud computing system because they feel like they’re surrendering control of their company’s information. Data inside the ‘cloud’ is outside a company’s firewall and that brings with it an intrinsic threat of risk, because services that companies outsource evade the physical, logical and
Velammal College of Engineering and Technology, Madurai
personnel controls that I.T. shops wield over data maintained in-house. Other fears include: 1. Risk of data breaching 2. Appeal to cyber crooks 3. Lack of specific standards for security and data privacy 4. Questions about jurisdiction. European concern about U.S. privacy laws led to creation of the U.S. Safe Harbor Privacy Principles, which are intended to provide European companies with a degree of insulation from U.S. laws 5. Data location. Cloud users probably do not know exactly where their data is hosted—not even the specific country. 8.
Conclusion
Cloud computing offers real alternatives to IT business for improved flexibility and lower cost. Markets are developing for the delivery of software applications, platforms, and infrastructure as a service to IT business over the “cloud”. These services are readily accessible on a pay-per-use basis and offer great alternatives to businesses that need the flexibility to rent infrastructure on a temporary basis or to reduce capital costs. Architects in larger enterprises find that it may still be more cost effective to provide the desired services in-house in the form of “private clouds” to minimize cost and maximize compatibility with internal standards and regulations. If so, there are several options for future-state systems and technical architectures that architects should consider to find the right trade-off between cost and flexibility. Using an architectural framework will help architects evaluate these trade-offs within the context of the business architecture and design a system that accomplishes the business goal. 9. References [1] Michael Miller, “Cloud Computing: Web-Based Applications That Change the Way You Work and Collaborate Online”, Que, 2008. [2] David Linthicum, “Cloud Computing and SOA Convergence in Your Enterprise: A Step-by-Step Guide”, 1st edition, Addison-Wesley Professional, 2009. [3] Toby Velte, Anthony Velte, Robert Elsenpeter, “Cloud Computing, A Practical Approach”, 1st edition, McGrawHill Osborne Media, 2009. [4]http://www.exforsys.com/tutorials/cloudcomputing/cloud-computing-users-perspective.html
Page 235
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[5]http://www.watblog.com/2008/03/25/yahoocomputational-research-laboratories-team-up-for-cloudcomputing-research/ [6]http://computersoftware.suite101.com/article.cfm/cloud _computing_and_virtualization_impact [7]http://www.microsoft.com/virtualization/en/us/privatecloud.aspx
Velammal College of Engineering and Technology, Madurai
Page 236
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
New Method For Solving Fuzzy Linear Programming With TORA S. Sagaya Roseline , A. Faritha Asma , E.C. Henry Amirtharaj, Department of Mathematics , Bishop Heber College (Autonomous) Tiruchirappalli -17, India. Email: [email protected]
[email protected]
Abstract - Since many real-world engineering systems are too complex to be defined in precise terms, imprecision is often involved in any engineering design process. Fuzzy linear programming problems have an essential role in fuzzy modeling, which can formulate uncertainty in actual environment. In this study we present a new method for solving FLPP with TORA software. Here the coefficient of objective function , constraint coefficient matrix, and right hand side of the constraints are in fuzzy nature. The coefficients of the objective function have taken as trapezoidal fuzzy number, the coefficients of the constraints and the right hand side of the constraints are considered as a triangular fuzzy number. Here our main scope is to find some non-negative vector x which maximizes the objective function z = c x so that A x = b . Finally the defuzzified LPP is solved by using TORA.
Keywords - Fuzzy sets, Fuzzy number, Trapezoidal fuzzy number, Triangular fuzzy number, Fuzzy linear programming.
I. INTRODUCTION Fuzzy set theory has been applied to many fields of Operations Research. The concept of fuzzy linear programming (FLP) was first formulated by Zimmermann. And then various types of the FLP problems are considered by many authors and several approaches for solving these problems are proposed. In this paper we consider a LPP in which the coefficients of the objective function, constraint coefficients and right hand side of the constraints are fuzzy. Here we explain the concept of the comparison of fuzzy numbers by introducing a linear ranking function. Our main contribution here is the establishment of a new method for solving the Fuzzy linear programming problems with TORA. Moreover we illustrate our method with an example. II. PRELIMINARY
[email protected]
A. FUZZY SETS Definition Let X be a classical set of objects called the universe whose generic elements are denoted by x. The membership in a crisp subset of X is often viewed as characteristic function μA(x) from X to {0, 1} such that μA(x) = 1 , if x Є A = 0 , otherwise. where {0, 1} is called valuation set. If the valuation set is allowed to be the real interval [0, 1], A is called a fuzzy set. μA(x) is the degree of membership of x in A. The closer the value of μA(x) is to 1, the more x belong to A. Therefore, A is completely characterized by the set of ordered pairs: A = { ( x, μA(x)) / x Є X } Definition The support of a fuzzy set A is the crisp subset of X and is presented as : Supp A = { x Є X / μA(x) > 0 } Definition The α level (α – cut ) set of a fuzzy set A is a crisp subset of X and is denoted by Aα = { x Є X / μA(x) ≥ α) } Definition A fuzzy set A in X is convex if μA( λx + (1-λ)y) ≥ min { μA(x) , μA(y) } x , y Є X and λ Є [0, 1] . Alternatively, a fuzzy set is convex if all α level sets are convex . Note that in this paper we suppose that X = R. B. Fuzzy numbers
In this section we review some necessary backgrounds of the fuzzy theory which will be used in this paper. Below we give definitions and notations.
Definition A fuzzy number A is a convex normalized fuzzy set on the real line R such that 1) it exists atleast one x0 Є R with μA(x0) = 1. 2) μA(x) is piecewise continuous.
Velammal College of Engineering and Technology, Madurai
Page 237
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Among the various types of fuzzy numbers, triangular and trapezoidal fuzzy numbers are of the most important. Definition A fuzzy number is a trapezoidal fuzzy number if the membership function of it be in the following form :
x > 0 , x a = ( xaL , x aU , x α, x β ) x < 0 , x a = ( xaU , xaL , - x β, - x α ) IV. RANKING FUNCTIONS A convenient method for comparing of the fuzzy numbers is by use of ranking functions. A ranking function is a map from F(R) into the real line. Now , we define orders on F(R) as following:
μ a
a ≥ b if and only if R (a) ≥ R (b) R x
1 x 0
aL - α
aL
aU
aU + β
We show that trapezoidal fuzzy number by a = (aL , aU , α, β ) where the support of a is ( aL - α , aU + β ) , and the modal set of a is [aL , aU ] . Let F(R) be the set of trapezoidal fuzzy numbers. Definition Triangular fuzzy number A can be represented by three real numbers, s, l, r whose meanings are defined as
1
l
r
R
a > b if and only if R (a) > R (b) R a = b if and only if R (a) = R (b) R where a and b are in F(R). It is obvious that we may write a ≤ b if and only if b ≥a R R Since there are many ranking function for comparing fuzzy numbers we only apply linear ranking functions. So, it is obvious that if we suppose that R be any linear ranking function, then i) a ≥ b if and only if a - b ≥ 0 if and only if - b ≥ - a R R ii) a ≥ b and c ≥ d , then a + c ≥ b + d R R R One suggestion for a linear ranking function is following: R (a) = aL + aU + (β - α) where a = ( aL , aU , α, β ) .
0 s-l
s
s+r
Using this representation we write A = (s, l, r ) Let F1 (R) be the set of all triangular fuzzy numbers. III. ARITHMETIC ON TRAPEZOIDAL FUZZY NUMBERS Let a = ( aL , aU , α, β ) and b = ( bL , bU , γ , θ ) be two trapezoidal fuzzy numbers and x Є R. Then the results of applying fuzzy arithmetic on the trapezoidal fuzzy numbers are shown in the following : U
L
Image of a : - a = (- a , - a , β, α ) Addition : a + b = ( aL + bL , aU + bU , α + γ, β + θ ) Scalar Multiplication:
Velammal College of Engineering and Technology, Madurai
V. FUZZY LINEAR PROGRAMMING PROBLEM Consider Fuzzy linear programming problem in which the coefficient of the objective function , the right hand side numbers and the coefficients of the constraint matrix are fuzzy numbers. Max z = xj R Sub to
xj ≤ Bi (i Є Nm) xj ≥ 0 (j Є Nn)
According to [2] the problem can be rewritten as Max z = R
xj
Page 238
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Sub to
,
,
)
( ti , ui, vi )
( i Є Nm )
Max z = ( 5, 8, 2, 5) x1 + ( 6, 10, 2, 6) x2
xj ≥ 0 ( j Є Nn )
Sub to 4x1 + 5x2 ≤ 24 4x1 + x2 ≤ 12 2x1 + 2x2 ≤ 19 3x1 +0.5x2 ≤ 6 5x1 + 6x2 ≤ 32 6x1 + 2x2 ≤ 15 x1 , x2 ≥ 0
where Aij = ( sij , lij , rij ) and Bi = ( ti , ui, vi ) are triangular fuzzy numbers.cj Є F(R) Summation and multiplication are operations on fuzzy numbers, and the partial order ≤ is defined by A ≤ B iff MAX ( A, B) = B. For any two triangular fuzzy numbers A = (s1, l1, r1) and B = (s2 , l2 , r2 ) A ≤ B iff s1 ≤ s2 , s1 - l1 ≤ s2 - l2 and s1 + r1 ≤ s2 + r2. Moreover, (s1, l1, r1) + (s2 , l2 , r2 ) = (s1+s2 , l1+l2 , r1+r2 ) and (s1, l1, r1)x = (s1x, l1x, r1x) for any non-negative real number x. Then the problem can be rewritten as Max z = R
R(z) = 86.515 Defuzzifing the objective function using ranking function , the FLPP becomes the following classical LPP.
xj
Sub to
Solving this problem we obtain, x1 = 1.23, x2 = 3.82 max z = ( 29.04 , 48, 10.09 , 29.04)
Max z = 14.5 x1 + 18 x2
≤ ti
Sub to 4x1 + 5x2 ≤ 24 4x1 + x2 ≤ 12 2x1 + 2x2 ≤ 19 3x1 +0.5x2 ≤ 6 5x1 + 6x2 ≤ 32 6x1 + 2x2 ≤ 15 x1 , x2 ≥ 0
≤ ti - ui ≤ ti + vi (i Є Nm) xj ≥ 0 ( j Є Nn ) Now the Fuzzy linear programming problem becomes Max z = c x R Sub to A x ≤ b
Solving this LPP using TORA we obtain, x1 = 1.23, x2 = 3.82 , max z = 86.52 (1)
x≥0 c T Є (F(R))n , b Є Rm , x Є Rn , R is a linear ranking function. Then according to [1] , the problem can be solved by using simplex method. By using ranking function , objective function of (1) can be defuzzified and then (1) is equivalent to the classical LPP which can be solved by using TORA.
VII. CONCLUSION In this paper we proposed a new method for solving the FLP problems with TORA. The significance of this paper is providing a new method for solving the fuzzy linear programming in which the coefficients of the objective function are trapezoidal fuzzy number and the coefficients of constraints, the right hand side of the constraints are triangular fuzzy number. We compared the solution of FLPP with the defuzzified LPP solution using TORA. REFERENCES
VI. A NUMERICAL EXAMPLE For an illustration of the above method we solve a FLP problem with TORA.
[1] [2]
Max z = ( 5, 8, 2, 5) x1 + ( 6, 10, 2, 6) x2
[3]
Sub to ( 4, 2, 1) x1 + ( 5, 3, 1) x2 ≤ ( 24, 5, 8) ( 4, 1, 2) x1 + ( 1, 0.5,1)x2 ≤ ( 12, 6, 3) x1 , x2 ≥ 0
[4] [5]
S. H. Nasseri, E. Ardil, A. Yazdani and R. Zaefarian, Simplex method for solving linear programming problems with fuzzy numbers, World academy of science, engineering and technology,2005 . George j. Klir / Bo Yuan, Fuzzy sets and fuzzy logic Theory and applications, 2006. S.H. Nasseri, A new method for solving fuzzy linear programming by solving linear programming, Applied mathematical sciences,vol.2, 2008. H. Rommelfanger, R. Hanuscheck and J. Wolf, Linear programming with fuzzy Objective funcion, Fuzzy sets and systems,1989. H.J. Zimmermann, Fuzzy programming and linear programming with several objective functions, Fuzy sets and systems, 1978.
we can rewrite it as
Velammal College of Engineering and Technology, Madurai
Page 239
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Latest Trends And Technologies In Enterprise Resource Planning – ERP Author–Dakshayani.B.S. Department of Computer Science, Amrita Vishwa Vidyapeetham, Mysore Campus, Karnataka. [email protected]
Abstract—This article deals with, a) The Latest trends in ERP, b) Recent technologies like Open Source, Web 2.0 services, SaaS, Cloud Computing and Green ERP Systems, Compatibility with dot net technologies c) Future of ERP Key words—Green ERP, Cloud Computing, SaaS, Wireless Technology, Web based ERP, Compatibility of ERP with Dot net technology
I INTRODUCTION ERP evolved from Manufacturing Resource Planning – MRP II (which originated from Material Resource Planning – MRP I). It has gained much prominence and utility with the intervention of open source, web enabled and wireless technologies. ERP has undoubtedly become an important business application to all industries. It has almost become a must for all organizations irrespective of the type of business manufacturing or service. In this context it becomes important to analyze the direction in which ERP is geared to progress or will ERP diminish in the future, emerging technologies etc. II LATEST TRENDS IN ERP ERP calls for constant modifications and up gradations. ERP developers are facing tremendous pressure both from vendors and companies. In this context it becomes important to analyze the ERP’s trends and modalities. The following are relevant issues in ERP. A. Need based application Organizations had to implement ERP throught their systems irrespective of the fact whether they help in all the functions or in one particular function. This was proving to be a big hurdle to the firms. In addition this remained as the main disadvantage or setback of ERP. They had to purchase the whole applications even if it meant that most of them would be idle except for the core function. The latest ERP software programs have overcome this menace. They offer need based applications.. They were given the liberty to purchase and install software Programs pertaining to that particular function. This advantage has
Velammal College of Engineering and Technology, Madurai
helped to increase the scope of ERP not only among large firms but also small and medium business as well. B. Expenditure ERP was a very costly affair. Thanks to the intrusion of internet and open source applications. This has helped S.M.E.'S to enter the market of prospective buyers. This has not only widened the horizon of S.M.E.'s but also increased the usage among large firms. These large firms were not able to invest huge money in spite of adequate funds. Now that the spending on ERP gets reduced there are no hesitations to show the green signal for monetary outlay. It is encouraging to notice the improving IT ERP trends. C. Implementation Time ERP was discouraged by companies because they took such a long time to get installed and set the whole process into action. Since this resource was spent excessively there were chances for reduction in potential business and losing manhours. The current day ERP applications are less complex to install and train. This has reduced the amount of time spent on ERP. Companies are thereby assured of spending lesser time for ERP. D. Open Source, Web enabled (Web 2.0 services), SaaS, Cloud Computing, Green ERP and wireless technologies These technologies are the results of finding solution to the issues discussed above. They have been dealt in detail in the following paragraphs. These technologies have helped in rejuvenation of the functioning of ERP and also in creating wide market base, III RECENT TECHNOLOGIES IN ERP A. Open Source ERP technologies Open Source Technologies have made the job of ERP easier. It has done away with the hassles of paying licence fee not only during installation but also whenever a modification is made. The company is relieved from depending even for mince matters. Following issues relating to open source ERP technology is worth discussing. 1) Influence of the cost
Page 240
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
factor: It is interesting to know the Price tag of Open Source ERP technologies. It is literally available at free of cost. All that the user has to do is to download the software and install it. An unbelievable fact is that even the source code comes freely. This in itself has encouraged lots of companies to go for ERP, as they are not burdened with investments. Of late, companies don't necessarily go for ERP because their attitude towards spending on ERP has undergone a sea change in the sense they don't mind to pay as long as they think ERP is worth the costs. Open Source accounting ERP and open source ERP payment are famous solutions. 2) Influence on the operational expenses: Open Source ERP technologies largely influence operational expenses. The company is relieved from paying the extra sum for facilities provided in moving to a new system. Similarly the company need not incur additional expenditures for renewal and purchase of licenses as and when some components are added in the framework. This gradually reduces the monetary outlay that has to be otherwise incurred for every update. Open Source accounting ERP has helped to simplify the financial practices. Open source ERP payment has helped in facilitating easy disbursement of cash. 3) Absence of Vendors help: Unlike the usual ERP applications it is not possible to avail the services of a vendor as the company handles everything independently. This has many dimensions. Firstly the company enjoys a sole liability. Secondly a simple error when not rectified (or the in-house personnel does not know to) could prove to be a costly affair for the company. Above all the company gets to learn from mistakes and without any external assistance 4) Litigations: Open source ERP has resulted in many lawsuit and incidental claims. There is still ambiguity in the copying aspects. The question of infringement and indemnification remains unanswered as seen from the previous cases. 5) Unsuitable for all applications: Open source has a limit when it comes to the places where they can be put to use. They don't find applicability in all areas especially for conventional practices. It is not appropriate to introduce open source in those areas without changing the way the systems work. In fact it could be a risky option to do it. This drawback discourages many functions from being Open source friendly
B. Web Based ERP This Web Based ERP is usually less expensive than a fullblown ERP software package which is usually priced out of the range for a small or medium company. Also with fullblown packages, there will probably be many modules that will not be implemented, even though those modules have been purchased. On the other hand, Web Based ERP can be purchased and implemented in stages. You only buy what you need. Since Web based ERP is based upon the web, it can be used from anywhere you have access to the internet or even
Velammal College of Engineering and Technology, Madurai
through an intranet. This means that financial information’s available whenever it is needed, from wherever you happen to be. This is an important criterion for people like Sales Reps to corporate CEO’s. Sales analysis, profitability, inventory, production, pricing, credit history: these are just some the information that’s available just by using the web to securely connect, in real-time, to a secure database SaaS – ERPshe Software as a Service (SaaS) model is a way of providing the same software to different customers via a network, usually the internet. In other words, the software is not hosted on the customer’s individual computers. Under the SaaS model, a vendor is responsible for the creation, updation, and maintenance of software. Customers buy a subscription to access it, which includes a separate license, or seat, for each person that who use the software. This model can add efficiency and cost savings for both the vendor and customer. Customers save time and money since they do not have to install and maintain programs. The customers do not have to hire staff, or use existing staff to maintain the software. They also generally do not have to buy any new hardware. This allows a customer to focus more resources on growing the business.Shifting the burden of software hosting and development to the vendor can also speed up the time it takes for the customer to see a return on the software investment. Using the SaaS model, the number of seats can be increased as the business grows. This is usually faster and cheaper than purchasing another license and adding to another computer, as with traditional software.Vendor usually only have to update and maintain the software on the network, versus updating different copies of the software on different computers. This allows the vendor to provide the latest updates and technology to each customer in a timely manner. The drawback for the customer is that they do not control the software and customization of programs may be limited. If an update is requested by a customer, it will most likely need to benefit other customers who are also using the same software. If the customer completely outgrows the software, however, the company can simply discontinue its subscription at the end of the current contract. In such a cancellation, applications typically do not have to be removed from the customer’s computers. Generally, the canceling customer maintains ownership of any proprietary data entered into the SaaS applications. SaaS model contracts may be terminated early with sufficient cause. The vendor not delivering the software in a timely manner, the software not working as specified in the contract, are all typically grounds for termination of the contract. With broadband technology more commonplace throughout the workforce, however, customers have many
Page 241
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
choices when it comes to software delivered under the SaaS model. Customers can research SaaS vendors thoroughly, and request current references, to help avoid any nondelivery issues. Plex Online Company utilizes SaaS model to deliver fullblown manufacturing ERP software functions in an ondemand mode. Plex online is a web based ERP solution, requiring only a web browser to access all of the analysis and management functions of a full-featured manufacturing ERP system, designed to meet the requirements of all departments within a quality-driven manufacturer, shop floor to top floor. Web 2.0 and SaaS software have captured a significant portion of the market led by legacy ERP systems. Even large enterprises (which only showered their love on ERP’s) are now accepting the value of SaaS application over legacy ERP’s. A survey about SaaS conducted in August 2009 by Business Week Research Service found that four out of five managers and senior executives in North America are either interested in, or in the process of, adopting the SaaS approach to IT. In fact roughly a third of the 326 respondent’s companies have already fully or partially adopted the SaaS approach for at least the one application. The key features of SaaS applications that are usually missing in traditional ERP software – usability, intuitive Graphic User Interface, collaboration, global accessibility, configurable, higher security levels and zero IT hassles for clients.
D. Cloud Computing Represents the next evolutionary step toward elastic IT. It will transform the way your IT infrastructure is constituted and managed, through consumable services for infrastructure, platform and applications. This will convert your IT infrastructure from a “factory” into a “supply chain”. Cloud computing is a general term for anything that involves delivering hosted services over the Internet. These services are broadly divided into three categories: Infrastructure-as-a-Service (IaaS), Platform-as-a-Service (PaaS) and Software-as-a-Service (SaaS). The name cloud computing was inspired by the cloud symbol that's often used to represent the Internet in flowcharts and diagrams. Cloud Computing SaaS is the next generation ERP and optimized for B2B/SMB’s with built-in CRM and SFA to cut costs and improve productivity.
Velammal College of Engineering and Technology, Madurai
The highest level of Cloud Computing is SaaS, BScaler Cloud computing SaaS is a leader in SaaS, using Web 2.0, Web Services, SOA, AJAX, XML, Java and SaaS 2.0 technologies to provide users with the highest security and data access performance via internet browsers. With BScaler ERM(Enterprise Resource Manager) SaaS, no upfront investment in hardware or software is required, customers do not have to operate their applications or IT infrastructure and maintenance and updates are no longer necessary! Furthermore, wherever you are in the world – at any time – your business will be with you. All you need is browser! This software also lets users focus on the businesses instead of an over-loaded, quickly aging data centers. E. Green ERP Scientists and intellectuals have become increasingly concerned with the human impact on our environment. Governments are now regulating businesses with strict environmental restrictions. Software systems such as Green ERP are gaining popularity as companies transform their practices. This concept is relatively new, so very few companies offer ERP systems that integrate environmental goals into the system’s structure. The best option is tailored software. Tailored ERP software utilizes the best features of both generic (turn-key) and custom software while avoiding all of the disadvantages. The concept, tailoring software was started to allow flexibility in a software system while avoiding the pit falls of a custom system such a high cost, having to rebuild the system to upgrade and long development time. The best tailored or configured software begins with an existing software structure. Key individuals at the company help to model the software’s functionality, but unlike custom coding, the lengthy process of developing the software is circumvented by high-end programming tools. Since the basic architecture’s not altered, upgrades as simple as with a generic, out-of-the-box software system. F. Wireless technology in ER: Wireless technology has helped enterprise operations in many ways. Firstly it has facilitated the stakeholders in getting upto-date information on enterprise operations as and when required through the use of modern communication devices like mobile phones, laptops (all connected to the internet).
Page 242
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1) Proper coordination with the organizational network: Wis ERP will give the best results only if it falls in line with proper communication channels. The communication channels should be improved in the organization to make it (Wireless) ERP friendly. The obsolete computers should be replaced with the latest phones. Mobile and telecommunication facilities should also be at par with the industry standards. It happens many a times that companies resort to wireless ERP without improving the communication facilities. This does not serve the purpose no matter whatsoever the money spent in establishing ERP or the resource persons employed to procure the same. In addition there is another major advantage in improving communication facilities. Apart from dissemination of information and inflating profits and improving productivity it also helps the companies to rise to professional standards in the market. This will also motivate the companies to improve all other facilities that directly or indirectly contribute to the working of ERP and make use of facilities like image Enabled ERP system and ERP data capture. 2) Privacy issues: Privacy is a burning issue that occupies significance whenever there is a technological explosion and ERP is no exception to this principle. These calls for more attention due to the (further) improvement namely wireless ERP (from ERP). Privacy always becomes a subject matter of conflict whenever things are brought under public domain. The company is poised to lose valuable information or prospective businesses when things are exposed unnecessarily .In those cases wireless ERP will be a disaster to the company (and not a boon). Wireless ERP should be provided with maximum security. The scrutiny should prevent third party access without prior permission and approval from the company. This might look impossible practically but it is strongly advocated to have them in force in order to safeguard the interests of the company. The security system will be successful if a department in the company is able to have a track of the details communicated. The viability of the company is also to be taken into account in this regard. Bigger companies can go for such system whereas the others should restrict depending on their capacity and needs. If they are required unconditionally the companies should resort to alternative arrangements .It should be kept in mind that none of these arrangements should interfere with the functioning of ERP. Image Enabled ERP system and ERP data capture will help to achieve this.
improvement. This will automatically help wireless ERP to succeed. 4) Multiple applications and connectivity speed: Wireless ERP should not depend on just one or two devices like laptop or mobile phone but make use of maximum appliances. This will help the companies to gain technological edge. In additions alternatives will be of great help especially in times of emergency and when a particular system fails. G. What makes dot net compatible to ERP? What are the probable reasons for preferring Dot net platform? They are supposed to be cheap when compared with J2EE. Since dot net is from Microsoft Corporation it goes without saying that they will be designed on application basis. Besides there are some other features that make dot net compatible to ERP. Some of the issues deserving attention are as follows: 1) Effectively checks for mistakes and procedural errors: Be it enterprise resource planning or software programs the chances of errors are often very high. These errors not only undermine the working of the software but also disturb the whole process in the company which even makes them to sigh " If only computers were not invited". No platform (including dot net and J2EE) can be guaranteed for controlling errors. However the unique codes set in dot net helps in knowing as and when errors occur and that too at the early stage. This is an important issue especially for ERP the functioning of which has to be constantly monitored to ensure accountability and performance. Since the platform immediately points out what is going wrong it is easier for the company to freeze the whole process rectify the errors and set back things to motion properly. 2) Flexibility to function simultaneously: Any platform or application poses a major challenge to the company. The more they are different and the larger are they in numbers the greater is the trouble for the company. The reason is that the simultaneous function leads to lot of collusions and confusions. Even though they work separately the functioning of one will clash with that of the other. This case could be worse if one of them is dependent on the other for one or more supportive function. Dot net has been designed to overcome this drawback. Dot net will neither have interference nor an influence with other applications/platforms unless they are desired in the course of work. Even in such cases dot net will only act as an inducer. This explains the reason for preferring dot net platform.
3) Message persistency: The frequency of message transfer is referred as message persistency. This rate is important in deciding the success of the communication systems and ERP functions. The organization should ensure that there is no slag in the systems or process or any other procedures that is likely to affect this rate. Companies should constantly monitor and suggest areas for
3) Safety and lesser risk: Computers and software programs become vulnerable to hacking and all sort of mishaps and no amount of technological advancements have helped to resolve this trouble. The security features of dot net platform are not only stringent but also a tough thing for the user himself. Unlike from other software programs which award
Velammal College of Engineering and Technology, Madurai
Page 243
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
licenses directly to the user the special facility is that they can have the liberty to award the license. 4) Communication, costs and other services: Dot Net software offers these facilities at an unbelievable cost that is predicted as low as 5% when compared with J2EE platforms. This is good news to companies and especially those who are unable to make use of the latest IT facilities due to budgetary constraints. This would have otherwise been a handicap to companies for not being able to serve their own requirements. Dot net also facilitates faster communication and renders many additional services to companies availing them. Since they are one among the latest the vendors take pains in ensuring that the companies understand the entire system and still continue preferring dot net platform.
Companies are now focusing on optimizing processes within those functions. Aligning employee roles and responsibilities with a process, as well as function, can reap big benefits by giving employees visibility to the process from end to end. Having a common picture of how a process works "stimulates tremendous innovation and improvement,". "It's a new way of looking at your business." ERP has thrown open opportunities for many companies to trade with foreign counter parts in the name of outsourcing, implementation and deployment of the existing ones. It has contributed lot to the economy. Academics also boast its own share of ERP relations. It has promoted lot of employment and educational opportunities. India happens to be a key beneficiary in this aspect.
IV FUTURE MARKET TRENDS OF ENTERPRISE RESOURCE PLANNING.
V LEADING ERP GIANTS:
The dream of enterprise resource planning systems is that a single application can track and monitor all of a business’s functions. In a perfect world, a manager opens a single ERP app to find data about any aspect of the business, from financials to HR to distribution schedules. Also, we’re not there yet – or at least most companies aren’t. Moreover, there are still a lot of gaps in ERP systems, particularly in industries where ERP functionality has grown up from its historic origins in manufacturing. There are even gaps in core ERP areas, Hamerman tells Datamation, “Where they just haven’t done a particularly good job, in areas like budgeting, and recruitment…where the vast majority of customer uses something other than their ERP vendor.” But despite the challenges, the movement toward a global ERP system is a key factor shaping the future of enterprise resource planning Companies looking to "shift from a functional to a business process orientation" will need a very long-term plan, and possibly a therapist on hand.
This is a niche market where only the best gets to thrive. Needless to mention the money involved and other factors ensure a tug of war between the players. Some of the leading ERP giants are as follows: 1) Microsoft: This software major holds a promising segment in the Small and medium enterprises market. They have been constantly upgrading the versions and are expert in manufacturing the products from the vendor's point of view (who knows all the practical difficulties of the stakeholders). They have been taking a very liberal stand when it comes to the market and competitors so as to be more compatible to the users. More innovations will take the company to great heights in the market around the globe. The biggest advantage of this software is that many companies will prefer them for one reason that they can be easily run in Microsoft application and platforms. 2) Oracle: They have been in the limelight ever since they have purchased People soft ERP software. Another
Velammal College of Engineering and Technology, Madurai
Page 244
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
encouraging fact for them is the mass campaign carried by UNDP (United Nations Development Program) to create an awareness of their products. They have also satisfies the customers (of both people soft and oracle) by offering many competitive features. 3) SAP: A specialized ERP product meant to solve technical and managerial issues. SAP ERP continues to be the choice of many companies around the world. It helps companies to redefine their business operations. Some of the features and advantages of this software program are at par with industry standards though some others require to be improved in order to make it compatible to the end user. 4) PeopleSoft: People soft ERP software program helps the companies to streamline their HR functions in an easy and effective manner. This software continues to be the undisputed leader in the market when it comes to HR operations anywhere in the globe. Oracles recent takeover of People soft has only been a boon to its customers who are now fable to access the e-Services of Oracle also as a result of this merger. 5) UPS: This ERP software has an important feature. They can work in combination with other applications like SAP, Oracle. This will enable them to get the best features from each software and put them in use as may be demanded. This has also hogged the limelight due to this advantage.
VI CONCLUSION The "soft" benefits of ERP provide good bottom-line savings and sustainable top-line growth, lower inventory and operating cost, more efficient production, improved quality with less waste, higher customer acquisition retention and better cash flow. The future of ERP holds an undisputed demand not only in the national level but also at the global level. If the technology can be improvised to the desired extent. ERP trends reflect positive signals for the ERP vendors and companies availing their service. It is important to remember the fact that both the vendor and the company will be able to make use of any advantage (including the modern facilities) only through proper coordination, teamwork and nurturing a cordial atmosphere. Mere IT ERP trends will not help in this aspect. Industry consultant Reed sums it up this way: "'Empower me. Give me the tools to create differentiating processes that allow me to define myself from my competitors. And make sure that it's easier for me to do, so I don't have to hire 100 programmers. Give me the building blocks to put that together quickly, so that it's just humming in the background, and leave me free to focus on what makes us better than other companies.' That's what customers are expecting now and really want”. REFERENCES www.erpwire.com www.whitepaper.techrepublic.com www.olcsoft.com www.eioupdate.com www.sdn.sap.com www.bscaler.com [7] www.plex.com [8] www.itmanagement.earthweb.com
[1] [2] [3] [4] [5] [6]
Velammal College of Engineering and Technology, Madurai
Page 245
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A New Semantic Similarity Metric for Handling all Relations in WordNet Ontology K.Saruladha#1, Dr.G.Aghila*2, Sajina Raj#3 #
Research Scholar, 3Post Graduate Student, Department of Computer Science & Engg. Pondicherry Engineering College, Pondicherry 605 014, India. 1
[email protected] 3 [email protected] *
Head of the department Department of Computer Science, Pondicherry University, Pondicherry 605 014, India. 2 [email protected] Abstract- Semantic similarity assessing methods play a central role in many research areas such as Psychology, cognitive science, information retrieval biomedicine and Artificial intelligence. This paper discuss the existing semantic similarity assessing methods and identify how these could be exploited to calculate accurately the semantic similarity of WordNet concepts. The semantic similarity approaches could broadly be classified into three different categories: Ontology based approaches (structural approach), information theoretic approaches (corpus based approach) and hybrid approaches. All of these similarity measures are expected to preferably adhere to certain basic properties of information. The survey revealed the following drawbacks The information theoretic measures are dependent on the corpus and the presence or absence of a concept in the corpus affects the information content metric. For the concepts not present in the corpus the value of information content tends to become zero or infinity and hence the semantic similarity measure calculated based on this metric do not reflect the actual information content of the concept. Hence in this paper we propose a new information content metric which provides a solution to the sparse data problem prevalent in corpus based approaches. The proposed measure is corpus independent and takes into consideration hyponymy and meronymy relations. Further the information content metric used earlier by Resnik, lin and Jiang and Cornath methods may produce better results with alternate corpora other than brown corpus. Hence the effect of corpus based information content metric on alternate corpora is also investigated. The proposed similarity function with noun pairs of R&G data set was tested for Resnik method and the correlation coefficient with human judgment were calculated. The results were promising.
Velammal College of Engineering and Technology, Madurai
Keywords-Ontology, similarity method, information retrieval, conceptual similarity, taxonomy, corpus based
I . Introduction The goal of Information retrieval process is to retrieve Information relevant to a given query request. The aim is to retrieve all the relevant information eliminating the nonrelevant information. An information retrieval system comprises of document representation, semantic similarity matching function and Query. Document representation comprises the abstract description of documents in the system. The semantic similarity matching function defines how to compare query requests to the stored descriptions in the representation. The percentage of relevant information we get mainly depends on the semantic similarity matching function we used. So far, there are several semantic similarity methods used which have certain limitations despite the advantages. No one method replaces all the semantic similarity methods. When a new information retrieval system is going to be build, several questions arises related to the semantic similarity matching function to be used. In the last few decades, the number of semantic similarity methods developed is high. This paper discusses the overall view of different existing similarity measuring methods used for ontology concept comparison. We also discuss about the pros and cons of existing similarity metrics. We have presented a new approach which is independent of the corpora for finding the semantic similarity between two concepts.
Page 246
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Section II, discusses the basic qualitative properties that a similarity measure is expected to satisfy. Section III discusses various approaches used for similarity computation and the limitations of those methods. In Section IV, comparison among different semantic similarity measures are discussed. In Section V we introduce a new semantic similarity approach and an algorithm for finding the similarity between all the relations in the WordNet taxonomy. The results based on new similarity measure is promising.
II. ONTOLOGY SIMILARITY In this section, a set of intuitive and qualitative properties that a similarity method should adhere to is discussed.[20] A. Basic Properties Any similarity measure must be compatible with the basic properties as they express the exact notion of property. o CommonalityProperty o Difference Property o Identity Property B. Retrieval Specific Properties The similarity measure cannot be symmetrical in case of ontology based information retrieval context. The similarity is directly proportional to specialization and inversely proportional to generalization. o Generalization Property C. Structure Specific Properties The distance represented by an edge should be reduced with an increasing depth. o Depth Property o Multiple Paths Property III. APPROACHES USED FOR SIMILARITY COMPUTATION In this section, we discuss about various similarity methods[20]. The similarity methods are A. B. C. D.
Path Length Approaches Depth-Relative Approaches Corpus-based Approaches Multiple-Paths Approaches
A. Path Length Approach The shortest path length and the weighted shortest path are the two taxonomy based approaches for measuring similarity through inclusion relation. 1) Shortest Path Length
Velammal College of Engineering and Technology, Madurai
A simple way to measure semantic similarity in a taxonomy is to evaluate the distance between the nodes corresponding to the items being compared. The shorter distance results in high similarity. In Rada et al. [1989][1][14], shortest path length approach is followed assuming that the number of edges between terms in a taxonomy is a measure of conceptual distance between concepts. distRada(ci; cj) = minimal number of edges in a path from ci to cj This method yields good results. since the paths are restricted to ISA relation, the path lengths corresponds to conceptual distance. Moreover, the experiment has been conducted for specific domain ensuring the hierarchical homogeneity. The drawback with this approach is that, it is compatible only with commonality and difference properties and not with identity property. 2) Weighted Shortest Path Length This is another simple edge-counting approach. In this method, weights are assigned to edges. In brief, weighted shortest path measure is a generalization of the shortest path length. Obviously it supports commonality and difference properties. - Similarity of immediate specialisation -Similarity of immediate generalisation P=(p1,…..,pn) where, Pi ISA pi+1 or Pi+1 ISA pi For each I with x=p1 and y=pn Given a path P=(p1,…..pn), set s(P) to the number of specializations and g(P) to the number of generalizations along the path P as follows: s(P)= |{i\pi ISA pi+1}| (1) g(P)=|{i|Pi+1 ISA pi}| (2) If p1,……pm are all paths connecting x and y, then the degree to which y is similar to x can be defined as follows: simWSP(x,y)=max{ s(pj) s(pj)} (3) j=1,….m The similarity between two concepts x and y, sim(x,y) WSP(weighted Shortest Path) is calculated as the maximal product of weights along the paths between x and y. Similarity can be derived as the products of weights on the paths. s(pj)
g(Pj
j
j
s(P ) and g(P ) = 0 Hence the weighted shortest path length overcomes the limitations of shortest path length wherein the measure is based on generalization property and achieves identity property. B. Depth-Relative Approaches Even though the edge counting method is simple, it limits the representation of uniform distances on the edges in the
Page 247
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
taxonomy. This approach supports structure specific property as the distance represented by an edge should be reduced with an increasing depth. This approach includes depth relative scaling method, conceptual similarity method and normalization based length method. 1) Depth Relative Scaling In his depth-relative scaling approach Sussna[1993][2] defines two edges representing inverse relations for each edge in a taxonomy. The weight attached to each relation r is a value in the range [minr; maxr]. The point in the range for a relation r from concept c1 to c2 depends on the number nr of edges of the same type, leaving c1, which is denoted as the type specifc fanout factor: W(c1→r c2)=maxr-{maxr--minr/nr(c1)} The two inverse weights are averaged and scaled by depth d of the edge in the overall taxonomy. The distance between adjacent nodes c1 and c2 are computed as: dist sussna(c1,c2)=w(c1→ r c2)+ (c1→ r’ c2)/2d (4) where r is the relation that holds between c1 and c2, and r’ is its inverse. The semantic distance between two arbitrary concepts c1 and c2 is computed as the sum of distances between the pairs of adjacent concepts along the shortest path connecting c1 and c2. 2) Conceptual Similarity Wu and Palmer [1994][3], propose a measure of semantic similarity on the semantic representation of verbs in computer systems and its impact on lexical selection problems in machine translation. Wu and Palmer define conceptual similarity between a pair of concepts c1 and c2 as: Sim wu&palmer(c1,c2)= (5) Where N1 is the number of nodes on the path from c1 to a concept c3. , denoting the least upper bound of both c1 and c2. N2 is the number of nodes on a path from c2 to c3. N3 is the number of nodes from c3 to the most general concept. 3) Normalised Path Length Leacock and Chodorow [1998][4], proposed an approach for measuring semantic similarity as the shortest path using is a hierarchies for nouns in WordNet. This measure determines the semantic similarity between two synsets (concepts) by finding the shortest path and by scaling using the depth of the taxonomy: Sim Leacock&Chaodorow(c1,c2)= -log(Np(c1,c2)/2D) (6) Np (c1,c2) denotes the shortest path between the synsets (measured in nodes), and D is the maximum depth of the taxonomy. C. Corpus-based Approach
Velammal College of Engineering and Technology, Madurai
The knowledge disclosed by the corpus analysis is used to intensify the information already present in the ontologies or taxonomies. In this method, presents three approaches viz information content method, J&C method and Lin’s universal similarity measure is discussed. All these methods incorporate corpus analysis as an additional, and qualitatively different knowledge source. 1) Information Content In this method rather than counting edges in the shortest path, they select the maximum information content of the least upper bound between two concepts. Resnik [1999] [5], argued that a widely acknowledged problem with edgecounting approaches was that they typically rely on the notion that edges represent uniform distances. According to Resnik's measure, information content, uses knowledge from a corpus about the use of senses to express nonuniform distances. Let C denote the set of concepts in a taxonomy that permits multiple inheritance and associates with each concept c 2 C, the probability p(c) of encountering an instance of concept c. For a pair of concepts c1 and c2, their similarity can be defined as: SimResnik (7) where, S(c1,c2): Set of least upper bounds in the taxonomy of c1 and c2 p(c) :Monotonically non-decreasing as one moves up in the taxonomy, p(c1) ≤ p(c2), if c1 is a c2. The similarity between the two words w1 and w2 can be computed as: wsimResnik (8) Where, s(wi): Set of possible senses for the word wi. Resnik describes an implementation based on information content using WordNet's [Miller, 1990][6], taxonomy of noun concepts [1999]. The information content of each concept is calculated using noun frequencies Freq(c) = where, words(c): Set of words whose senses are subsumed by concept c. (c)=freq(c)/N where , N: is the total number of nouns. The major drawback of the information content approach is that they fail to comply with the generalization property, due to symmetry. 2) Jiang and Conrath's Approach(Hybrid Method) Jiang and Conrath [1997][7]proposed a method to synthesize edge-counting methods and information content
Page 248
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
into a combined model by adding the information content as a corrective factor. The edge weight between a child concept cc and a parent concept cp can be calculated by considering factors such as local density in the taxonomy, node depth, and link type as, Wt(cc,cp)= (9) Where, d(cp) : Depth of the concept cp in the taxonomy, E(cp) :Number of children of cp (the local density) (¹E) : Average density in the entire taxonomy, LS(cc, cp) : Strength of the edge between cc and cp, T(cc,cp) : Edge relation/type factor The parameters 0 and ,0 1 control the in°uence of concept depth and density, respectively. strength of a link LS(cc; cp) between parent and child concepts is proportional to the conditional probability p(cc|cp) of encountering an instance of the child concept, cc, given an instance of the parent concept, cp: LS(cc, cp) = -log p(cc|cp) Resnik assigned probability to the concepts as p(cc \cp) = p(cc), because any instance of a child concept cc is also an instance of the parent concept cp. Then: p(Cc|Cp) = p(Cc|Cp) = If IC(c) denotes the information content of concept c. then: LS(Cc,Cp) = IC(Cc)-IC(Cp) Jiang and Conrath then defined the semantic distance between two nodes as the summation of edge weights along the shortest path between them [J. Jiang, 1997]: distjiang&conrath(C1,C2)= (10) Where, path(c1, c2) : the set of all nodes along the shortest path between c1 and c2 parent(c) : is the parent node of c LSuper(c1, c2) : is the lowest superordinate (least upper bound) on the path between c1 and c2.. Jiang and Conrath's approach made information content compatible with the basic properties and the depth property, but not to the generalization property. 3) Lin's Universal Similarity Measure Lin [1997; 1998][8][13] defines a measure of similarity claimed to be both universally applicable to arbitrary objects and theoretically justified. He achieved generality from a set of assumptions. Lin's information-theoretic definition of similarity builds on three basic properties, commonality, difference and identity. In addition to these properties he assumed that
Velammal College of Engineering and Technology, Madurai
the commonality between A and B is measured by the amount of information contained in the proposition that states the commonalities between them, formally: I(common(A,B)) = -log p(common(A,B) where, I(s): Negative logarithm of the probability of the proposition, as described by Shannon[1949]. The difference between A and B is measured by: I(description(A,B)) – I(common(A,B)) Where, description(A;B) : Proposition about what A and B are. Lin proved that the similarity between A and B is measured by, simLin(A,B)= (11) The ratio between the amount of information needed to state the commonality of A and B and the information needed to describe them fully. Lin’s similarity between two concepts in a taxonomy ensures that: SimLin(c1,c2)= (12) where, LUB(c1, c2): Least upper bound of c1 and c2 p(x) : Estimated based on statistics from a sense tagged corpus. This approach comply with the set of basic properties and the depth property, but would fail to comply with the generalization property as it is symmetric. D. Multiple-Paths Approaches This approach solves the problem with single path approach. Single path as a measure for the similarity, fails to truly express similarity whenever the ontologies allow multiple inheritance. In multiple-path approach measurement is made by taking into account all the semantic relations in ontologies, considering more than one path between concepts. Attributes should influence the measure of similarity, thus allowing two concepts sharing the same attribute to be considered as more similar, compared to concepts not having this particular attribute. This approach includes Hirst & St-Onge method, generalized shortest path method, shared nodes and weighted shared node methods. 1) Medium-Strong Relations Hirst and St-Onge [Hirst and St-Onge, 1998; St-Onge, 1995] [9][15], distinguishes the nouns in the Wordnet as extra-strong, strong and medium-strong relations. The extra-strong relation is only between a word and its literal repetition. A strong relation between two words exists if: 1. They have a synset in common
Page 249
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2. There is a horizontal link (antonymy, similarity) between a synset of each word 3. There is any kind of link at all between a synset of each word or if one word is a compound or phrase that includes the other word. Medium-strong exists between words when a member of a set of allowable paths connects the two words. A path is allowable if it contains no more than five edges and conforms to one of the eight patterns described in Hirst and St-Onge [1998]. 2) Generalised Weighted Shortest Path The principle of weighted path similarity can be generalized by introducing similarity factors for the semantic relations. However, there does not seem to be an obvious way to differentiate based on direction. Thus, we can generalize simply by introducing a single similarity factor and simplify to bidirectional edges. This method solves the symmetry problem by introducing weighted edges. 3) Shared Nodes This approach overcomes the limitation of single path length approach. Multiple paths are considered for measuring the similarity. The shared nodes approach with similarity function discussed above complies with all the defined properties. 4) Weighted Shared Nodes Similarity It is found that when deriving similarity using the notion of shared nodes, not all nodes are equally important. Assigning weights to edges is very important, as it generalizes the measure so that it can be make use for different domains with different semantic relations. It also allows differentiating between the key ordering relation, ISA and the other semantic relations when calculating similarity. The weighted shared nodes measure complies with all the defined properties. IV. COMPARISON OF DIFFERENT SIMILARITY MEASURES In this section we discuss about the results of comparison of the measures to human similarity judgments. The first human similarity judgment was done by Rubinstein and Goodenough [1965][11], using two groups totaling 51 subjects to perform synonymy judgments on 65 pairs of nouns and this in turn been the basis of the comparison of similarity measures.
Velammal College of Engineering and Technology, Madurai
Miller and Charles [1991][12] repeated Rubinstein and Goodenough's original experiment, they used a subset of 30 noun pairs from the original list of 65 pairs, where ten pairs were from the high level of synonymy, ten from the middle level and ten from the low level. TABLE I REPLICA OF THE RUBINSTEIN AND GOODENOUGH AND THE MILLER AND CHARLES EXPERIMENTS
Word1
Word2
Replica
R&G
M&C
Car
Automobile
3.82
3.92
Gem
Jewel
3.86
3.84
3.92 3.94
Journey
Voyage
3.58
3.54
3.58
Boy
Lad
3.10
3.76
3.84
Coast
Shore
3.38
3.70
3.60
Asylum
madhouse
2.14
3.61
3.04
Magician
Wizard
3.68
3.50
3.21
Midday
Noon
3.45
3.42
3.94
Furnace
Stove
2.60
3.11
3.11
Food
Fruit
2.87
3.08
2.69
Bird
Cock
2.62
3.05
2.63
Bird
Crane
2.08
2.97
2.63
Tool
implement
1.70
2.95
3.66
Brother
Monk
2.38
2.82
2.74
Lad
Brother
1.39
1.66
2.41
Crane
Implement
1.26
1.68
2.37
Journey
Car
1.05
1.16
1.55
Monk
Oracle
0.90
1.10
0.91
Cemetery
Woodland
0.32
0.95
1.18
Food
Rooster
1.18
0.89
Coast
Hill
1.24
0.87
1.09 1.26
Forest
Graveyard
0.41
0.84
Shore
Woodland
0.81
0.63
0.90
Monk
Slave
0.36
0.55
0.57
Coast
Forest
0.70
0.42
0.85
Lad
Wizard
0.61
0.42
0.99
Chord
Smile
0.15
0.13
0.02
Glass
Magician
0.52
0.11
0.44
Rooster
Voyage
0.02
0.08
0.04
Noon
String
0.02
0.08
0.04
1.00
Page 250
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The correlation between these two experiments is 0.97. An experiment were performed by taking the replica of the Miller and Charles experiment that included 30 concept pairs and an additional ten new compound concepts pairs and human judgment for all the 40 pairs has been performed. The correlations for the measures done by Resnik, Jiang and Conrath; Lin, Hirst and St-Onge; and Leacock and Chodorow are shown in table given below. [Budanitsky, 2001][18][19][20] TABLE II CORRELATION BETWEEN DIFFERENT SIMILARITY MEASURES & HUMAN SIMILARITY JUDGMENTS FROM THE MILLER AND CHARLES EXPERIMENT
The collection of more than 1 million words was manually tagged with about 80 parts of speech. The presently available list of WordNet concepts tagged in the brown corpus includes approximately 420000 words. There are more number of concepts in WordNet which are not tagged in Brown Corpus. We ran a small experiment on the brown corpus and found that the words soccer, fruitcake, world trade centre, CPU etc are not found in the Brown corpus. Information content of a concept in corpus based approaches is calculated using the formula IC(C)=-log p(c) For the nouns for which p(c) value is zero the IC value becomes zero or infinity. This issue is addressed as sparse data problem. A concept (noun) other than non informative words(stop words) is expected to have information content value>0. But the existing IC calculation method returns a zero for a concept(noun) which is unacceptable.
Approach
Correlation
Resnik
0.744
Jiang and Conrath
0.850
Lin
0.829
Hirst and St-Onge
0.744
BROWN CORPUS CATEGORIES Press : reportage (Political Sports,Society,Spot
Leacock and Chodorow
0.816
News,Financial,Cultural
TABLE IV
Press: Editorial (Institutional Daily, Personal, Letters to the
The table given below shows the correlations between the replica and the two previous experiments. The correlations between the replica experiment and the previous experiments are fairly good[20]. TABLE III CORRELATION BETWEEN THE THREE HUMAN SIMILARITY JUDGMENT EXPERIMENTS Correlation Rubinstein Goodenough Rubinstein & Goodenough Miller & Charles
Editor) Press: Reviews (Theatre, Books, Music, Dance) Religion (Books, Periodicals, Tracts) Skill and Hobbies (Books, Periodicals) Popular Lore (Books, Periodicals) Belles-Lettres (Books, Periodicals)
Miller & Charles
0.97
Replica
0.93
Replica
0.95
Miscellaneous: US Government & House Organs (Government Documents, Foundation Reports, College Catalog, Industry House organ) Learned (Natural Sciences, Medicine, Mathematics, Social and Behavioral Sciences, Political Science, Law, Education,
A. Sparse Data Problem in Corpus based Approach Following the standard definition from Shannon and Weaver’s information theory [1949], the information content of c is −log p(c). Information content in the context of WordNet is drawn from the brown university standard corpus. This corpus refers to a collection of documents of widely varying genres collected in 1961 which was updated in 1971and 1979 to reflect new literatures shown in table IV.
Humanities, Technology and Engineering) Fiction: General (Novels, Short Stories) Fiction: Mystery and Detective Fiction (Novels, Short Stories) Fiction: Adventure and Western (Novels, Short Stories) Fiction: Romance and Love Story (Novels, Short Stories) Humor (Novels, Essays)
B. Significance of the work We have proposed an algorithm based on new information content which takes into consideration the meronymy
Velammal College of Engineering and Technology, Madurai
Page 251
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
relations. The Information content metric proposed by Pierro [2009] takes into concern the holonym relations that a concept has with other concepts. All of the concepts of the WordNet will not have all of the relation types. Some of the concepts will have holonym relation and few with meronym relations etc. The information content metric which considers holonymy relation alone, if it ignores the other type of relations and if it produces a zero value, it means that the concept has no information which is not correct. Hence we decided to consider both the meronym and holonym relations. If a concept do not have holonymy relation then the information imparted by meronomy is taken into consideration and the information content will become zero when the concept has no relations with other concepts which is a very rare case.
2. To investigate the existing IC metric on corpora other than brown corpus. TABLE V WORDNET MERONYMY/HOLONYMY RELATION Relation
Example
Meronym(Part-of)
Engine is a Meronym of Car
Holonym(Has-part) Car is a Holonym of Engine
Meronym(Has-Member)
Team has member player
Holonym Player is the member of Team (Member-of)
V. PROPOSED WORK The semantic similarity measures are mostly based on the information content. Most of the corpus based similarity methods like Lin[8], Jiang Cornath[7] and Resnik[5] are IC based and the IC calculation is done using Brown corpus. All concepts of WordNet are not present in the Brown Corpus. The noun such as autograph, serf and slave are not present in the Brown Corpus. Similarity measures that rely on information content can produce a zero value for even the most intuitive pairs because the majority of WordNet concepts occur with a frequency of zero. This makes the Lin method and Jiang cornath method to return zero or infinity in the continuous domain and hence the similarity measure is not true or reliable. Hence the computation of Information content should be computed in a different way so that the similarity measure becomes reliable.
A. Proposed Corpora Independent Measure to Solve Sparse Data Problem
Similarity
The objective of this work is twofold. 1. To design a new information metric which solves sparse data problem which is corpus independent and is based on semantic information available in well known information source like WordNet.
Velammal College of Engineering and Technology, Madurai
The Existing Similarity methods do not consider the Holonymy/Meronymy relationships defined in Wordnet. Hence we propose to devise a new similarity measure which considers these relationships and experimentally evaluate and compare it with the existing similarity measures using R&G data set and extended data set. The fig.1 shows architecture. Google based corpus Reuters Corpus Brown Corpus
the
semantic
Compute Frequency of Wordnet Concepts Semantic Similarity computation Resnik Jiang & Conrath Lin Pirro&Seco
similarity
system
Human Judgments for Extended R&G Data set
Compute Correlation Coefficient
WordNet Taxonomy Rubenstein & Goodenough Extended DataSet
Performance Analysis
Fig. 1 Semantic Similarity System Architecture
Page 252
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
B. Proposed IC Calculation Involving Meronym
IC (LCS (c1,c2))--hyper
The intrinsic IC for a concept c is defined as: IC(c) = 1- log(hypo(c)+mero(c)+1
---(16)
Log(maxcon) where, Hypo(c)-Function returns the number of hyponyms Mero(c)-Function returns the number of meronyms Maxcon- is a constant that indicates the total number of concepts in the considered taxonomy Note: Assume hypo(c), mero(c)>=0 & maxcon>0 The function hypo and mero returns the number of hyponyms and meronyms of a given concept c. Note that concepts representing leaves in the taxonomy will have an IC of one, since they do not have hyponyms. The value of one states that a concept is maximally expressed and cannot be further differentiated. Moreover maxcon is a constant that indicates the total number of concepts in the considered taxonomy.
(c1 c2)
----(17)
where, the function LCS finds the lowest common subsumer of the two concepts c1 and c2 and the function hyper finds all the hypernyms of c1 and c2 upto the LCS node. The Proposed IC formula will be used in existing semantic similarity methods such as Resnik, Lin and Jiang and cornath for computation of similarity measure using R&G and M&C data sets. The influence of meronymy and hyponymy relations in calculation of similarity measure will be studied. We will test the existing Resnik, Lin and Jiang cornath method using alternate corpora like reuters and Google Based Corpora. Some researches have taken this issue into consideration and have proposed corpus independent information content calculation. But those calculation takes into consideration only the holonymy relations [21] excluding other semantic relations that exists between concepts.
C. Proposed Similarity Function Based On Proposed IC According to our new formula, the similarity between the two concepts c1 and c2 can be defined as, Simext(c1,c2) =
Velammal College of Engineering and Technology, Madurai
Page 253
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Input (noun1, noun2), Output (similarity score) Step 1: For each sense c1 and c2 in noun1 and noun2 If c1 and c2 are hyponyms Calculate IC (c1) for hyponyms i.e
VI. Conclusion This paper has discussed the various approaches that could be used for finding similar concepts in an ontology and between ontologies. We have done a survey to exploit the similarity methods for ontology based query expansion to aid better retrieval effectiveness of Information retrieval models. The experiments conducted by early researches provide better correlation values which gives promising direction of using them in Ontology based retrieval models. A new semantic similarity metric has been introduced which overcomes the shortcomings of existing semantic similarity methods mainly sparse data problem. We are working with the new similarity function which will combine the advantages of the similarity methods discussed in this paper and we will test it with ontologies of particular domain. Since we are considering all the relationship present in the WordNet taxonomy, the new metric gives accurate results.
IC(c) = 1- log(hypo(c)+1) log(maxcon) go to step 3 Else if c1 and c2 are meronyms Calculate IC(c1) for meronyms i.e IC(c) = 1- log(mero(c)+1) log(maxcon) go to step 3 Step 2: For IC value for both hyponyms and meronyms using the proposed IC formula,
VII. REFERENCES IC(c) = 1- log(hypo(c)+mero(c)+1) [1] Roy Rada, H. Mili, Ellen Bicknell, and M. Blettner. “Development and application of a metric on semantic nets” IEEE Transactions on Systems, Man, and Cybernetics, 19(1), 17{30}, January 1989. [2] Michael Sussna: “Word sense disambiguation for tree-text indexing using a massive semantic network” In Bharat Bhargava [3]
Zhibiao Wu and Martha Palmer. “Verbs semantics and lexical
selection”, In Proceedings of the 32nd annual meeting on Association for Computational Linguistics”, Association for Computational Linguistics, 1994. pages 133{138, Morristown, NJ, USA [4] Claudia Leacock and Martin Chodorow: “Combining local context and wordnet similarity for word sense identification”. In Christiane Fellbaum, editor, WordNet: An Electronic Lexical Database. MIT Press, 1998. [5] Philip Resnik.: “Semantic similarity in a taxonomy: An information-based measure and its application to problems of ambiguity in natural language”, 1999. [6] George A. Miller, Richard Beckwith, Christiane Fellbaum, Derek Gross, and K.J. Miller.[Resnik, 1995] Philip Resnik. “Using information content to evaluate semantic similarity in a taxonomy”. In IJCAI, pages 448{453, 1995. [7] D. Conrath J. Jiang.: “Semantic similarity based on corpusstatistics and lexical taxonomy”. In Proceedings on International Conference on Research in Computational Taiwan, pages 19{33, 1997.
Linguistics,
Velammal College of Engineering and Technology, Madurai
log(maxcon) go to step 3 Step 3: Call the existing semantic similarity function Simres(c1,c2), SimLin(c1,c2), SimJ&C(c1,c2) And then go to step 4. Step 4: Call the proposed semantic similarity function for the given concepts c1 & c2 Simext(c1,c2) = IC(LCS(c1,c2))hyper(c1 c2) Step 5: Collect human judgments and save it as a separate table for the R&G and M&C data sets Step 6: Calculate the correlation coefficients between results of the similarity measures and human judgments Step 7: Compare the similarity measures for R&G data set using the proposed IC and proposed similarity existing similarity measures.
Page 254
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[8]
Dekang
Lin.:
“An
information-theoretic
definition
of
A
similarity”. In Jude W. Shavlik, editor, ICML, pages 296{304. Morgan Kaufmann, 1998. ISBN 1-55860-556-8. [9]
Graeme
Hirst and
David St-Onge: “Lexical chains as detection and correction
representation of context for the
malapropisms”In Christiane Fellbaum, editor, WordNet: An Electronic Lexical Database. MIT Press, 1998. [10] Troels Andreasen, Rasmus Knappe, and Henrik Bulskov: ” Domain-specifiic similarity and retrieval”. In Yingming Liu, GuoqingChen, and Mingsheng Ying, editors, 11th International FuzzySystems Association World Congress, volume 1, pages 496{502,Beijing,China, 2005. IFSA 2005, Tsinghua University Press. [11] Rubinstein and Goodenough, H. Rubinstein and J. B. a Goodenough: “Contextual correlates of synonymy”. Communications of the ACM, 8(10), 1965. [12] George A. Miller and W. G. Charles: “Contextual correlates of semantic similarity”. Language and Cognitive Processes, 6(1):1{28, 1991. [13] Dekang Lin: “Using syntactic dependency as local context to resolve word sense ambiguity”. In ACL, pages 64{71, 1997. [14] Roy Rada and Ellen Bicknell: “Ranking documents with a thesaurus”. JASIS, 40(5):304{310,1989.Timothy Finin, and Yelena Yesha, editors, Proceedings of the 2nd International Conference on Information and Knowledge Management, pages 67{74,New York, NY, USA, November 1993. ACM Press. [15] Alexander Budanitsky: “University of Toronto Graeme Hirst, University of Toronto Evaluating WordNet-based Measus Of Lexical Semantic Relatedness”, Association for Computational Linguistics,2006 [16] Philip Resnik: “Using Information Content to Evaluate Semantic Similarity in a Taxonomy”, Sun Microsystems Laboratories, 1995 [17]
Henrik Bulskov Styltsvig: “Ontology-based Information Retrieval Computer Science Section Roskilde University”, 1996
[18] A. Budanitsky. “Lexical semantic relatedness and its application in natural language processing”, 1999. [19] A. Budanitsky. Semantic distance in wordnet: An experimental, application-oriented evaluation of ¯ve measures, 2001. [20] H. Bulskov: “Ontology-based Information Retrieval”, PhD Thesis Andreasen, T., Bulskov, H., & Knappe, R. (2006). [21] N. Seco, T. Veale, J. Hayes: “ An intrinsic information content metric for semantic similarity in WordNet, in : Proceedings of ECAI, 2004, pp. 1089–1090. [22] Giuseppe Pirró: “A semantic similarity metric combining features and intrinsic information content”, Data & Knowledge Engineering 68(2009) 1289–1308 2009.
Velammal College of Engineering and Technology, Madurai
Page 255
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fault Prediction Using Conceptual Cohesion in Object Oriented System V.Lakshmi#1,P.V.Eswaripriya*2 ,C.Kiruthika*3 and M.Shanmugapriya*4 #1
Faculty of Kamaraj College of Engg and Technology, Virudhunagar, Tamilnadu. 1
*2
[email protected]
. Final Year CSE,Kamaraj College of Engg and Technology, Virudhunagar Tamilnadu. 2
*3
[email protected]
. Final Year CSE,Kamaraj College of Engg and Technology, Virudhunagar Tamilnadu *4 . Final Year CSE,Kamaraj College of Engg and Technology, Virudhunagar Tamilnadu Abstract-High cohesion is desirable property in software systems to achieve reusability and maintainability. In this project we are measures for cohesion in Object-Oriented (OO) software reflect particular interpretations of cohesion and capture different aspects of it. In existing approaches the cohesion is calculate from the structural information for example method attributes and references. In conceptual cohesion of classes, i.e. in our project we are calculating the unstructured information from the source code such as comments and identifiers. Unstructured information is embedded in the source code. To retrieve the unstructured information from the source code Latent Semantic Indexing is used. A large case study on three open source software systems is presented which compares the new measure with an extensive set of existing metrics and uses them to construct models that predict software faults. In our project we are achieving the high cohesion and we are predicting the fault in Object Oriented Systems. Keywords- Software cohesion, textual coherence, prediction, fault proneness,, Latent Semantic indexing.
fault
I.INTRODUCTION Software modularization, Object oriented decomposition is an approach for improving the organization and comprehension of source code. In order to understand OO software, software engineers need to create a well connected representation of the classes that make up the system. Each class must be understood individually and, then, relationships among classes as well. One of the goals of the OO analysis and design is to create a system where classes have high cohesion and there is low coupling among them. These class properties facilitate comprehension, testing, reusability, maintainability, etc. A.Software Cohesion Software cohesion can be defined as a measure of the degree to which elements of a module belong together. Cohesion is also regarded from a conceptual point of view. In this view, a cohesive module is a crisp abstraction of a concept or feature from the problem domain, usually described in the requirements or specifications. Such
Velammal College of Engineering and Technology, Madurai
definitions, although very intuitive, are quite vague and make cohesion measurement a difficult task, leaving too much room for interpretation. In OO software systems, cohesion is usually measured at the class level and many different OO cohesion metrics have been proposed which try capturing different aspects of cohesion or reflect a particular interpretation of cohesion. Proposals of measures and metrics for cohesion abound in the literature as software cohesion metrics proved to be useful in different tasks, including the assessment of design quality, productivity, design, and reuse effort, prediction of software quality, fault prediction modularization of software and identification of reusable of components. B.Approaches for Cohesion Most approaches to cohesion measurement have automation as one of their goals as it is impractical to manually measure the cohesion of classes in large systems. The tradeoff is that such measures deal with information that can be automatically extracted from software and analyzed by automated tools and ignore less structured but rich information from the software (for example, textual information). Cohesion is usually measured on structural information extracted solely from the source code (for example, attribute references in methods and method calls) that captures the degree to which the elements of a class belong together from a structural point of view. These measures give information about the way a class is built and how its instances work together to address the goals of their design. The principle behind this class of metrics is to measure the coupling between the methods of a class. Thus, they give no clues as to whether the class is cohesive from a conceptual point of view (for example, whether a class implements one or more domain concepts) nor do they give an indication about the readability and comprehensibility of the source code. Although other types of metrics were proposed by researchers to capture
Page 256
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
different aspects of cohesion, only a few metrics address the conceptual and textual aspects of cohesion. We propose a new measure for class cohesion, named the Conceptual Cohesion of Classes (C3), which captures the conceptual aspects of class cohesion, as it measures how strongly the methods of a class relate to each other conceptually. The conceptual relation between methods is based on the principle of textual coherence. We interpret the implementation of methods as elements of discourse. There are many aspects of a discourse that contribute to coherence, including co reference, causal relationships, connectives, and signals. The source code is far from a natural language and many aspects of natural language discourse do not exist in the source code or need to be redefined. The rules of discourse are also different from the natural language.
• Lack of cohesion implies classes should probably be split into two or more sub/classes. • Any measure of disparateness of methods helps identify flaws in the design of classes. • Low cohesion increases complexity, thereby increasing the likelihood of errors during the development process.
C3 is based on the analysis of textual information in the source code, expressed in comments and identifiers. Once again, this part of the source code, although closer to natural language, is still different from it. Thus, using classic natural language processing methods, such as propositional analysis, is impractical or unfeasible. Hence, we use an Information Retrieval (IR) technique, namely, Latent Semantic Indexing (LSI), to extract, represent, and analyze the textual information from the source code. Our measure of cohesion can be interpreted as a measure of the textual coherence of a class within the context of the entire system. Cohesion ultimately affects the comprehensibility of source code. For the source code to be easy to understand, it has to have a clear implementation logic (that is, design) and it has to be easy to read (that is, good language use). These two properties are captured by the structural and conceptual cohesion metrics, respectively.
III.PROPOSED C3
II.RELATED WORKS A.STRUCTURAL METRICS We study the cohesion metric LCOM B.Henderson-Sellers [2] of object oriented programs. The LCOM metric is based on the assumption, that if methods and instance variables of a class are interconnected, then the class is cohesive. If a method x uses instance variable y, then x and y are interconnected. Henderson-Sellers defined the metric LCOM* that is similar to LCOM, but has a fixed scale. Briand.et al observed that the scale of LCOM* is from 0 to 2, and gave a refined version of this metric with scale from 0 to 1. Other similar cohesion metrics are LCC and TCC, and CBMC. Good surveys of cohesion metrics have been made by Briand et al and Chae et al . • Cohesiveness of methods within a class is desirable, since it promotes encapsulation.
Velammal College of Engineering and Technology, Madurai
B.Formula for Lcom5 lcom=((((1/a)*Mu)-m)/deno); Mu-count for fields, m-methods length, a-fields Length, deno=1-m; lcom51=lcom*lcom; lcom5=Math.sqrt (lcom51);
LSI [4] is a corpus-based statistical method for inducing and representing aspects of the meanings of words and passages (of the natural language) reflective of their usage in large bodies of text. LSI is based on a Vector Space Model (VSM) as it generates a real-valued vector description for documents of text. Results have shown that LSI captures significant portions of the meaning not only of individual words but also of whole passages, such as sentences, paragraphs, and short essays. The central concept of LSI is that the information about the contexts in which a particular word appears or does not appear provides a set of mutual constraints that determines the similarity of meaning of sets of words to each other. LSI was originally developed in the context of IR as a way of overcoming problems with polysemy and synonymy that occurred with VSM approaches. Some words appear in the same contexts and an important part of word usage patterns is blurred by accidental and inessential information. The method used by LSI to capture the essential semantic information is dimension reduction, selecting the most important dimensions from a co-occurrence matrix (words by context) decomposed using singular value decomposition (SVD) . As a result, LSI offers a way of assessing semantic similarity between any two samples of text in an automatic unsupervised way. LSI relies on an SVD of a matrix (word _ context) derived from a corpus of natural text that pertains to knowledge in the particular domain of interest. According to the mathematical formulation of LSI, the term combinations that occur less frequently in the given document collection tend to be precluded from the LSI subspace. LSI does “noise reduction,” as less frequently co-occurring terms are less mutually related and, therefore, less sensible. Similarly, the most frequent terms are also eliminated from the analysis. The formalism behind SVD is rather complex and too lengthy to be presented here.
Page 257
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Once the documents are represented in the LSI subspace, the user can compute similarity measures between documents by the cosine between their corresponding vectors or by their length. These measures can be used for clustering similar documents together to identify “concepts” and “topics” in the corpus. This type of usage is typical for text analysis tasks.
This metric provide as an accurate calculation of cohesion of classes by using c3.Figure 4.2 reveals the measurement of cohesion using C3.
A.Advantages 1. We can predict the Cohesion. 2. We can predict the particular system is cohesive or not. B.Formula for C3 For every class C, the conceptual similarity between the methods mk and mj. C3=Mk*Mj /[Mk]2*[Mj]2 K=1, 2, 3……n; n: up to n comments j=2, 3…..n; n: up to n comments IV.COMPARISION Structural metrics are calculated from the source code such as references and data sharing between methods of a class belong together for cohesion. It define and measure relationships among the methods of a class based on the number of pairs of methods that share instance or class variables one way or another for cohesion. This metric bring only approximate value of cohesion .Figure 4.1 reveals the measurement of cohesion using Lcom5 .
Fig.4.1 PROCESS OF LCOM MEASURE
In proposed System unstructural information is retrieved from the source code like comments and identifiers. Information is retrieved from the source code using Latent Semantic Indexing.
Velammal College of Engineering and Technology, Madurai
FIG 4.2 PROCESS OF C3 MEASURE
. Here the cohesion is calculated for structured information using LCOM5 method and for unstructured information using Conceptual Cohesion of Classes .Then the results are compared and C3 method gives the accurate cohesion value. With the help of C3 and existing metrics we are achieving the high cohesion and low coupling V.CONCLUSION Classes in object-oriented systems, written in different Programming languages contain identifiers and comments which reflect concepts from the domain of the software system. This information can be used to measure the cohesion of software. To extract this information for Cohesion measurement, Latent Semantic Indexing can be used in a manner similar to measuring the coherence of natural language texts. This paper defines the conceptual cohesion of classes, which captures new and complementary dimensions of cohesion compared to a host of existing structural metrics. Principal component analysis of measurement results on three open source software systems statistically supports this fact. In addition, the combination of structural and conceptual cohesion metrics defines better models for the prediction of faults in classes than combinations of structural metrics alone. Highly cohesive classes need to have a design that ensures a strong coupling among its methods and a coherent internal description. ACKNOWLEDGMENT The authors are grateful to the management of Kamaraj College of Engineering and Technology, Virudhunagar, India for granting permission to undertake this work. Our thanks are due to the Head of the Department of Computer Science and Engineering of Kamaraj College of Engineering and Technology for allowing us the use of the laboratories and computing facilities.
Page 258
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VI.REFERENCES [1] Andrian Marcus, Denys Poshyvanyk, and Rudolf Ferenc,”Using Conceptual of Classes for Fault prediction in object oriented system”.IEEE TRANSACTIONS ON SOFTWARE ENGINEERING, VOL. 34, NO. 2, MARCH/APRIL 2008. [2] B. Henderson-Sellers, Software Metrics. Prentice Hall, 1996. [3] E.B. Allen, T.M. Khoshgoftaar, and Y. Chen, “Measuring Coupling and Cohesion of Software Modules: An Information-Theory Approach,” Proc. Seventh IEEE Int’l Software Metrics Symp.,pp. 124-134, Apr. 2001. [4] A. Marcus, A. De Lucia, J. Huffman Hayes, and D. Poshyvanyk, “Working Session: Information-Retrieval-Based Approaches in Software Evolution,” Proc. 22nd IEEE Int’l Conf. Software Maintenance, pp. 197-199, Sept. 2006. [5]P.W. Foltz, W. Kintsch, and T.K. Landauer, “The Measurement of Textual Coherence with Latent Semantic Analysis,” Discourse Processes, vol. 25, no. 2, pp. 285-307, 1998.
Velammal College of Engineering and Technology, Madurai
Page 259
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
On the Investigations of Design,Implementation, Performance and Evaluation issues of a Novel BDSIIT Stateless IPv4/IPv6 Translator 1
Hanumanthappa.J.1,Manjaiah.D.H.2,Aravinda.C.V.3 Teacher Fellow, Dos in CS, University of Mysore, Manasagangothri, Mysore, INDIA. 2 Reader &Chairman, Mangalore University, Mangalagangothri, Mangalore, INDIA. 3 M.Tech. KSOU. Manasagangotri, Mysore, INDIA. [email protected] [email protected] [email protected]
Abstract. Today, the Internet consists of native IPv4 (IPv4-only), native IPv6 and both IPv4/IPv6 dual networks. Currently both IPv4 and IPv6 existing protocols are considered as incompatible protocols. Unfortunately IPv4 and IPv6 are incompatible protocols when both the versions are available and the users of internet want to connect without any restrictions, a transition mechanism is required. Since a huge amount of resources have been invested on current IPv4 based Internet, how to smoothly transit the Internet from IPv4 to IPv6 is also a great tremendous and interesting research topic. During the time of migration from IPv4 to IPv6 networks, a number of transition mechanisms have been proposed by IETF to ensure smooth, stepwise and independent changeover. The development of Internet Protocol Version 6(IPv6), in addition to being a fundamental step to support growth of Internet is at the base of the increase IP functionality and Performance. It will enable the deployment of new applications over the Internet, opening a broad scope of technological development.BD-SIIT is one of the transitional technique which is mentioned in the IETF draft to perform the transition from IPv4/IPv6.We implement the BD-SIIT transition algorithm. The Stateless Internet Protocol/Internet Control Messaging Protocol Translation (SIIT) [RFC2765] is an IPv6 transition mechanism that allows IPv6-only hosts to talk to IPv4-only hosts. The mechanism involves a stateless mapping or bidirectional translation algorithm between IPv4 and IPv6 packet headers as well as between Internet Control Messaging Protocol version 4(ICMPv4) and ICMPv6 messages. SIIT is a stateless IP/ICMP translation, which means that the translator is able to process each conversion individually without any reference to previously translated packets. Most IP header field translations are relatively simple however; there is one issue, namely, how to translate the IP addresses between IPv4 and IPv6 packets. The NS-2 simulator is used to implement the BD-SIITmechanism
Velammal College of Engineering and Technology, Madurai
I.
INTRODUCTION.
1 .Importance of IPv6 based 4G Networks and On-going work. The IPv4/IPv6 transition process always occurs in deploying IPv6 based services crosses the IPv4 Internet. The IETF Next Generation Transition Working Group (NGtrans) has proposed many transition mechanisms to enable the seamless integration of IPv6 facilities into current networks. It also addresses the performances of various tunneling transition mechanisms used in different networks. The effect of these mechanisms on the performance of end-to-end applications is explored using metrics such as transmission latency, throughput, CPU utilization, and packet loss. This synopsis also looks into usage of Internet protocol version 6(IPv6) in one of the network architecture defined by Moby Dick research project. The Moby Dick has more straight forward approach. It takes Physical and Data link layer transmission functionalities of different access network as given exchange all existing all higher-level tasks by fully IPbased mechanisms. The Moby Dick was a three year European Union (EU) information Society Technologies (IST) project finished in December 2003.Project name Moby Dick has derived from the word “Mobility and Differentiated Services in a Future IP Network”. The main aim of the Moby Dick was to study and test IPv6 based and QoS enabled mobility architecture comprising Universal Mobile Telecommunication System (UMTS), Wireless Local Area Networks (WLAN), and wired Local Area networks (LAN) based on Ethernet. The Moby Dick project presents field evaluation results of an IP-based architecture for a 4G “True-IP” network. Moby Dick demonstrated the seamless integration of three disciplines like QoS, AAA and IP Mobility over a heterogeneous network infrastructure focusing on three access technologies (WLAN, Ethernet, and TD-CDMA).The Moby dick architecture incorporates Mobile IPv6, Fast handovers, AAA- control (Authentication, Authorization and Accounting).The migration from circuit-switched to IP
Page 260
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
based technologies and the growing role of mobility pave the way to a next generation Integrated network. The Importance of IP-based communication has already been identified in UMTS as well as in EDGE/IMT-2000 which proves an IP packet service using tunneling mechanisms but still employing access mechanisms of second generation networks.4G networks scenarios IP is used to glue all different link-level technologies, deploying technology un-aware protocols for mobility or quality of service (QoS).4G architectures should be able, thus, to embrace almost any wireless access technology available. This paper presents the field results obtained in the MobyDick architecture, an IP-based 4th Generation(4G)architecture for heterogeneous environments covering UMTS–like TD-CDMA wireless access technology, wireless and Ethernet LAN’s.This is one of the first implemented leader approaches to a 4G network thus a key contribution to this research work. The MIND (Mobile IP based network developments) project which was the follow-up of the BRAIN (Broadband radio access over IP networks) project was focused in mobility aspects, as well as ad-hoc, self-organizing, and meshed networks. The LONG project [ ] on IPv6 transition and deployment issues and Moby dick profited from some of the outcomes of that project. Nevertheless LONG didn’t aim at deploying a native IPv6 4G system, as Moby Dick did. 2. The IPv6 Transition. An IPv6 transition mechanism is a method to connect the hosts/networks using the same or different IP protocols under some specific IPv6 transition environment. The Internet protocol was started in early 1980.In the early s1990’s the Internet was growing some temporary solutions were offered in order to cover the fast growth in number of Internet user’s such as NAT, CIDR.At the Same time IETF began working to develop a new Internet Protocol namely IPv6 which was designed to be a Successor to the IPv4 Protocol. The main reason for designing this new Internet protocol(IPv6) was the need to increase the number of addresses(address spaces).The IPv6 address was designed with a 128-bit address scheme instead of 32-bit scheme in IPv4.So,the number of possible addresses in IPv6 is 3.4X1038 unique addresses.IPv6 will have enough to uniquely address every device (example Tele phone, Cell phone,mp3 player,hosts,routers,bridges etc) on the surface of earth with full end-to-end connectivity(about 32 addresses per square inch of dry land).In addition IPv6 is designed to support IPSec,Scalability,Multimedia transmissions,Security,Routing, Mobility, Real time applications like audio,video,Cryptography techniques like encryption and Decryption, Large address space, Better support for QoS,Stateless and Stateful address configuration, Enhanced support for Mobile IP and Mobile computing devices, Increased number of multicast addresses and improved support for multicast, New
Velammal College of Engineering and Technology, Madurai
protocol for neighboring node interaction, and New header format that is designed to keep header overhead to a minimum etc.Over all IPv6 was carefully thought out designed and was designed with future applications in mind. But the question now is whether it is possible for these two IPv4, IPv6 to work together smoothly and in an easy way? The answer is yes because the IETF IPng has proposed several mechanisms to be suitable for both Internet protocols to coexist. Some of the transition mechanisms work by encapsulating IPv6 packets in IPv4 packets then transporting them via an IPv4 network infrastructure and others work by operating dual IPv4/IPv6 stacks on the hosts or edge routers to allow the two versions of IPv4 and IPv6 to work together. Researchers designing new transition mechanisms in order to develop a proper mechanism, to be used in the transition from IPv4 from IPv6 and vice versa which overcomes some of the shortcomings that may appear in some proposed mechanisms. Some of the transition mechanisms work by encapsulating IPv6 packets in IPv4 packets then transporting them via an IPv4 network infrastructure and others work by operating dual IPv4/IPv6 stacks on the hosts or edge routers to allow the two versions of IPv4 and IPv6 to work together. Researchers designing new transition mechanisms in order to develop a proper mechanism, to be used in the transition from IPv4 from IPv6 and vice versa which overcomes some of the shortcomings that may appear in some proposed mechanisms. III. Related Work. 3.1. How the BD-SIIT (Bidirectional Stateless Internet Protocol/Internet control messaging Protocol Translation (SIIT) works. In this paper we proposed a new transition algorithm called BD-SIIT.This is also a new type of transition mechanism. As we Know that SIIT(Stateless Internet Protocol/Internet control messaging Protocol Translation(SIIT) is an IPv6 transition mechanism that allows IPv6 only hosts to talk to IPv4 only hosts. This mechanism contains a Stateless mapping or a bidirectional mapping (bidirectional translation algorithm) between IPv4,IPv6 packet headers as well as ICMPV6 and ICMPv4.Our new proposed BD-SIIT(new transition system) depends on the understanding of the received datagram, capturing the header, Identifying the header, Verification of the header, Transformation of the datagram to the destination environment, and then transmitting the datagram to the destination address.In,fact the proposed system deals with a Bidirectional operation that leads to convert the received datagram to the destination environment, depending on the address mapping ID value generated by DNS46 in order to be used in our proposed transition system between IPv4 and IPv6 protocols [29] [30] [31].BD-SIIT
Page 261
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
requires the assignment of an IPv4 address to the IPv6only host, and this IPv4 address is used by the host in forming a special IPv6 address that includes this IPv4 address. The mechanism which is intended to preserve IPv4 address rather than permanently assigning IPv4 addresses to the IPv6 only-hosts. The method of
assignment is one of the scope of BD-SIIT and also RFC 2765 also suggests that DHCP is basis for the Temporary IPv4 address assignment
Fig.1: BD-SIIT Translation process
The BD-SIIT has been proposed and designed in order to support the Bi-Directional communication sessions to be initiated by an IPv6 node which is located in the native IPv6 network and vice versa. The BD-SIIT Translation algorithm consists of the following two steps. 1. The First component explains V4-V6 Domain name System (DNS46) server that identifies the two public IPv4 and IPv6 addresses statistically or dynamically for each IPv4/IPv6 communicating system.
Stage-1: IPv6 Packet Transmission. Stage-2:IPv6-to-IPv4 mapping calculation and Address mapping. Stage-3:IPv6-in-IPv4 Header Translation Stage-4:IPv4 Packet Transmission. 2. The Second component explain the V4-V6 enabled gateway does the address mapping between an IPv4 and an IPv6 addresses as Well as header conversion between IPv4 and IPv6 packet headers
3.2. BD-SIIT Data Transmission from Source to Destination (end to end) Process:-
Fig.3: Data Packet Transmission from Source to Destination (End-toEnd). Fig.2: Translation of IPv6 header to IPv4 header.
The Data Packet transmission process for BD-SIIT is clearly defined in Fig.3.The BD-SIIT Translation mainly occurs due to the following four stages 1, 2, 3, and 4.
Velammal College of Engineering and Technology, Madurai
3.2. An Architecture of BD-SIIT. As mentioned in Fig.1.the BD-SIIT architecture consists of the DNS46(IPv4-IPv6 DNS),the IPv4-IPv6 enabled router, the IPv4 hostnames which are located in the IPv4 zone only, and the IPv6 host names which are located in the IPv6 zone only.
Page 262
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The BD-SIIT translation process can perform by using two techniques like (1) directly with in the end system. (2)Within a
Network-based device.
Fig.4: Network based BD-SIIT Translation process.
As shown in Fig.4.the BD-SIIT the upper layer protocols like TCP and UDP in transport layer can be passed through the translator relatively unscathed. For ex:-The BD-SIIT translation has been designed such that that TCP and UDP pseudo header checks sums is not affected by the translation process. While using BD-SIIT translation can cause some problems with various applications like FTP and embedded IP address in higher layer protocols and it requires the necessity of addition application specific application layer gateways in the translation process [29]. 3.3. Working Procedure of BD-SIIT Transition. As we know that BD-SIIT is a stateless IP/ICMP translation. The BD-SIIT translator is able to process each conversion individually, without any reference to previously translated packets. Although most of the IP header field translations are relatively very simple to handle, however one of the issue 0
79 80-Zero bits
Sl.No 1 2 3
80
related with BD-SIIT translator is how to map the IP addresses between IPv4 and IPv6 packets. A. The role of IPv6 mapped address in BD-SIIT Translator. BD-SIIT resides on an IPv6 host and converts an outgoing IPv6 headers into IPv4 headers, and incoming IPv4 headers into IPv6.To perform this operation the IPv6 host must be assigned an IPv4 address either configured on the host, or obtained via a network service left unspecified in RFC 2765.When the IPv6 hosts wants to communicate with an IPv4 host, based on DNS resolution to an IPv4 address, the BD-SIIT algorithm recognizes IPv6 address as an IPv4 mapped-address as shown in Fig.3.The one of the mechanism to translate the resolved IPv4 address into an IPv4 mapped address is provided by Bump-in-the stack(BIS)or Bump-inthe API(BIA) techniques[30].
95 FFFF(16 bits)
96
127 32 bits(IPv4 Address)
Fig.5:IPv4-mapped-IPv6 address. Table.1: Address mapping IPv6/IPv4. IPv6 Address IPv4 Address Address mapping value ABC2::4321 195.18.231.17 1 ABC2::4321 195.18.231.17 2 ABC2::4321 223.15.1.3 37
1 2`
Table.2:DNS46 Corresponding to IPv4 and IPv6. Address IPv4 Address IPv6 Address DNS mapping value 212.17.1.5 ---B 4 223.15.1.3. 1C::DACF Y 37
Sl.No
Table.3: Address mapping IPv4/IPv6. IPv4 Address IPv6 Address Address mapping value
Sl.No
1
195.18.231.17
2
210.154.76.91
3
223.15.1.3.
ABC2::4321 ABC2::4321 ABC2::4321
3.4. The BD-SIIT A Novel Transition Address-mapping algorithm for the forward operation.
Velammal College of Engineering and Technology, Madurai
1 2 37
Based on the presence of IPv4-mapped-IPv6 address as the destination IP address the BD-SIIT algorithm performs the header translation as described in Algorithm-1 and
Page 263
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Algorithm-2 to obtain an IPv4 packet for transmission via data link and Physical layers. The Fig.4.shows the Protocol .
stack
view
of
BD-SIIT
Fig.5: Protocol Stack view of BD-SIIT.
The following steps show the forward operation during the migration from IPv4/IPv6. Algorithm-1:-IPv4->IPv6: Forward operation. 1. When Host X belongs to X-zone initiates a request (Query message) to DNS46 in order to get the IP-address of Host Y which belongs to IPv6 zone. 2. When DNS46 receives a query message as a request then its checks its (Table-2) to identify whether Host Y has an IPv6 address which is unknown for the Host X.The DNS46 knows that the whole IPv6 zone has a public IPv4 address (like NAT method) i.e. 195.18.231.17 address in the destination address field of IPv4 header then forwards it via a network. 3. Simultaneously, the DNS46 sends another message to V4-V6 enabled router in order to update table-3. 4. When the Host X receives requested address of Host Y, immediately creates the IPv4 packet, inserting 195.18.231.17 address in the destination address field of IPv4 header then forwards it via network. 5. When the IPv4 packet is arrived to V4-V6 enabled router, then the router identifies a packet and verifies that a destination Address is a public address with mapping value 2 that indeed refers ABC2::4321 IPv6 address (as shown in table3).Then the V3-V4 enabled router which updates Table-1 then creates its new IPv6 packet which is based on IPv4 packet, and forwards it to Its destination in IPv6 Zone. 6. When Host Y accepts the packet then it starts to process it successfully without getting any problem. Algorithm-2:-IPv6->IPv4: Feedback operation. The following steps show the feedback operation from IPv6-IPv4 zone one which is illustrated in Figs.1. And 5.
Velammal College of Engineering and Technology, Madurai
[Note:-Consider Host X as a Client and Host Y as a Server. If the Client A sent a HTTP get command to retrieve a web page from the server Y.].[As shown in Fig.7.steps 1- 6. 1. As a response for the received command from a client X, server Y creates packet(s) then forwards them via a network to the client X using the public IPv6 zone address(ABC2::4321) as a destination address. 2. When V4-V6 enabled router receives a packet which has been sent by Server Y, then it verifies its Table-1,Table3,depending on the addressing mapping value like 37 in our scenario, it refers to 220.12.145.10 as a Sender address in Table-2 and 223.15.1.3 as a destination address in Table-2 instead of instead of 1C:: DACF and ABC2::4321 IPv6 rely. 3. After that, the V4-V6 enabled router creates a new IPv4 packet, based on the accepted IPv6 packet then forwards it to the destination (Client X). 4. When the Client X receives the IPv4 packet, its starts to process successfully without any problem. Algorithm-3:-BD-SIIT Algorithm of IPv4 header Conversion to IPv6 header conversion. 1. When BD-SIIT allows IPv6 hosts which do not contain permanently, assigned IPv4 address 2. When the IPv6 host tries to communicate with IPv4 host 3. Then BD-SIIT handles the IP address translation between IPv4 and IPv6. 4. If IPv6 address type=0: FFFF: v4 {print “IP address as an IPv4-mapped-IPv6 address”. It is mainly used to mapping IPv4 host addresses to IPv6 addresses.} 5. If the IPv6 address type=0: FFFF: 0:v4 {Print “IP address as an IPv4 translated (compatible) address.}
Page 264
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.6: BD-SIIT, IP Forward address translation, IPv6 to IPv4.
The Fig.6.shows the IP address translation process IPv6 to IPv4 with IPv4-mapped-IPv6 address and an IPv4 translated (compatible) address. The IPv6 host has obtained a temporary address-v4temp-for use in communicating with the IPv4 host. The Fig.6.very briefly illustrates the operation of IP address translation in going from the IPv6 host, using an IPv4 translated address to the IPv4 host. The translation of the remaining field is straight forward with a couple of exceptions. If there is no IPv6 fragment header, the IPv4 header fields are set as follows [31] [32]. Algorithm-4:-BD-SIIT.Algorithm for the Forward Header conversion IPv6->IPv4. If version=4 {proceed to set Internet Header Length=5,} Else if {proceed to set type of service and precedence= by default copied from the IPv6 header traffic class field},
Else if {proceed to set total length=pay load length value from IPv6 header+IPv4 header length.}, Else if {proceed to set Identification=0}, Else if {proceed to set Flags=the more fragments flag is set=0} Else if {proceed to set Don’t Fragment flag=1}, Else if {proceed to set Fragment offset=0}, Else if {proceed to set TTL=Hop limit value copied from IPv6 header - 1.}, Else if {proceed to set Protocol=next header field copied from IPv6 header}, Else if {proceed to set Header Checksum=Compute once IPv4 header has been created. Else if Source IP address=low order 32 bits of IPv6 Source address field (IPv4 translated address field) Else Destination IP address=Low order 32 bits of IPv6 Destination address field (IPv4 mapped address field)}
Fig.7: BD-SIIT, IP Reverse address translation, IPv6ÅIPv4.
Algorithm-5: BD-SIIT.Algorithm for the Reverse Header conversion IPv4->IPv6. If version=6{proceed to set version=6} Else if {proceed to set Traffic Class= IPv4 header ToS bits}, Else if {proceed to set Flow label=0}, Else if {proceed to set Pay Load Length=IPv4 header Total length value-(IPv4 header length+IPv4 options}, Else if
Velammal College of Engineering and Technology, Madurai
{proceed to set next header =IPv4 header protocol field value} Else if {proceed to set hop limit=IPv4 TTL field value-1}, Else if {proceed to set Fragment offset=0}, Else if {proceed to set source address =0:0:0:0: FFFF: /80.Concatenated with IPv4 header source IP address}, Destination IP address=0:0:0:0:0: FFFF: /96 concatenated with IPv4 header destination address.}[31][32]
Page 265
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.8: BD-SIIT hosts connection IPv6-IPv4 operation using TCP Protocol.
IV. Implementation & Performance Metrics. We implemented our proposed BD-SIIT translation system and tested it by varying the number of packets sent as well as the transmission rate for each communication system. We have tested comparison between EED(end-enddelay)of IPv4-IPV4 and IPv6-IPv6 communication session, A Comparison between the throughput(v4-to-v4 , and v6-to-v6) communication sessions, Round trip time(RTT) for each data packet of varying sizes, DNS response time, the total transmission time with TCP and UDP Protocols. In this paper we are mainly measured two important metrics like throughput and End–to End delay for BD-SIIT. Throughput: We calculated the throughput performance metric in order to identify the rate of received and processed data at the intermediate device (router or gateway)during the simulation period. The mean throughput for a sequence of packets of specific size is calculated by using equations 1 and 2. --------(1) Where Thrj=
------------ (2)
Where Thrj is the value of the throughput when the packet j is received at intermediate device like DSTM gateway, BDSIIT Gateway, v4/v6 Router and N is the number of received packets at intermediate device, Pr is the no of packets. At intermediate device and Pg is the number of packets created by source host. 2. End-to-End Delay (EED) As we know that to calculate the performance evaluation of various novel transition mechanisms EED plays an important role. The mean EED for a sequence of packets of specific size is calculated as follows. -------(3) Where as EEDk=Tdk-Tsk-----(4)
Where EEDk is the End-to-End delay of packet “k”, Tsk is the generated time of packet “k” at source workstation and Tdk is the received time of the packet “k” at the destination workstation, Nr is the total number of received packets at the destination workstation and Mean EED. 4.1. Simulation Results and Discussions. The following table (Table-1)shows the simulation parameters, which are used to calculate the performance measurements using ns-2 simulation environment when each packet arrival follows a Poisson process with rate λ=2
.Table-4: Simulation parameters. Sl.No 1 2 3 4 5
Simulation parameters. Buffer Size Propagation delay Payload Size Very traffic Loads Queue Management Scheme
Scenario-1–The simulation results for scenario one present in the EED and throughput result when a direct link connection between either IPv4 workstations in IPv4 only network or IPv6
Velammal College of Engineering and Technology, Madurai
Value 500 Packets. 10ms 200 Bytes 6~240 Nodes. Drop tail
workstations in IPv6 network is conducted through IPv4/IPv6 router. The results are illustrated in Fig.9.andFig.10.
Page 266
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.9: The comparison between EED of v4-to-v4 and v6-to-v6 communication sessions.
source and destination hosts compared with the IPv4 that allows these processes to be performed by all the Intermediate devices, which fall between the source and destination hosts. The Fig.9.also shows that the difference between the v4-to-v4 EED and v6-to-v6 EED results is very small and the differences increases when the number of connected workstations increases as well. The v6-to-v6 EED is suddenly increases because of network congestion and when the number of connected nodes becomes more i.e. greater than 200.On the other hand the following Fig.10. Shows a Comparison between the v4-to-v4 and v6to-v6 throughput results.
From Fig.9.the simulation results shows that the v4-tov4 EED is definitely less than v6-to-v6 EED.In fact there may be so many reasons which explains why this happens. The first reason is the difference between IPv6 header (40 Bytes) and the IPv4 header (20 Bytes) and it causes more traffic overhead especially when the IP packet payload is small. The Second reason is the size of IP packet payload is fixed, in all communication sessions, this means no fragmentation process is needed, which leads to reduce the benefit from the IPv6 feature that allows the fragmentation and defragmentation processes to be performed by only
Fig.10: A Comparison between the throughput (v4-to-v4, and v6-to-v6) communication sessions.
VI. References.
4.2. DNS Response Time:The DNS Response Time (DNSRT) metric shows that the time needed to calculate, the communication session between the two end systems which are located the two heterogeneous networks. The DNS Response time can be calculated by using the equation (1)
Where Transtime = Transmission time of a Packet is the link number between the two nodes, and Proctime = Processing time, j is the node number and M is the total number of nodes. V. Conclusions. In this paper we have proposed our novel BD-SIIT Transition mechanism. The Simulation results shows the impact of the translation process, that contains performing of address mapping as well as the header translation process which are needed for each incoming and outgoing packet to and from the BD-SIIT translator. The BD-SIIT novel transition concentrated on identifying, determining, 5. translating and forwarding packets between the two different network environments (IPv6 and IPv4 zones).Our proposed new BD-SIIT reduces the size of packet compared with the 6. encapsulation method in the tunneling algorithm.
Velammal College of Engineering and Technology, Madurai
1. Hanumanthappa.J.,Manjaiah.D.H.,”IPv6 and IPv4 Threat reviews with Automatic Tunneling and Configuration Tunneling Considerations Transitional Model: A Case Study for University of Mysore Network”, International Journal of Computer Science and Information(IJCSIS)Vol.3.,No.1,July-2009,ISSN 1947-5500,Paper ID: 12060915] of IPv4 Network 2. Hanumanthappa.J.,Manjaiah.D.H.,”Transition Applications to IPv6 Applications”[TIPv4 to TIPv6],Proceedings of IEEE International Conference on emerging trends in computing(ICETiC2009),Virudhunagar,Tamilnadu 8-10,January 2009,INDIA.[Paper ID 234].
3. Hanumanthappa.J. Manjaiah.D.H., Tippeswamy.K. “An Overview of Study on Smooth Porting Process Scenario during IPv6 Transition” [TIPv6] , Proceedings of IEEE International Conference on the IEEE International Advance Computing Conference IACC-2009 on March 5-8 at Patiala, Punjab [Paper ID IEEE-APPL-1278] 1) 4. Hanumanthappa.J.,Manjaiah.D.H.,“A Study on Comparison and Contrast between IPv6 and IPv4 Feature Sets” Proceedings of International Conference on Computer Networks and Security(ICCNS2008),Pune,Sept 27-28th,2008,[Paper code CP 15]. 5.Ra’Ed AlJa’afreh, John Mellor, Mumtaz Kamala,”A Novel IPv4/IPv6 transition mechanism which support transparent connections”. 6.Behrouz A.Forouzan,Third Edition,“TCP/IP Protocol Suite” .Atul Kahate,“Cryptography and Network.Security“,Tata McGraw-
Page 267
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hill,2003,pp-8-10.Kurose.J.& Ross.K.(2005)Computer Networking:A top-down approach featuring the Internet.3rd ed,(Addison Wesley).S.Hagen:“IPv6 essentials“,Orielly,July 2002,ISBN-0-59600125-8. 7.Moby dick project: “Mobility and Differentiated services in a future IP network:” Final project report, 4.4.2004. 7. J.Wiljakka(ed.,)“Analysis on IPv6 transition in 3GPP networks”,draft-ietf-v6ops-3gpp-analysis- 04.txt,Internet draft, work in progress 8. MIND:Mobile IP based network developments, IST Project, http://www.mind-project.org. [20].NOMAD: Integrated networks for Seamless and Transparent service Discovery, IST Project, http://www.ist-moby Dick.org. 9. Moby Dick: Mobility and Differentiated services in a future IP Network, IST Project, www.ist.mobydick.org. 10. John.J.Amoss and Daniel Minoli, Handbook of IPv4 to IPv6 Transition: methodologies for institutional and Corporate Networks. Auerbach Publications. 11. S.G.Glisic Advanced Wireless Communications, 4G Technology. John Wiley Chichester,2004. 12. Juha wiljakka ,Jonne Soninnen,Managing IPv4 –to –IPv6 Transition Process in Cellular Networks and Introducing new Peer-to-Peer Services. 13. Ioan R,Sherali.Z.2003.Evaluating IPv4 to IPv6 mechanism.IEEE,West Lafayette,USA,v(1):1091–1098
Transition
14. Gilligan. & Nodmar .E. (1996) Transition Mechanisms for IPv6 Hosts and Routers. OMNeT++Discrete Event Simulation.
15. L. Toutain
16. 17.
18. 19. 20.
, H.Afifi, Dynamic Tunneling: A new method for IPv4 to IPv6 transition. Internet draft, <draft-ietf-ngtranns-dti-00.txt> E.Nordmark, Stateless IP/ICMP translation algorithm (SIIT), RFC 2765, February 2000. Ra’ed AlJa’afreh, John mellor,Mumtaz Kamala,”A Novel IPv4/IPv6 transition mechanism which support transparent connections”. TimRooney,IPv4/IPv6Transition strategies ,Director, Product management, BT Diamond IP.John.J.Amos and Daniel Minoli,Handbook of IPv4 to IPv6 Transition Methodologies for Institutional and Corporate Networks Jivika Govil,Jivesh Govil, Navkeerat Kaur, Harkeerat Kaur, An examination of IPv4 and IPv6 Networks: constraints, and various transition mechanisms. K.K.Ettikan, et al.”Application Performance Analysis in Transition mechanism from IPv4 to IPv6,”Multimedia University(MMU),Jalan Multimedia, June 2001. porject.SHISA,2006,http://www.mobileip.jp/
21. J.Bound,
Mysore-06 and currently pursuing Ph.D in Computer Science and Engineering, from Mangalore University under the supervision of Dr.Manjaiah.D.H on entitled “Design and Implementation of IPv6 Transition Technologies for University of Mysore Network (6TTUoM)”. His teaching and Research interests include Computer Networks,Wireless and Sensor Networks, Mobile Ad-Hoc Networks, Intrusion detection System, Network Security and Cryptography,Internet Protocols,Mobile and Client Server Computing,Traffic management,Quality of Service, RFID,Bluetooth,Unix internals, Linux internal, Kernel Programming,Object Oriented Analysis and Design etc.His most recent research focus is in the areas of Internet Protocols and their applications.He received his Bachelor of Engineering Degree in Computer Science and Engineering from University B.D.T College of Engineering,Davanagere,Karnataka(S),India(C),Kuve mpu University,Shimoga in the year 1998 and Master of Technology in CS&Engineering from NITK Surathkal,Karnataka(S ),India (C) in the year 2003.He has been associated as a faculty of the Department of Studies in Computer Science since 2004.He has worked as lecturer at SIR.M.V.I.T,Y.D.I.T,S.V.I.T,of Bangalore. He has guided about 250 Project thesis for BE,B.Tech,M.Tech,MCA,MSc/MS.He has Published about 15 technical articles in International ,and National Peer reviewed conferences. He is a Life member of CSI, ISTE,AMIE,IAENG,Embedded networking group of TIFAC–CORE in Network Engineering,ACM,Computer Science Teachers Association(CSTA),ISOC,IANA,IETF,IAB,IRTG,etc. He is also a BOE Member of all the Universities of Karnataka,INDIA.He has also visited Republic of China as a Visiting Faculty of HUANG HUAI University of ZHUMADIAN,Central China, to teach Computer Science Subjects like OS and System Software and Software Engineering,Object Oriented Programming With C++,Multimedia Computing for B.Tech Students. in the year 2008.He has also visited Thailand and Hong Kong as a Tourist.
Dual Stack Transition mechanism,draft-bound-dstm-exp-
1.txt,april,2004,http://www.dstm.info.
Mr.Hanumanthappa. J. is Lecturer at the DoS in CS,University of Mysore, Manasagangothri,
Velammal College of Engineering and Technology, Madurai
Dr.Manjaiah.D.H D.H. is currently Reader and Chairman of BoS in both UG/PG in the Computer Science at Dept.of Computer Science,Mangalore University, and Mangalore.He is also the BoE Member of all Universities of Karnataka and other reputed universities in India.He received Ph.D degree from University of Mangalore,
Page 268
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
M.Tech. from NITK,Surathkal and B.E.,from Mysore University.Dr.Manjaiah.D.H D.H has an extensive academic,Industry and Research experience.He has worked at many technical bodies like IAENG,WASET,ISOC,CSI,ISTE,and ACS. He has authored more than -25 research papers in international conferences and reputed journals. He is the recipient of the several talks for his area of interest in many public occasions. He is an expert committee member of an AICTE and various technical bodies. He had written Kannada text book,with an entitled, ”COMPUTER PARICHAYA” ,for the benefits of all teaching and Students Community of Karnataka.Dr.Manjaiah D.H’s areas interest are Computer Networking & Sensor Networks, Mobile Communication, Operations Research, E-commerce, Internet Technology and Web Programming
.
Aravinda.C.V.,currently pursuing M.Tech(I.T) K.S.O.U., Manasagangotri, Mysore-06.He received M.Sc ., M.Phil in Computer Science.He has worked as a Lecturer in the following institutions. 1.CIST, Manasagangotri,Mysore, 2.Vidya Vikas Institute of Engg and Technology,Mysore. 3.Govt First Grade college,Srirangapatna and Kollegal. He has published two papers in National Conference hosted by NITK,Surathkal,Mangalore.
Velammal College of Engineering and Technology, Madurai
Page 269
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The Role of IPv6 over Fiber (FIPv6): Issues, Challenges and its Impact on Hardware and Software 1
Hanumanthappa.J.1, Manjaiah.D.H.2, Aravinda.C.V.3 Teacher Fellow, Dos in CS, University of Mysore, Manasagangothri, Mysore, INDIA. 2 Reader & Chairman, Mangalore University, Mangalagangotri, Mangalore, INDIA. 3 M.Tech. KSOU. Manasagangotri, Mysore, INDIA. [email protected] [email protected] [email protected]
Abstract The Internet protocol was started in early 1980.In the earlys1990’s the Internet was growing some temporary solutions were offered in order to cover the fast growth in number of Internet user’s such as Network address translation (NAT), CIDR (Classless Inter domain Routing).At the Same time IETF began working to develop a new Internet Protocol namely IPv6 which was designed to be a Successor to the IPv4 Protocol. This paper proposes the new concept of error handling at network layer (layer-3) instead of data link layer (layer-2) in ISO/OSI reference model by adopting new capabilities and by using IPv6 over Fiber. This paper also shows how to reduce the over head In terms of header processing at data link layer by eliminating cyclic redundancy check (CRC) field by using IPv6 over Fiber. Therefore we can also prove that ISO/OSI model contains 6 layers instead of 7 layers. Key words: IPv4, IPv6, CRC, Ethernet, FCS.
I.
Introduction to IPv6
In the last 20 years, the internet undertook a huge and unexpected explosion of growth [].There was an effort to develop a protocol that can solve problems in the current Internet protocol which is in the current internet protocol which is in Internet protocol version 4(IPv4).It was soon realized that the current internet protocol the IPv4, would be inadequate to handle the internet’s continued growth. The internet Engineering task force (IETF) was started to develop a new protocol in 1990’s and it was launched IPng in 1993 which is stand for Internet protocol Next Generation. So a new generation of the Internet Protocol (IPv6) was developed [7], allowing for millions of more IP addresses. The person in charge of IPng area of the IETF recommended the idea of IPv6 in 1994 at Toronto IETF[1].But mainly due to the scarcity of unallocated IPv4 address the IPv4 protocol cannot satisfy all the requirements of the always expanding Internet because
Velammal College of Engineering and Technology, Madurai
however its 32 bit address space being rapidly exhausted[2] alternative solutions are again needed[3].It is reported that the unallocated IPv4 allocated IPv4 addresses will be used with 6 to 7 years short period[2].The Long term solution is a transition to IPv6[5]which is designed to be an evolutionary step from IPv4 where the most transport and application layer protocol need little or no modification to the work. The deployment of NAT [3] can alleviate this problem to some extent but it breaks end to end characteristic of the Internet, and it cannot resolve the problems like depletion (exhaustion) of IPv4 addresses. 1.1. Features of IPv6. The main reason for designing this new Internet protocol(IPv6) was the need to increase the number of addresses(address spaces).The IPv6 address was designed with a 128-bit address scheme instead of 32-bit scheme in IPv4.So,the number of possible addresses in IPv6 is 3.4X1038 unique addresses.IPv6 will have enough to uniquely address every device (example Telephone, Cell phone,mp3 player,hosts,routers,bridges etc) on the surface of earth with full end-to-end connectivity(about 32 addresses per square inch of dry land).In addition IPv6 is designed to support IPSec, Scalability, Multimedia transmissions, Security, Routing, Mobility, Real time applications like audio,Video,Cryptography techniques like encryption and Decryption, Large address space, Better support for QoS. 1.2. Data Link error handling function. In this paper performance can be measured in many ways, including transit time and response time. Transit time is the amount of time for a message to travel from one device to another
Page 270
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
. Fig.1:An Exchange using the OSI model.
Transit Time (TT) =Time taken from Source to Destination. [17].Response Time (RT) =Elapsed time between an inquiry and response [17].Performance is also evaluated by two networking metrics like throughput and delay
Fig.2: The Interaction between layers in the ISO/OSI model.
1.3. Layered architecture. The OSI model is composed of seven ordered layers physical(layer1),Data link(layer2),Network(layer3), Transport(layer4), Session(layer5), Presentation(layer6), Application layer(7), To reduce the design complexity, computer networks follow a layered architecture[1].Each layer clearly defines based on the previous layers and has a set of well defined functions with clear cut boundaries. Also with layered architecture the implementation details of each layer is independent of other layers.Fig.2.shows the layers involved when a message is sent from device A to device B.As a message travels from A to B it may pass through many intermediate nodes. These intermediate nodes usually involved only the first three layers of OSI model. Each layer defines a family of functions distinct
Velammal College of Engineering and Technology, Madurai
from those of the other layers. By defining and localizing functionality in this fashion, the designers created an architecture that is both comprehensive and flexible. Most importantly the ISO/OSI model allows complete interoperability between otherwise incompatible systems. Within a single machine, each layer calls upon the services of the layer just below it. The processes on each machine that communicates at a given layer are called peer-to-peer processes. The passing of the data and network information down through the layers of the sending device and back up through the layers of the receiving device is made possible by an interface between each pair of adjacent layers. Layers 1, 2 and 3–physical, data link and network are the network support layers. Layers 5, 6, and 7–session, presentation and application can be thought of as the user support layers. The upper OSI layers are almost always implemented in software; lower layers are a combination of hardware and
Page 271
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
software except for the physical layer, which is mostly hardware. In Fig.1. Which gives an overall view of the OSI layers, D7 means the data unit at layer 7.The process starts at layer 7, then moves from layer to layer in descending, sequential order.At each layer commonly a header or trailer can be added to the data unit. At each layer the packet is encapsulated with a header that contains control information to handle the data received at other side by the corresponding layer. This paper also states that how take care of error handling function very efficiently by reducing the overall packet processing time, and thus improve the transmission of IPv6 packets. This can be achieved by
utilizing the characteristics or capabilities of the communication medium used to transfer data, and by improving the existing error handling mechanisms at the lower layer. In Data communication and networking and computer networks the errors are broadly divided into single bit error and Burst errors. The term single-bit error means that only 1 bit of a given data unit (such as a byte, character or packet) from 0 to 1.The Fig.4.shows Burst error with Length 5.The most important common approaches to detect the errors are parity check(pc),Cyclic redundancy check(CRC) and Checksum.[17][19]
Fig.3: Single-bit error.
Fig.4: Burst error of Length 5.
Fig.5: Error Detection methods.
The errors that are detected by either redundancy, parity check, or cyclic redundancy check, or checksum can be corrected by with two types of mechanisms called Automatic repeat request(ARQ),and Forward error correction(FEC).This paper is organized as follows: We briefly described Introduction to IPv6 in 4G networks and its on-going work in Section 1.We described,IPv6 transition in terms of 4G in section-2.Section-3 clearly specifies related work meant for the translation of IPv4/IPv6 BD-SIIT a novel transition algorithms. Finally we concluded the whole paper in section 4.
Velammal College of Engineering and Technology, Madurai
III. The ISO/OSI Packet Transmission. The ISO/OSI and TCP/IP are the two most important popular network architectures that have been widely used. The ISO/OSI reference model has remained as a popular model for its simplicity, and clarity of its functions where the TCP/IP was a more working model that is popularly used over the Internet. The Fig.6.shows the Transport layer communication process. The data from the user on the sender side, passes through a series of layers before it is transmitted over the internet to reach the other machine(recipient side) .On the receiver side data received
Page 272
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
by the Physical layer goes up to the application layer, by neglecting the header in each layer. It can be observed from
the Fig.6.the original length of the data remains same, but the length of the header increases with each layer.
Fig.6: The Transport layer communication process.
3.1. IPv6 Packet format
. Fig.7: The IPv6 Packet format.
The IPv6 packet consists of IPv6 base header, extension headers, and upper layer protocol data unit.IPv6 base header is of fixed size, and is of 40 bytes in length. Payload length may change due to the presence of extension headers. The IPv6 packet format can also support for
multiple extension headers and the use of extension header is optional. The IPv6 based header consists of 8 fields as shown in the Fig.7.Extension headers are inserted into the packet only if options are needed. They are processed in the order in which they are present.
Fig.8: The Ethernet Frame Format.
3.2. Error handling aspects and its issues by using CRC in Data link layer.
1011000 is a code word and when cyclically left shifted then 0110001 is also a code word.
CRC is one of the most popular error detection mechanisms that are currently being used at data link layer. Cyclic codes are special linear block codes with one extra property. In a cyclic code if a code word is cyclically shifted then the result is another code word. For ex – if a
3.2.1. Cyclic redundancy check.
Velammal College of Engineering and Technology, Madurai
CRC is a cyclic code method to correct errors in networks like LAN and WANs.CRC is a popular method used simple to implement in binary hardware, and easy to analyze mathematically and are particularly good at
Page 273
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
detecting common errors caused by noise in transmission channels. In our proposed method by removing CRC from the Ethernet frame format and fix it in the IPv6 packet
format of the extension header at network layer of ISO/OSI model. The Fig.9.shows the novel IPv6 packet format [17].
IV. Network layer error checking Methodology (FIPv6EH)
Fig.9:IPv6 novel packet format with CEH (CRC Extension header).
3.3. Applications of CRC. 1.Cyclic codes performing very good performance in detecting single-bit errors, double errors, an odd number of errors, and burst errors. 2. CRC is easy to implement in hardware and software. 3. CRC is very fast to correct errors when implemented in hardware. 4. Cyclic code has made a good candidate for many networks [17].
In this, paper we considered to place error detection in the network layer to check IPv6 whole packet including header and payload. The error detection (check) method used will be same as CRC method that is currently being used with the existing systems. The CRC extension header (CRCEC) is a new IPv6 extension header to handle error detection for the entire IPv6 packet shown in Fig.9. Simulation Scenario-1: The first simulation states that when sender generates an IPv6 packet, and the corresponding FIPv6EC (CRCEC) code to be inserted to the IPv6 packet format as CRCEC.The packet with CRCEH is sent through a network with a topology as mentioned in Fig.10.The routers are mainly used to connect sender host and a recipient host will not verify the FIPv6H (CRCEH) instead it will identify the next route of the packet. When the receiver upon receiving the packet will verify the CRCEH (FIPv6EH) in its network layer to check whether the packet is error free, and then deliver to the upper layers. If suppose the received packet contains an error it will be discarded and wait for retransmission [21].
Fig.10: Novel FIPv6EH Error detection model in Network Layer.
Simulation-2: Second simulation also covers error control in data link layer, the sender generates the IPv6 packet without, any extension header inside. It is embedded in data link layer with header and trailer. The FCS in the trailer is actually CRC-32 code generated from the whole frame. The Fig.10.shows when the intermediate node-1 which is also called router-1 receives a packet, and verifies the CRC code inside, when the verification process generated no code in the packet then only the packet is processed to next router-2 of a network layer. Each time the bad packets are rejected and then the receiving station will be waiting for the retransmission of a packet. This process will continue in all the intermediate routers, which
Velammal College of Engineering and Technology, Madurai
connects the sender with receiver. We have analyzed from the simulation scenario-1 and scenario-2 there is no error control in data link layer and it takes place only at the end workstations, and will not be done at every routers. V. Performance Evaluation Metrics and Simulation Parameters. In this paper we have calculated two performance evaluation metrics like throughput, end-to-end delay. The throughput can be calculated as follows. The throughput is a measure of how fast we can actually send data through a network. Although at first glance throughput and
Page 274
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
bandwidth in bits per second seem the same, but totally there are different. The bandwidth is a total measurement of a link; the throughput is an actual measurement of how fast we can send data. We measured the throughput performance metric value in order to find out the rate of received and processed data at the router (intermediate device) during the time of simulation. The mean throughput for a sequence of packets of specific size can be calculated by using the formula
-----(1) whereas Thri=Paccept/Pcreated*100%----(2) Where Thri is the throughput value when the packet “i” is accepted at the intermediate device like router. and “n” is the total number of packets received packets at the router, and Prec is the number of received packets at router and Pcrea The below Table-1 shows the Simulation results that are mainly used to calculate, the Performance measurements .
is the number of packets created by the source hosts, and the mean throughput is the mean value for each communication. Latency (Delay): The latency or delay defines how long the entire message takes to completely arrive as at the destination from the time the first bit is sent out from the source. Therefore we can conclude latency is made up of four important components: Propagation time, Transmission time, Queuing time and Processing delay. Delay=Propagation time (Pt) +Transmission time (Tt) +Queuing time (Qt) + Processing delay (Pd). Where Pt=Distance/Propagation speed. Transmission Time=Message Size/Bandwidth. Queuing time=Time needed for each Intermediate or n devices to hold the message before it can be processed
using the NCTUNS5.0 Simulator and emulator.The Simulation parameters as shown in the Table-1
Table-1: Simulation parameters. Simulation Value Parameters 1.Buffer Size 200 packets 2.Delay 5ms 3.Pay load size 100 Bytes 4.Vary traffic load 6~150 nodes. 5.Queue mngt Drop tail Scheme V. Challenges, necessity and Impact of IPv6 over Fiber on Hardware and Software. 5.1. Challenges of IPv6 over Fiber. 1. To minimize the header overheads by handling efficiently similar or redundant functionalities among layers. 2. To maximize the performance of data transmission in terms of packet data rate, throughput, and bandwidth utilization. 3. To maximize the efficiency of operations for data transmission. 4.To propose a new new frame work by simplifying the existing network models like ISO/OSI and TCP/IP model through possible changes in data link layer and network layer.
Velammal College of Engineering and Technology, Madurai
5. By collapsing data link layer and network layer for similar functions and try to eliminate redundant functions. 6. It is also possible to enhance performance of IPv6 packet transmission in terms of frame rate, throughput and bandwidth. 7. To Simplify the ISO/OSI layered architecture by reducing the size of the data link layer header or removing it completely [21]. 5.2. Necessity of IPv6 over Fiber. 1. Ellimination of MAC addresses. 2. Discard Data Link layer. 3. No framing process. 4. Increase speed transfer. 5. Reduce Worm and Virus outbreak. 6. Efficiency can be increased and Buffering can be reduced.
Page 275
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
7. ARP currently being used is not required. 8. Data is encapsulated into many headers. 9.26 bytes of Data Link Layer header for each frame.Max data size in each frame is 1500 bytes. Some amount of frame size is allocated for the headers and it is waste of bandwidth [21].
3. Layer 3 switches will experience faster packet processing 4. Routing Table will consist of IPv6 Address and smaller routing table entry. 5. No MAC address to IP mapping is required resulting in router performing faster and more efficient [21].
5.3. Impact of IPv6 over Fiber on Hardware.
5.4. Impact of IPv6 over Fiber on Dual Stack.
The impact of IPv6 over Fiber can updates some of the changes in hardware. The following are some of the changes in IPv6 over Fiber. 1. Only drivers can be re-written for NIC.No changes required for manufacturing NIC. 2. Layer 2 switches will be obsolete or will have to run dual stack.
1. Once a new architecture is in its placeman dual stack is required to process the packets transmitted from both current architecture as well as new architecture. 2. The initiated protocol will automatically choose the appropriate stack. 3. This will be an international standard to be used by all Internet applications especially for real time applications [21].
Fig.11:IPv6 over Fiber. VI. Conclusions. In this paper we have proposed our new concept of error handling mechanism like FIPv6EH at network layer instead of at Data link layer with the help of various salient features of IPv6 protocol and the various characteristics features of high speed communication networks like Fiber. This paper not only enhances the error handling issues at network layer along with that its exploits very briefly Challenges and Impact of IPv6 over Fiber on Hardware as well as Software. The proposed FIPv6EH reduces the overhead by eliminating CRC field from its frame header and placing it in the extension header in the Network layer instead of doing header processing at data link layer. The proposed concept will enhance the performance of packet transmission. As we know that Data link layer and Network layer in ISO/OSI model are the two important lower layers where the IPv6 protocol is in between neighboring nodes. This paper also states that error checking at the packet will be more efficient than at frame level by using advanced technology and today’s faster routers. References [1]. Manjaih .D.H. Hanumanthappa.J,2008,A Study on Comparison and Contrast between IPv4 and IPv6 Feature sets.In Proceedings of ICCNS’08, 2008,Pune,297-302. [2]. Manjaih.D.H.,Hanumanthappa.J.,2008,Transition of IPv4 Network Applications to IPv6 Applications,In Proceedings of ICETiC- 09,2009,S.P.G.C. Nagar,VirudhaNagar-626 001,TamilNadu,INDIA-35-40.
Velammal College of Engineering and Technology, Madurai
[3] Manjaih.D.H.Hanumanthappa.J. 2009, IPv6 over Bluetooth: Security Aspects, Issues and its Challenges, In Proceedings of NCWNT-09, 2009, Nitte-574 110, Karnataka, INDIA–18-22. [4] Manjaih.D.H.,Hanumanthappa.J.,2009,Economical and Technical costs for the Transition of IPv4 –to-IPv6 Mechanisms [ETCTIPv4 to ETCTIPv6],In Proceedings of NCWNT-09,2009,Nitee574110,Karnataka,INDA-12-17. [5]. Manjaih.D.H. Hanumanthappa.J. 2009, Smooth porting process scenario during the IPv6 transition, in Proceedings of IACC’09, March 6-7, Patiala, Punjab, INDIA. [6] Manjaiah. D. H. and Hanumanthappa.J. IPv6 over IPv4 QoS metrics in 4G–Networks: Delay, Jitter, Packet Loss Performance, Throughput and Tunnel Discovery mechanisms. Proceedings of NCOWN-2009, RLJIT, Doddaballapur, Bangalore, Karnataka, INDIA, August-21-22-pp.122-137. [7] Hanumanthappa. J. And Manjaiah.D.H.”A Study on IPv6 in IPv4 Static Tunneling threat issues in 4G Networks using OOAD Class and Instance Diagrams” in Proceedings of the International Conference on Emerging trends in Computer science and Information technology(CSCIT-2010)organized by Dept.of.CS and Information Technology,Yeshwant ahavidyalaya,Nanded,Maharashtra,India,(2010). [8] Hanumanthappa.J. and Manjaiah.D.H.”An IPv4–to-IPv6 Threat reviews with Dual Stack Transition mechanism Considerations a Transitional threat model in 4G Wireless “in Proceedings of the International Conference on Emerging trends in Computer science and Information technology (CSCIT-2010)organized by Dept of CS and Information Technology,Yeshwant Mahavidyalaya,Nanded,Maharashtra,India,(2010). [9] Manjaiah.D.H. and Hanumanthappa.J,”IPv6 and IPv4 Threat review with Automatic Tunneling and Configuration Tunneling Considerations Transitional Model: A Case Study for University of Mysore Network”International Journal of Computer Science and Information Security (IJCSIS), Vol.3, (2009). [10] S.Deering and R. Hinden “Internet Protocol Version 6(IPv6) Specification”, RFC 2460, December 1998. [11]. J.Postel, INTERNET PROTOCOL, RFC 0791, September 1981.
Page 276
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[12] S.Tanenbaum, “Computer Networks”, Third Edition, Prentice Hall Inc., 1996, pp.686, 413-436,437-449. [13] Beerhouse A.Forouzan, Third Edition, “TCP/IP Protocol Suite”. [14] Atul Kahate, “Cryptography and Network Security”, Tata McGrawHill, 2003, pp-8-10. [15] Kurose.J. & Ross.K. (2005)Computer Networking: A top-down approach featuring the Internet.3rd Ed, (Addison Wesley). [16] S.Hagen:“IPv6 essentials“, Orielly, July 2002, ISBN-0-5960-01258. [17] Behrouz A Forouzan, Fourth edition Data communication and Networking, pp-7-8. [18] S.Shukla and N.W.Bergmann, “Single bit error correction implementation in CRC-16 on FPGA”, in proceedings of The IEEE International conference on Field-Programmable technology, pp-319-322,204. [19] J.Davies, “Understanding IPv6”, Microsoft press, Washington: Microsoft press, 2003. [20] T.Parnell, Building High Speed networks, First edition, Mc-GrawHill, 1999. [21] Raja Kumar Murugesan, Dr.Rahmat Budiarto, Prof.Sureswaran Ramadass, Enhanced performance of IPv6 packet transmission over Fibre.
Mr.Hanumanthappa. J. is Lecturer at the DoS in CS,University of Mysore, Manasagangothri, Mysore-06 and currently pursuing Ph.D in Computer Science and Engineering, from Mangalore University under the supervision of Dr.Manjaiah.D.H on entitled “Design and Implementation of IPv6 Transition Technologies for University of Mysore Network (6TTUoM)”. His teaching and Research interests include Computer Networks, Wireless and Sensor Networks, Mobile Ad-Hoc Networks, Intrusion detection System, Network Security and Cryptography, Internet Protocols, Mobile and Client Server Computing,Traffic management,Quality of Service,RFID,Bluetooth,Unix internals, Linux internal, Kernel Programming,Object Oriented Analysis and Design etc.His most recent research focus is in the areas of Internet Protocols and their applications.He received his Bachelor of Engineering Degree in Computer Science and Engineering from University B.D.T College of Engineering ,Davanagere,Karnataka(S),India(C),Kuvempu University,Shimoga in the year 1998 and Master of Technology in CS&Engineering from NITK Surathkal,Karnataka(S ),India (C) in the year 2003.He has been associated as a faculty of the Department of Studies in Computer Science since 2004.He has worked as lecturer at SIR.M.V.I.T,Y.D.I.T,S.V.I.T,of Bangalore. He has guided about 250 Project thesis for BE,B.Tech,M.Tech,MCA,MSc/MS.He has Published about 15 technical articles in International ,and National Peer reviewed conferences. He is a Life member of CSI, ISTE,AMIE,IAENG,Embedded networking group of TIFAC–CORE in Network Engineering,ACM,Computer Science Teachers
Velammal College of Engineering and Technology, Madurai
Association(CSTA),ISOC,IANA,IETF,IAB,IRTG,etc.H e is also a BOE Member of all the Universities of Karnataka,INDIA.He has also visited Republic of China as a Visiting Faculty of HUANG HUAI University of ZHUMADIAN,Central China, to teach Computer Science Subjects like OS and System Software and Software Engineering,Object Oriented Programming With C++,Multimedia Computing for B.Tech Students. in the year 2008.He has also visited Thailand and Hong Kong as a Tourist. Dr.Manjaiah.D.H D.H. is currently Reader and Chairman of BoS in both UG/PG in the Computer Science at Dept.of Computer Science,Mangalore University, and Mangalore.He is also the BoE Member of all Universities of Karnataka and other reputed universities in India.He received Ph.D degree from University of Mangalore, M.Tech. from NITK,Surathkal and B.E.,from Mysore University.Dr.Manjaiah.D.H D.H has an extensive academic,Industry and Research experience.He has worked at many technical bodies like IAENG,WASET,ISOC,CSI,ISTE,and ACS. He has authored more than -25 research papers in international conferences and reputed journals. He is the recipient of the several talks for his area of interest in many public occasions. He is an expert committee member of an AICTE and various technical bodies. He had written Kannada text book,with an entitled,”COMPUTER PARICHAYA”,for the benefits of all teaching and Students Community of Karnataka.Dr.Manjaiah D.H’s areas interest are Computer Networking & Sensor Networks, Mobile Communication, Operations Research, E-commerce, Internet Technology and Web Programming.
Aravinda.C.V.,currently pursuing M.Tech(I.T) K.S.O.U., Manasagangotri, Mysore-06.He received M.Sc ., M.Phil in Computer Science. He Worked has Lecturer at Various Institutions :1) CIST University Of Mysore Manasagangothri Mysore.2) Vidya Vikas Institute Of Engineering and Techonology Mysore.3) Govt.First Collge Srirangapatna,Mandya District .4) Govt First College Kollegal ,Chamrajanagar Distirct. 5) Worked has Technical Co-ordinator NIIT Gandhi Bazar,Bangalore. He has published two National Papers hosted by NITK Surthkal Mangalore, DK.
Page 277
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Localized CBIR for Indexing Image Databases D.Vijayalakshmi 1, P. Vijayalakshmi 2 1,2
Faculty of Kamaraj College of Engg and Technology, Virudhunagar, Tamilnadu, India. [email protected], [email protected], 3
P.Jeyapriya 3
Final B.Tech, IT, Kamaraj College of Engg and Technology, Virudhunagar.
Abstract: At present, research has been focused on image indexing techniques based on global feature distributions. However, these global distributions have limited discriminating power because they are unable to capture local image details. Localized image retrieval is required when it is preferred to index image databases based on the similarity metrics. Wavelet represents images at various resolution levels. A high wavelet coefficient at a coarse resolution corresponds to a region with high global variations and vice versa. In this paper, we present wavelet based localization for indexing image databases employing color and texture features. The feature vectors thus computed is unique and has sharp discrimination power in various image regions. The new system performs better in terms of computational time and storage requirements as compared to the global feature approaches. The potential applications include automatic image classification, video retrieval and video data management Keywords, Wavelet Decomposition, Image indexing, feature extraction, Matching Criteria, localization.
• I. INTRODUCTION Many picture libraries use keywords as their main form of retrieval – often using indexing schemes developed inhouse, which reflect the special nature of their collections. A good example of this is the system developed by Images to index their collection of stock photographs .Users should be able to explore images in a database or video clips by visual similarities. Structuring and visualizing digital images are based on their content similarities. Currently, many content based image retrieval techniques have been developed to incorporate higher-level feature extraction capabilities, but a lot of work remains to be done. Ultimately, feature-extraction techniques, combined with other techniques are expected to narrow down the gap between relatively primitive features extracted from images and high-level semantically-rich perceptions by humans. Many picture libraries use keywords as their main form of retrieval – often using indexing schemes developed in-house, which reflect the special nature of their collections. Index terms are assigned to the whole image, and the main objects. When discussing the indexing of images and videos, one needs to distinguish between systems which are geared to the formal description of the image and those concerned with subject indexing and
Velammal College of Engineering and Technology, Madurai
retrieval. The former is comparable to the bibliographical description of a book. However, there is still no one standard in use for image description, although much effort is being expended in this area by a range of organizations such as the Museum Documentation Association. Access to a desired image from a repository might thus involve a search for images depicting specific types of object or scene, evoking a particular atmosphere, or simply containing a specific texture or pattern. Potentially, images have many types of attribute which could be used for retrieval, including: • the presence of named individuals, locations, or events (e.g. the Queen greeting a crowd); • subjective emotions one might associate with the image (e.g. happiness); meta data such as who created the image, The last decade has seen the appearance of a number of commercial image data management systems. These systems normally store representations of pictorial documents (such as photographs, prints, paintings, drawings, illustrations, slides, video clips, and so on) in static archival databases, and incorporate multimedia database management systems in the storage of, and provision of wider access to, these repositories. Current image indexing techniques have much strength. Keyword indexing has high expressive power – it can be used to describe almost any aspect of image content. It is in principle easily extensible to accommodate new concepts, and can be used to describe image content at varying degrees of complexity. There is a wide range of available text retrieval software to automate the actual process of searching. But the process of manual indexing, whether by keywords or classification codes, suffers from two significant drawbacks. Firstly, it is inherently very labor-intensive. Indexing times quoted in the literature range from about 7 minutes per image for stock photographs at Getty Images, using their in-house system, to more than 40 minutes per image. Secondly, manual indexing does not appear to be particularly reliable as a means of subject retrieval of images. Markey had found that, in a review of inter-indexer consistency, there were wide disparities in the keywords that different individuals assigned
Page 278
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
to the same picture. The advantage of all these techniques is that they can describe an image at varying levels of detail (useful in natural scenes where the objects of interest may appear in a variety of guises), and avoid the need to segment the image into regions of interest before shape descriptors can be computed. Despite recent advances in techniques for image segmentation, this remains a troublesome problem. Wavelet Decomposition The continuous wavelet transform of a 1-D signal f(x) is defined as
(W a f )(b ) = ∫ ( x )Ψ
*
a, b
( x )dx
----------1 where the wavelet b , a Ψ is computed from the mother wavelet Ψ by translation and dilation,
Ψa , b( x ) =
1 Ψ (( x − a ) / b ) |a| ----------
-2 the mother wavelet Ψ satisfies the constraint of having zero mean. The eq. (1) can be discretized by restraining ’a’ and ’b’ to a discrete lattice. Typically it is imposed that the transform should be non-redundant, complete and constitutes a multi resolution representation of the original signal. In practice, the transform is computed by applying a separable filter bank to the image where denotes the convolution operator, ) 2 , 1 ↓ ( 1 , 2 ↓ denotes the down sampling along the rows (columns) and I = A0 is the original image, H and G are low pass and high pass filters, respectively. The n A is obtained by low pass filtering and is referred to as the low resolution (Approximation) image at scale n. The 3 n 2 n 1 n D , D , D are obtained by band pass filtering in a specific direction(Horizontal, Vertical and Diagonal, respectively) and thus contain directional detail information and is referred to as high resolution(Detail) images at scale n. The original image I is thus represented by a set of sub images at several scales. This decomposition is called “Pyramidal wavelet transform” decomposition or discrete wavelet decomposition (DWT). Every detail sub image contains information of a specific scale and orientation. The spatial information is retained within the sub image. In the present paper, the features are obtained using Haar Wavelet (Fig. 1.), which is given by
⎧1 0<=t <=1/2 ⎪ Ψ(t) =⎨−1 1/2<=t <=1 ⎪0 otherwise ⎩ ---------------3
Velammal College of Engineering and Technology, Madurai
2. RELATED WORK Visualization technique of images with more continuous scenes described in this article have a wide range of potential applications, for example, data mining in remote sensing images and image retrieval from film and video archives. This methodology is suitable on a sample of images with more continuous scenes, for example, video segments, so that we will be able to keep track of the impact of various feature-extraction techniques more closely [1]. A proposal is presented for constructing an efficient image sharing system, in which a user is able to interactively search interesting images by content-based methods in a peer-to-peer network. A distributed variant of the Tree-Structured Self- Organizing Maps (TS-SOM) by incorporating the essence of DHT is presented. Compared with the existing approaches, this method has the following advantages: many operations can be performed on the 2-D lattice instead of the highdimensional feature space, which significantly saves the computing resource; the TS-SOM-like hierarchical search speeds up the information lookup and facilitates the use of large Self-Organizing Maps; a query can involve more than one feature simultaneously. The current design is considered for the tree-structured index with fixed levels and fixed size at each level. It is not feasible for some huge other network communities during their evolving stages. Many details, for instance the image normalization and the system strategy have not been considered due to the limit of time and space [2]. In another paper, an extraction of signatures, to determine a comparison rule, including a querying scheme and definition of a similarity measure between images is proposed. Several querying schemes are available: regionbased searching, where the retrieval is based on a particular region in the image, or searching by specifying the color histogram or object shape of the images, significant semantic information is lost. Principal component analysis is used to represent and retrieve images on the basis of content which reduces the dimensionality of the search to a basis set of prototype images that best describes the images. Each image is described by its projection on the basis set; a match to a query image is determined by comparing its projection vector on the basis set with that of the images in the database, the detection performance is transform variant [3]. Here, images are normalized to line up the eyes, mouths and other features. The eigenvectors of the covariance matrix of the face image vectors are then extracted. These eigenvectors are called eigen faces. Raw features are pixel intensity values Ai. Each image has n × m features, where n, m are the dimensions of the image. Each image is encoded as a vector G of these features. Compute “mean” face in training set and detected [4]. A technique for automatic face recognition based on 2D Discrete Cosine Transform (2D-DCT) together with Principal Component Analysis (PCA) is suggested. Applying the DCT to an input sequence decomposes it into a
Page 279
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
weighted sum of basis cosine sequences. A block size of 16×16, 32×32 and 64×64 are chosen to be further transformed using PCA. The DCT transformed face images using the different block sizes are transformed into a lower order feature size using the PCA to get a feature vector of 25 features. Two other block sizes are chosen to be further transformed using PCA; namely a 32×32 and a 64×64 DCT coefficients so as to see the effect of changing the block size on the recognition rate. Also, a feature vector of 25 features is calculated in each case. The recognition rate depends on a chosen threshold based on the security level of the application. Two important aspects are investigated namely the False Accept rate (FAR) and the False Reject Rate (FRR). The FAR is the success probability of an unauthorized user to be falsely recognized as a legally registered user [5], [6].A technique for representing shapes for the purpose of image retrieval is proposed. The proposed representation is a contour based approach. Canny operator is used to detect the edge points of the image. The contour of the image is traced by scanning the edge image and re-sampling is done to avoid the discontinuities in the contour representation. After the object contour has been detected for an object, the central point of the object, the centroid of an object is computed. The resulting image is swept line by line and the neighbor of every pixel is explored to detect the number of surrounding points and to derive the shape features. In this method, when the color image database is considered for comparison 71.17% retrieval efficiency is obtained for top 20 retrievals. Shape feature requires extraction of a region or object, in itself a complicated problem. Accurate shape detection requires human intervention because methods like segmentation are very difficult to completely automate [7]. Retrieving images based on color similarity is achieved by computing a color histogram for each image that identifies the proportion of pixels within an image holding specific values. Current research is attempting to segment color proportion by region and by spatial relationship among several color regions. The main idea of appearance- based approach is to represent the images in terms of their projection onto a low-dimensional space called eigen space is presented. The projection of the images onto this space is called the eigen images. Principal components as sequence of image vectors are computed incrementally, without estimating the covariance matrix and transforming these principal components to the independent directions that maximize the non–Gaussianity of the source. This algorithm takes input image finds the non-Gaussian vector (eigenvector) which is passed as input to the ICA (independent component analysis) algorithm. The nonGaussian components will be updated using an updating rule from the previous component values in a recursive way. IPCA (incremental principal component analysis) returns the estimated eigenvectors as a matrix that represents subspaces of data and the corresponding eigen values as a
Velammal College of Engineering and Technology, Madurai
row vector, Fast ICA searches for the independent directions where the projections of the input data vectors will maximize the non-gaussianity. The nearest neighbor algorithm is used to evaluate the object recognition technique [8], [9]. Four different approaches to color texture analysis are tested on the classification of images from the VisTex database. The first method employs multi spectral approach, in which texture features are extracted from each channel of the RGB color space. The second method uses HSV color space in which texture features are extracted from the luminance channel V and color features from the chromaticity channels H and S. The third method uses YCbCr color space, in which texture features are extracted from the luminance channel Y and color features from the chromaticity channels Cb and Cr. The last one uses gray scale texture features computed for a color Image [10]. In another paper, we present a localized CBIR system, ACCIO! , that uses labeled images in conjunction with a multiple-instance learning algorithm to first identify the desired object and weight the features accordingly, and then to rank images in the database using a similarity measure that is based upon only the relevant portions of the image [11]. A number of indexing schemes use classification codes rather than keywords or subject descriptors to describe image content, as these can give a greater degree of language independence and show concept hierarchies more clearly. Local feature-based image matching was done in two steps. The first step of the feature detection is key point or interest point detection. The second step involved computing descriptors for each detected interest point. Descriptors were used for matching key points of the input image with key points in the logo database. For logo recognition the descriptor had to be distinctive and at the same time robust to changes in viewing conditions. The features were to be resilient to changes in illumination, image noise, uniform scaling, rotation, and minor changes in viewing direction. The descriptor must minimize the probability of mismatch and finally, it should also be relatively easy and fast to extract the features and compare them against a large database of local features. Based on these requirements and the reported robustness in other research projects, Scale Invariant Feature Transform (SIFT) was used in [12] 3. PROPOSED WORK Here , we have used YUV colour spaces in order to obtain one channel containing the luminance information and two others containing chrominance information. The conversion of RGB to YCbCr color model is shown in the eqn (4). Discrete wave let decomposition is employed and is shown in the Fig 3.1. Here, we have chosen Haar wavelet for wavelet decomposition. And its representation is shown in the Fig 3.2. The four quadrants coefficients of the images are obtained. The correlation between approximation and horizontal coefficients, approximation and vertical coefficients and approximation and diagonal coefficients are computed. A
Page 280
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
high wavelet coefficient at a coarse resolution corresponds to a region with high global variations and vice versa. In this paper, we present wavelet based localization for indexing image databases employing color and texture features In addition to colour and texture features and other statical features are also computed. The feature vectors have sharp discrimination power in various image regions. Texture features are computed from the luminance channel. The first order statistical features namely, mean and standard deviation, are computed from the chrominance channel. The following equations transform RGB in [0, 1] to Y Cb Cr in [0,255].
4. FEATURE EXTRACTION The co-occurrence histograms are constructed across different wavelet coefficients of an image and its complement decomposed up to 1-level. All combinations of the image quadrants of wavelet coefficients of image and its complement are considered to compute the correlation between various coefficients. From the normalized cumulative histogram, color and texture features are extracted employing equation ( 5), (6) and (7). 256
Mean :
μ
∑ nchi nch =
i =1 256
………….5
Slope : S nch = Slope of the regression line fitted
….....4
. across the sample points - --------------6 256
Mean Deviation: Dnch =
∑|nchi −μnch|
i=1
256
-----------7 • • • • • • • • • • FIG . 3.1. WAVELET DECOMPOSITION
Algorithm indexing and Retrieval Get image Apply multilevel wavelet decomposition Extract color and texture features Compute feature vector Train all images Get Test image Localize a Image segment to be indexed Compute test feature vector Apply matching criteria and search Retrieve and index all images that matches 5. ANALYSIS OF EXPERIMENTAL RESULTS
In this paper, wavelet based localization for Indexing Image databases are implemented. The paper implemented in RSI IDL 6.3 environment. The project is tested on standard images for indexing the color face images shown in the Fig 5.1. Localized portion of the face images used to index the image database. They are nose and eye images and they are shown in Fig 5.2 and 5.3. The experimental results of indexing for nose and eye image are shown in Fig 5.5 and 5.6. The objective metrics of retrieval algorithm are computed and are shown in Table 5.2. Also, success and failure rate of retrieval algorithm are computed and shown in Table in 5.1. FIG 3.2 HAAR WAVELET REPRESENTATION
Velammal College of Engineering and Technology, Madurai
Page 281
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
FIG 5.3 TRAINED AND UNTRAINED TEST EYE IMAGES
FIG 5.4 SNAPSHOT OF OUTPUT SCREEN
FIG 5.1 TRAINED FACE IMAGE DATA BASE
FIG 5.5 INDEXED OUTPUT FOR NOSE IMAGES
Fig 5.2 TRAINED AND UNTRAINED NOSE IMAGES
Velammal College of Engineering and Technology, Madurai
Page 282
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[5] Application of DCT Blocks with Principal Component Analysis for Face Recognition MAHA SHARKAS, Proceedings of the 5th WSEAS Int. Conf. on Signal, Speech And Image Processing, Corfu, Greece, August 17-19, 2005 (pp107-111). [6] Face Recognition Using the Discrete Cosine Transform, Ziad M. Hafed and Martin D. Levine, International Journal of Computer Vision 43(3), 167–188, 2001, c _ 2001 Kluwer Academic Publishers, Netherlands [7] Image Retrieval using Shape Feature, S.Arivazhagan, L.Ganesan, S.Selvani dhyananthan , International Journal Of Imaging Science And Engineering (IJISE) , ISSN: 1934-9955,Vol.1,No.3, July 2007. [8] Appearance Based 3D Object Recognition Using IPCA-ICA, V.K Ananthashayana and Asha. V, International Archives of the Photogrammetry, Remote Sensing and spatial Information Sciences. Vol. XXXVII. Part B1, Beijing 2008. [9] Statistical Appearance Models for Automatic Pose Invariant Face Recognition, M.Saquib Sarfraz and Olaf Hellwich, 978-1-4244-2154-1/08/ c2008 IEEE. [10] Localized Content Based Image Retrieval, Rouhollah Rahmani, Sally A. Goldman, Hui Zhang, Sharath R. Cholleti, and Jason E. Fritts IEEE Transactions On Pattern Analysis And Machine Intelligence, Special Issue, November 2008.
FIG 5.6 INDEXED
OUTPUT FOR
EYE IMAGES
6. CONCLUSION A novel method for indexing the image databases employing the wavelet features obtained from the coefficients of an image has been presented. The experimental results obtained is good in Y Cb Cr color spaces and provides better retrieval results. The Haar wavelets are chosen since they are more effective in texture representation compared to other wavelets. From the Table 5.1 and 5.2 , the robustness of the proposed feature set for indexing image data base is proved. The experimental results demonstrate the worth of the proposed method for indexing of image databases and the results are encouraging. It can be enhanced in future as an intelligent system by employing genetic algorithm or neural Networks. 7. REFERENCES [1] Similarity-Based Image Browsing, Chaomei Chen, George Gagaudakis, Paul Rosin Brunel University, Cardiff University, U.K, funded by the British research council EPSRC (GR/L61088 and GR/L94628). [2]
Interactive Content-based Image Retrieval in the Peer-to-peer Network Using SelfOrganizing Maps, Zhirong Yang, HUT T-110.551 Seminar on Internetworking 2005- 04-26/27 [3] Principal Component Analysis for Content-based Image Retrieval, Usha Sinha, and Hooshang Kangarloo, Radio graphics 2002; 22:12711289.), © RSNA, 2002
[11] Wavelet Based Features for Color Texture Classification with Application to Cbir P.S.Hiremath, S. Shivashankar, and Jagadeesh Pujari, IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.9A, September 2008 [12] Content-based Image Retrieval, John Eakins, Margaret Graham, University of Northumbria at Newcastle
Precision rate %
Recall Rate %
CBIR
90.2
92
FGNET
85.3
89
IRDS
85
90
Database
TABLE 5.1 COMPARISON OF EXPERIMENTAL RESULTS NOSE
Database
Success
Failure rate
rate in %
in %
CBIR
90.2
9.8
FGNET
85.3
14.7
IRDS
85
15
[4].Content-Based Image Retrieval, http://www.cs.cmu.edu/afs/cs/academic/class/15385-06/lectures/.
Velammal College of Engineering and Technology, Madurai
Page 283
Ta
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Architecture Evaluation for Web Service Security Policy Joshi.vinayak.B [1]
Dr. Manjaiah D. H [2]
Hanumathappa.J [3]
Research scholar
Reader Computer Science
Lecture, computer Science
Mangalore University
Mangalore University
University of Mysore
Mangalore, India [email protected]
Mangalore, India [email protected]
Mysore, India [email protected]
Abstract Web Services can be published, discovered and invoked over the web. Web Services can be implemented in any available technology but they are accessible through a standard protocol. With web services being accepted and deployed in both research and industrial areas, the security related issues become important. In this paper, architecture evaluation for web service on negotiating a mutually acceptable security policy based on WSDL (web service description language). It allows a service consumer to discover and retrieve a service-provider’s security policy for service requests and allows a service-consumer to send its own security policy for service responses to the service-provider. The service consumer combines its own policy for service requests with that of the service provider to obtain the applied security policy for requests, which specifies the set of security operations that the consumer must perform on the request. The combining takes place in such a way that the applied security policy is consistent with both the consumer’s and provider’s security policies. The service provider also combines its own policy for responses with that of the consumer, to obtain the applied security policy for responses [1].
Nayak.Ramesh.Sunder [4] A.P, ShreeDevi Institute of Technology kenjar, Mangalore, India [email protected]
have more control for the access and the updating of information, and the reliability of the registry content. We concentrate here on one key issue, providing security in Web services architecture. In this paper, we evaluated a technique for deriving mutually acceptable quality of protection for exchanges between a service provider and a service consumer. The WSDL document of a web service would include a security policy description representing the types of security operations that are required and supported by the Web-service for its SOAP message exchanges with consumers. However, since Web-service consumers may themselves be implemented as a web service, both the consumer and provider of the service may have a security policy defined in their WSDL documents. This implies that there is a need for the web service consumers and providers to agree on the applied security policy to be used to protect the SOAP message exchange between them [6].
2. System Overview 1. Introduction
Web services are reusable Web components with their programmatic interfaces described in WSDL [4].WSDL is a XML format standard for describing the interface of a web service. The WSDL description gives information about what exactly a web service does, how to invoke its functions and where to find it. Universal Description, Discovery, and Integration (UDDI) is a registry standard, which allows organizations to publish and discover Web Services using standardised methods. The UDDI is an industry initiative to provide a platform-independent framework for creating a UDDI Business Registry [9]. There are currently several providers of UDDI registers called UDDI Operators. The UDDI specification defines a set of datastructures and an Application Programming Interface (API) for registering and finding businesses. The UDDI specification also allows organizations to create their own UDDI registries in order to
Velammal College of Engineering and Technology, Madurai
2.1 Agent Based Web Service Discovery Web service discovery can be performed based on a web service security policy using agents. It consists of a ser-vice provider, a service consumer and a UDDI to include a discovery agent and security agent and use an augmented UDDI that contains security policy information to allow secure web service discovery (as shown in Figure1). The discovery agent acts as a broker between a service consumer, a UDDI registry and a security policy that helps to discover secure web services that satisfy the consumer security requirements.
2.2 Security Agent The security agent describes the security requirement that service provider needs to be registering their WSDL into the
Page 284
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
registry. Web service security test case describe a testing methodology for web service security and outline a process that can be adopted to evaluate web service security requirements. Test case can be classified according to different categories of threat faced by web services. Security policy can be represented in the UDDI registry by a tModel, which is typically used to specify the security policy details of a web service. In the tModel, each security policy is represented by a keyed Reference, which contains the name of a security attribute as key Name, and associated method.
3. Process Model The model works with the exception that the containers hosting the consumer and provider classes emit a SOAP message, which is intercepted by the security service. The consumer and provider classes could provide the <Security Mechanisms> and <Security Services> elements to their security services, in a WSS header, with the security service module identified as the target role. Alternatively, the security service could obtain the <Security Mechanisms> and <Security Services> elements directly on its own. WSDL binding to support the publication of the security policy in the case that a provider offers a secured interface. Specifically, elements called <Security Mechanisms> and <Security Services> are associated with message definitions in the service’s WSDL instance. In addition, we specify a WSS header for conveying the consumer’s policy for service responses using the same element definitions. The <Security Mechanisms> element describes a set of security mechanism, which may be applied to one or more nodes of the SOAP document. Additionally, parameters of a security mechanism may be specified in the element [1].
Figure 1. Web service discovery using agents 2.2 Discovery Agent A discovery agent receives service requests containing specifications for functional and security requirements from the service consumer, finds the services that meet the specified criteria, and then returns a list of services to the consumer in the order of priority. Discovery should be based on web service security polices for concerned request. The list of available services will be return to the service consumer in order. This avoids the overhead of discovery mechanism to search secure web services over UDDI registry for consumers needs. This approach that allows a service consumer to discover and retrieve a service provider’s security policy for service requests and allows a service consumer to send its own security policy for service responses to the service provider. The service consumer combines its own policy for service requests with that of the service provider to obtain the applied security policy for requests, which specifies the set of security operations that the consumer must perform on the request. The combining takes place in such a way that the applied security policy is consistent with both the consumer and provider security policies [1]. Likewise, the service provider combines its own policy for responses with that of the consumer, to obtain the applied security policy for responses.
Velammal College of Engineering and Technology, Madurai
Figure 2. Model for web service security policy
Page 285
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Input: User request with specified security criteria Output: Secure match set of services from UDDI u(h): Select all the services which matches the functionality requirements of user request that exists in UDDI. Let u(h)={ws1,ws2…..wsn}wss (h): Choose the set of services which have been registered in UDDI with security specifications. Let wss(h)={ws1(s), ws2(s), ….wsi(s)} Step 1 : For each web services wsi in u(h) //find the services that match the QOS requirements Step 2: QoS based Selection=Qos_Match (u(h) , QWS Parameters); Step 3 : If wss(h) requirements specified then Step4 :{Secuirty_Search=Security_Match (QoS_Search,wss(h) specified); Step5 : If wss(h) ratings found then //find the services that matches security criteria Step6 : return output of available services in wssi in u (h) according to criteria rank} Step7 :{Else return the output of available services wsi in u (h)}
step), testing (2step), and reporting (1step). It generates a number of outputs such as: a prioritised list of quality attributes, a list of architectural decisions made, a map linking architectural decisions to quality attributes, lists of risks and non-risks, and lists of sensitivities and tradeoffs. 4.2 ATAM Process Phase 1 Step 1 - Presenting the ATAM Process ATAM stands for Architecture Tradeoff Analysis Method. It is a method that tells how well an architecture satisfies particular goals by providing insight of how quality goals interact and how they trade off. Step 2 - Present Business Drivers •Due to the increase of business-to-business communication between different organizations over internet resources, the current architecture will provide secure service connection establishment between service consumer and provider with added security policy. •Suggest the service provider to accept the service consumer requirements to add new security features to perform secure tasks. Architecture Drivers The major quality attribute are as below Pri ority 1
Quality Attribute Driver Security
The proposed architecture is evaluated by the Software Architecture Tradeoff Analysis Method (ATAM).All the scenarios corresponding to each application of the secure web service discovery and retrieval are listed and evaluated. The steps corresponding to ATAM for evaluating the web service security are described as follows
2
Interoperability
4.1 ATAM: Secure Web Service Discovery
3
Availability
We put ATAM to the test on our architecture and discuss the findings based on the outputs generated which include lists of risks, non-risks, sensitivities, and tradeoffs made. The findings show that secure web service discovery and retrieval architecture can greatly benefit from using ATAM by revealing its strengths and weaknesses before evolving the architecture further. ATM comprises of 9 steps grouped into four phases: presentation (3step), investigation and analysis (3
4
Performance
Figure 3. Service discovery algorithm 4. Evaluation of Proposed Architecture
Velammal College of Engineering and Technology, Madurai
Rationale It is a major concern to this area of the architecture because it should support authentication, encryption, Integrity and non reputation over different communication channel and platform model. The registry must be able to interact with the provider servers to be able to operate. The service should be in need to run at any time even system failure occurs over registry or service provider. Continues user request will affect the system response. we will establish the user connection based on
Page 286
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
token request.
Step 3 - Presenting the ATAM Process Step 4 - Identify Architecture Approaches
Integrity
Authentication
Non-reputation
Layering
Rationale
Trade-offs
It organizes the system in hierarchical structure that allows for easy system modification.
Security potentially reduced risk
Scenario#: 1 Attribute(s) Environment Stimulus Arch decision Sensitivity Tradeoff
Step 5 - Quality Attribute Utility Tree I=Importance D=Difficulty to achieve H, M, L = high, medium, low
Risk Quality Attribut e Performance
Attribute Refinement
Scenarios
(I, D)
Response Time
Overheads of trapping user events must be imperceptible to users. At peak load the service should be able to perform transaction with limited time. Users' information shall only be visible to
(M, L)
Throughput
Security
Confidentiality
(H, M)
(H, M)
(H, M)
Step 6 - Architecture elicitation and analysis
Important Approaches of the Secure Web Service Architectural Approach
administrative users of the system and it is encrypted before transmitting to the server. The System resists unauthorized intrusion and modification of data. This enables the user to access the service with required token It verifies the signed information from valid user
(M,M)
Non risk
Scenario#: 2 Attribute(s) Environment Stimulus
(H, L)
Velammal College of Engineering and Technology, Madurai
Response
Scenario: Separate User Interface Usability Normal operations Service Consumer can add their own required security policy to the existing security policy of service provider Reasoning The present of physical buttons and clear messages display. Complexity of the physical interface design. Due to the limited space on panel or the box, design must have to be simple and neat but understandable to the user. The label on the buttons fades out after being pressed many times. This may apply same policy to server more than once. Not apply here.
Scenario: Language neutral Security Normal operations The proposed architecture should support for implementation in any language. Changing the implementation language should not affect any components in architecture. Not apply here
Page 287
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Arch decision Sensitivity Tradeoff Risk Non risk
Reasoning It support all data format conversion between different platform Need different data format for any platform New data format cannot be supported Not apply here.
Arch decision Sensitivity Tradeoff Risk Non risk
Scenario#: 3 Attribute(s) Environment Stimulus Response Arch decision Sensitivity Tradeoff
Risk
Non risk Scenario#: 4 Attribute(s) Environment Stimulus
Response Arch decision Sensitivity Tradeoff Risk Non risk Scenario#: 5 Attribute(s) Environment Stimulus Response
Scenario: Confidentiality Security Normal operations Certificate authority has to provide security token to authenticate Intermediary has no way to read the message while establishing connection with service provider Reasoning The encryption algorithm. More computation time and resource used. Performance is the tradeoff with Security. Not apply to architecture, but the Encryption algorithm itself, if it is not complex enough, could be hacked by brute force. Not apply here. Scenario: Integrity Security Normal operations Unauthorized user without security token cannot able to access the service available in the registry Identity Certificate are required to verify the user authentication Reasoning Identity certificate Need resource to map data, Performance, but not too much. Provide certificate to user in more secret Not apply here. Scenario: Authorization Security Normal operations Service ticket has way to establish trust relationship with more than one security domain utility certificate are required to
Velammal College of Engineering and Technology, Madurai
Scenario#: 6 Attribute(s) Environment Stimulus Response Arch decision Sensitivity Tradeoff Risk Non risk
verify the user authorization Reasoning Utility certificate More computation time and resource used, Performance, but not too much. Provide certificate to user in more secret Not apply here. Scenario: Non-reputation Security Normal operations Utility has key certificate to form signed message to verify the user utility key certificate are required to verify the user sign information Reasoning Utility key certificate Need signed key information for operation response Provide certificate to user in more secret Not apply here.
Risk themes: •Heavily load on the service provider web server leads to return of result to the consumer must be unavailable. This leads to server unavailable. •Certificate Verifier should verify their security token before access the service available in the registry. While during modification of the web service by the service provider add on failure may occur. Required service may be unavailable to the consumer for particular time. Inconsistencies: •Security has no way to identify the web security policies in efficient manner. •Service provided has to provide secure interface to add new security features with the existing provided security policy for applied security policy otherwise consumer has no way to add their own security requirements. •It is risky for transferring service ticket to consumer for accessing the service that available in the registry. This may be available to other without proper criteria.
Step 7: Scenario Prioritization The following table prioritizes the Quality Scenarios for the secure web service discovery and retrieval architecture. The Scenario # refers to the scenario being referenced.
Page 288
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
5. Conclusions Universal Description Discovery and Integration has no way to identify the secure web services when multiple service providers are now providing similar functional services. With an increasing number of web services providing similar functionalities, security has become one of the most critical issues. An evaluated architecture called agent based web service discovery to automate secure web service discovery for negotiating a mutually acceptable security policy based on WSDL for both consumer and provider in dynamic nature. 6. References [13] Zahid Ahmed, Martijn de Boer,, Monica Martin, Prateek
Mishra, Dale Moberg, “Web-Services Security Quality of Protection”, Version 0.9 22 Nov 2002.
[14] Kassem Saleh and Maryam Habil, “The Security
Requirements Behavior Model for Trustworthy Software”, International MCETECH Conference on e-Technologies 2008 pp 235 - 238.
[21] Tiantian Zhang, Junzhou Luo, Weining Kong “An Agent-
based Web Service Searching Model”, 9th International Conference on Computer Supported Cooperative Work in Design Proceedings pp 390 – 295.
JOSHI VINAYAK.B is currently doing research in the computer science at Dept. of Computer Science, Mangalore University, and Mangalore. He received M.Tech. From Mysore university. and master of computer science from Karnataka university Dharwad joshi has already published 4 international conference and 5 national conference papers..He has an extensive academic, Industry and Research experience. His areas interest are Computer Networking & Sensor Networks, Mobile Communication, security and reliability, E-commerce, Internet Technology and Web Programming
[15] Janette Hicks, Madhusudhan Govindaraju, Weiyi Meng,
“Enhancing Discovery of Web Services through Optimized Algorithms” IEEE International Conference on Granular Computing 2007 pp 685 - 698.
[16] Colin Atkinson, Philipp Bostan, Oliver Hummel and
Dietmar Stoll, “A Practical Approach to Web Service Discovery and Retrieval”, IEEE International Conference on Web Services (ICWS 2007).
[17] Janette Hicks, Madhusudhan Govindaraju, Weiyi Meng,
“Search Algorithms for Discovery of Web Services”, IEEE International Conference on Web Services (ICWS 2007).
[18] Artem Vorobiev and Jun Han, “Specifying Dynamic
Security Properties of Web Service Based Systems”, IEEE Computer Society 2006.
[19] Slim Trabelsi Jean-Christphe Pazzaglia Yves Roudier,
“Secure Web Service discovery: overcoming challenges of ubiquitous computing”, Proceedings of the European Conference on Web Services (ECOWS'06).
Dr. MANJAIAH.D.H is currently Reader and Chairman, Department of Computer Science Mangalore University, Mangalore. He received PhD degree from University of Mangalore, M.Tech. From NITK, Surathkal and B.E., from Mysore University. Dr.Manjaiah D.H has an extensive academic, Industry and Research experience. He has worked at many technical bodies like IAENG, WASET, ISOC, CSI, ISTE, and ACS. He has authored more than - 45 research papers in international conferences and reputed journals. He is the recipient of the several talks for his area of interest in many public occasions and International and National conferences. He is an expert committee member of an AICTE and various technical bodies. Dr .Manjaiah. D.H’s areas interest are Computer Networking & Sensor Networks, Mobile Communication, Operations Research, E-commerce, Internet Technology and Web Programming.
[20] David Geer, “Taking Steps to Secure Web Services”,
Technology News October 2003.
Velammal College of Engineering and Technology, Madurai
Page 289
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Mr.HANUMANTHAPPA.J. is Lecturer at the DoS in CS,University of Mysore,Manasagangothri,Mysore-06 and currently pursuing Ph.D in Computer Science and Engineering, from Mangalore University under the supervision of Dr.Manjaiah.D.H on entitled “Design and Implementation of IPv4–to- IPv6 Transition Scenarios for 4G-Networks”.
Nayak Ramesh Sunder is currently working as Assistant Professor in department of Cs & E in Shreedevi Institute of technology, Kenjar, Mangalore. He received M.Tech. From Mysore University. and Bachelor of Engineering from Karnataka university Dharwad. Ramesh has already published 2 national conference papers..He has 10years of academic, and Research experience. His areas interests are Computer Networking & Mobile Communication, security and reliability, E-commerce, Internet Technology and Web Programming
Velammal College of Engineering and Technology, Madurai
Page 290
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Rule Analysis Based On Rough Set Data Mining Technique P.Ramasubramanian Professor & Head, Dept. of CSE Dr.G.U.Pope College of Engg., Sawyerpuram – 628251. Tamilnadu – India. [email protected]
V.Sureshkumar Professor, Dept. of CSE Velammal Engg. College, Chennai. Tamilnadu – India. [email protected]
P.Alli Professor & Head Dept. of CSE Velammal College of Engg & Tecchnology, Madurai. Tamilnadu – India. [email protected]
ABSTRACT Rough Set theory is a new mathematical tool to deal with representation, learning, vagueness, uncertainty and generalization of knowledge. It has been used in machine learning, knowledge discovery, decision support systems and pattern recognition. It is essential for professional colleges to improve competitive advantage and increase in placement performance. Students are a key factor for an institution’s success. In this paper, we use a ROSE system to implement the rule generation for students placements. Keywords: Data Mining, Decision Support, Human resource development, Knowledge Discovery, Rough set theory.
1. INTRODUCTION Many researchers provide important methods for human resource management. Since 80’s people have been product-oriented; most employees just need skills to produce, which is the idea of job-based human resource management. In 1973, the evaluation idea and technology was "work as the center". Today students’ do not need to pursue only their studies but also to undergo soft skills training. The objective of this research is to find out a compromising solution that satisfies both the institution and students, and finds out what kind of features and behaviors of students who can build good relationships with an institution. The results of this research will be used to guide institution as they are useful in providing a good strategy for human resource development and customer relationship management This paper is organized as follows: Section 2 reviews basic concepts of rough sets. The mathematical model employed here is briefly illustrated. Section 3 is spent to explain the problem tracked in this paper. Its results are discussed in Section 4. At the end in section 5 several remarks of this paper is given. 2. INTRODUCTION TO ROUGH SET
Velammal College of Engineering & Technology, Madurai
Z. Pawlak, a Polish mathematician, put forward Rough Sets Theory (RST) in 1980, which is a data reasoning method. In recent years, it has been rapidly developed in the field of intelligent information processing. Rough sets theory has many advantages. For instance, it provides efficient algorithms for finding hidden patterns in data, finds minimal sets of data (data reduction), evaluates significance of data, and generates minimal sets of decision rules from data. It is easy to understand and offer straightforward interpretation of results [3–7]. The rough sets theory has fundamental importance in artificial intelligence (AI) and cognitive science, especially in the areas of machine learning, knowledge acquisition, decision analysis, knowledge discovery inductive reasoning, pattern recognition in databases, expert systems, decision support systems, medical diagnosis, engineering reliability, machine diagnosis, travel demand analysis and data mining. Rough set theory provides a new different mathematical approach to analyze the uncertainty, and with rough sets we can classify imperfect data or information easily. We can discover the results in terms of decision rules. So in this research, we use rough set theory to analyze the human resource problem. Information System: An information system, denoted by IS, and defined by IS = (U,A,V,f), where U consists of set of finite objects and is named a universe and A is a finite set of attributes {a1,a2, …,an}. Each attribute a belongs to a set A, that is, a ∈ A. fa: U → Va, where Va is a set of values of attributes. It is named a domain of attribute a. In Table 1, there are 10 objects. The set of condition attributes is {English Medium, Financial Status, and Soft Skills), the set of decision attributes is {Placement}.
Page 291
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Example: Table 1: Sample Data Set
Objects
Attributes a1 English Medium
1 2 3 4 5 6 7 8 9 10
Yes No Yes Yes No Yes Yes Yes Yes No
a2 Financial status Low Low Medium High Medium Medium Low Medium Medium High
The value of the English Medium has two values namely Yes or No. The value of the Financial Status is calculated on the basis of Parents monthly income and additional income. The value of the Personality Development is calculated on the basis of student performances in Subject depth, Communication skills, Participation in seminar, Public, social activities, and Co-curricular activities. In the above table, U= {1,2, 3, 4, 5, 6, 7, 8, 9, 10} A = {a1, a2, a3} a1 = {Yes, No} a2 = {Low, Medium, High} a3 = {Low, Medium, High} D = {Low, Medium, High} Lower and Upper Approximations A method to analyze rough set is based on the two basic concepts that are lower and upper approximations. Let U be any finite universe of discourse. Let R be any equivalence relation defined on U. The pair (U,R) is called the approximation space. This collection of equivalence classes of R is called as knowledge base. Let X is a subset of elements in universe U, that is, X ⊂ U. Now, we consider a subset B in Va, i.e., B ⊆ Va. The lower approximation of P, denoted as PX , can be defined by the union of all elementary sets xi contained in X.
Velammal College of Engineering & Technology, Madurai
a3 Soft Skills
Decision D1 Placement
Medium Low High Low Low Medium High High Low Medium
High Low Medium Low Low Medium High High Medium Low
The upper approximation of P, denoted as PX , can be defined by a non-empty intersection of all elementary sets xi contained in X. The lower and upper approximations are defined respectively as follows:
{ } PX = {x i ∈ U [x i ]ind (p) ∩ X ≠ φ}
PX = x i ∈ U [x i ]ind (p ) ⊂ X B X = {X ∈ U | K B ( x) ⊆ X }
B X = {X ∈ U | K B ( x) ∩ X ≠ φ } where xi is an elementary set contained in X, 1,2,…,n. The ordered pair
i =
(PX, PX ) is called rough set. In
general, PX ⊆ X ⊆ PX. If PX = PX then X is called exact. The lower approximation of X is called the positive region of X and is denoted by POS(X) and the complement of upper approximation of X is called the negative region of X and is denoted by NEG(X). Its boundary is defined as BND(X) = PX − PX. Hence, it is trivial that if BND(X)= ϕ , then X is exact. In the theory of rough sets, the decision table of any information system is given by T=(U,A,C,D), where U is the universe of discourse, A is a set of primitive features, C and D are the subsets of A, called condition and decision features respectively.
Page 292
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
For any subset P of A, a binary relation IND(P), called the indiscernibility relation is defined as IND(P)={(x,y) ∈ U × U : a ( x) = a ( y ) for all a in P} The Rough set in A is the family of all subsets of U having the same lower and upper approximations. Usually rough set theory are explored using the decision tables i.e., information tables. Core and Reduct of Attributes: Core and reduct attribute sets are two fundamental concepts of a rough set. Reduct can minimize subset and make the object classification satisfy the full set of attributes. The core concept is commonly used in all reducts [11]. Reduct attributes can remove the superfluous attributes and give the decision maker a simple and easy information. There may be more than one reduct attributes. If the set of attributes is dependent, we are interested in finding all possible minimal subsets of attributes which have the same number of elementary sets. The reduct attribute set affects the process of decision making, and the core attribute is the most important attribute in decision-making. If the set of attributes is indispensable, the set is called the core, which is defined as
RED( P) ⊆ A COR(B) = ∩ RED(P) Decision Rules: The information system in RST is generally expressed in terms of a decision table. Data in the table are got through observation and measurement. Its rows represent objects, and its columns represent attributes divided into two sets, i.e, a condition attributes set and a decision attributes set. RST reduces the decision table, and then produces minimum decision rules sets. The general form of RST rules can be expressed by RULE : IF C THEN D where ‘C‘ represents conditional attributes and their values, ‘D’ represents a decision attributes and their values. Though the decision rules we can minimize the set of attributes, reduct the superfluous attributes and group elements into different groups. In this way we can have many decision rules, each rule has meaningful features. The stronger rule will cover more objects and the strength of each decision rule can be calculated in order to decide the appropriate rules.
Velammal College of Engineering & Technology, Madurai
3. CASE STUDY ON HRD OF PROFESSIONAL COLLEGES A questionnaire has been built based on the several analysis we provided, in an educational institution. In this research we obtained answers of 25 questions of 60 students from various disciplines. These questions are named attributes which characterize each student’s. When students have good command over English language it always helps an institution to get a placement and to earn more admissions in the next academic year. Nevertheless, it is difficult to find out whether the students have a good command over English language. This issue is an important one in every institution, but for the head of an institution/ department is hard to recognize who has good command over English and his interaction with institutions, his/her class mates and faculties. Therefore, we place a focus on the answer of the question: “Do you have a good command over English with your class mates/faculty? Let us denote an answer of student i to this question (i = 1,2,...,10) as Ai . The objective of this research is to clarify what kind of pattern of other answers to the questionnaire’s results in some value of Ai , that is, “Yes, I have a good command over English” , “no, I have not a good command over English”, “High, I have a good soft skills and Personality development, and Good financial constraints”, “Medium, I have a little bit knowledge of soft skills and financial constraints” and “Low, I have no soft skills and financial constraints”. We investigate all the answers of 10 selected students based on his/her class performance in various disciplines and find out the latent structure of the answers as head of the institution and institution can provide effective functions to motivate their students. In the analysis, we firstly processed the data that are obtained from questionnaires by ROSE [8,9]. Table 1 shows the lower and upper approximations obtained by a rough set analysis. This result has the accuracy of 1.000. This means the target set is definable on the basis of an attribute set [3,4]. Attribute Ai ( i = 1, 2, …, 10) “ Do you have a good command over English with your teachers and class mates?” is named as a decision attribute. The values that Ai (i = 1, 2, …, 10) takes are 1 (yes) and 2 (no). There are 7 YESs and 3 NOs. Figure 2: ROSE with the result of decision rule
Page 293
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Class number 1 2
Number of objects 3 7
Lower approx. 3 7
Upper approx. 3 7
accuracy 1.0000 1.0000
Table 2 : Lower and Upper Approximations Table 2 illustrates that the upper and lower approximations are equivalent. Therefore, there is no uncertainty in the classification of classes D =1 and D = 2. When decision rules are obtained, the decision rules can help a decision maker to get more information about human resource.
students’ behavior, in this way the head of an institution can differentiate features into many groups and each group of its own policies. Through the ROSE, we can obtain the results, which are shown in figure 2.
Core and Reducts In rough set analysis, the core contains the attributes that are the most important one in the information table. In this study all three condition attributes are included in the core. This indicates that any attribute that is necessary for perfect approximation of the decision classes and removal of any of them leads to the decrease of the quality of approximation. Reducts, also called minimal sets, contain no redundant information. In this study, 4 reducts, shown in Table 3, were obtained for the coded information table. The length of reducts is 2~3 and is smaller than that of the original set. They represent a reduced set of attributes that provide the same classification quality of the decision attribute as the original set of condition attributes. Table 3: Reducts for the coded information Table No. Reduct {English, Soft Skills} #1 {English, Finance} #2 {Finance, Soft skills} #3 {English, Finance, Soft Skills #4 Decision Rules induction We used LEM2 methodology in ROSE system to reduce the information table and got 7 rules. Rough set analysis provides computation intelligence to the problem of classification. Each rule has its own elements. Those elements are the features of these rules. We want to find out the typical rules that can cover most of the students and help decision makers to get the ideal picture of
Velammal College of Engineering & Technology, Madurai
There are 5 decision rules that have good placement with organizations and 2 decision rules that have bad placement with organizations. This decision table stands for students’ behaviors in the English Language communications, Financial stability of parent/Guardian, and Soft skills development in an institutions. More than half of the students in an institution, have better communication skills, soft skills and such features, when they have good placement in organizations. 4. RESULTS AND DISCUSSIONS From the rules 6 and 7 we can understand that when students have better placement opportunity in an institution in the IT corporations, they also have positive thinking and behaviors. They understand a whole system of an institution and contact with people of their teachers to learn more soft skills, personality development, and improve more communication skills. So, the head of an institution should pay attention on how to make them contribute continually and increase their motivation to make more placement opportunity for their students. In rules 1 and 2 the students with bad communication and soft skills development with their teachers, classmates and public. Adequate training is required for those students. Those who do not improve his/her skills, they never get a placement. They are advised to attend various lectures or seminars and carrier-up, and knowledge of various training etc. Those negative features mean that students do not care about their jobs and do not want to make achievements. The institution of such students must find out the reason and push them back to work. Those positive and negative decision rules let us
Page 294
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
know what kind of features and behaviors can bridge between institution and placement. Students have good relationship with staff and engage themselves to gain recognition and have activities that give the person a sense of contribution. As students try to become better, they work harder to find out their abilities. Those students always want to learn new things and know more information about their technology. Therefore, such students will play a central role in their placements. Students, who have a bad relationship with teachers, only care about themselves whether they will be employed or not. It means that, they think everything for themselves. Therefore, they do not want placement. They also do not think that they are students. Therefore, those students have no interest in doing anything out of their duty. The head of the institution has to assure placement of their students. In this way, an institution can have effective staff. The institution also has to take care of the students who have good relationship with teachers, help them do their work correctly, sometimes give them some challenges and rewards to increase their motivation. No matter whether the students have good or bad relationship with teachers, they are the human resource of an institution. So, the head of the institutions can give those two groups with different functions. But it is always not easy to divide people into groups, so the heads of the institutions can recognize their staff according to the result of rough set theory. In this research we give a clear function to make students differentiable. 5. CONCLUSION In summary, rough sets theory is a new method for analyzing, inducing, studying and discovering of data. It puts forward a method of student placement assistance based on rough set theory. We have collected data for one college corresponding to the attributes and formulated the problem of finding placements using ROSE system. This formulation can be extended to a large number of colleges. The rule analysis in ROSE system for placement of students is equivalent to the rule generation in rough set theory that was presented in[10]. REFERENCES [1]. Azibi, R., and Vanderpooten, D. “Construction of rule-based assignment models,” European Journal of Operational Research, Vol. 138, No. 2. Pages 274-293, 2002.
Velammal College of Engineering & Technology, Madurai
[2]. Cascio, W.C. Managing Human Resources: Productivity, Quality of Work Life, Profits, McGraw- Hill, New York, 1992. [3]. Pawlak, Z. “Rough sets,” International Journal of Computer and Information Science, Vol. 11, No. 5. Pages 341–356, 1982. [4]. Pawlak Z., Rough Sets, Kluwer Academic Publishers, 1991. [5]. Pawlak, Z, “Rough set and data analysis,” Proceedings of the Asian 11-14 Dec.. Pages 1 – 6, 1996. [6]. Pawlak, Z.. “Rough classification,” Int. J. Human-Computer Studies, Vol. 51, No. 15. Pages 369-383, 1999. [7]. Pawlak, Z. , “Decision networks,” In: Rough Sets and Current Trends in Computing by Shusaku Tsumoto, Roman Slowinski, Jan Komoroski, Jerzy W. Grzymala-Busse (Eds.), Lecture Notes in Artificial Intelligence (LNAI), Vol. 3066, No. 1. Pages 1–7, 2004. [8]. Predki, B., Slowinski, R., Stefanowski, J., Susmaga, R. and Wilk, Sz., “ROSE - software implementation of the rough set theory,” In: L.Polkowski, A.Skowron (eds.), Rough Sets and Current Trends in Computing, Lecture Notes in Artificial Intelligence, Vol. 1424, Springer-Verlag, Berlin. Pages 605-608, 1998. [9]. Predki, B., and Wilk, Sz., “Rough set based data exploration using ROSE system,” In: Z.W. Ras, A. Skowron (eds.). Foundations of Intelligent Systems, Lecture Notes in Artificial Intelligence, Vol. 1609. Springer-Verlag, Berlin. Pages 172-180, 1999. [10]. Ramasubramanian.P, Iyakutti.K, Thangavelu.P, “Mining Analysis of Computing Rule Confidence in Rough Set Theory”, Proceedings of National Conference on Networks, Image and Security, Organized by Noorul Islaam College of Engineering in association with The Institution of Engineering and Technology, YMS – Chennai Network, Pages .102 – 106, 2008. [11]. Shinya Imai et. Al: Rough set approach to Human Resource Development, vol. 9, No.2, May 2008. About the Authors P.Ramasubramanian, is Professor and Head in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. He obtained his Bachelor and Master degree in Computer Science and Engineering from M.K.University, Madurai in the year 1989 and 1996 respectively. He is doing Ph.D in Madurai Kamaraj
Page 295
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
University, Madurai. He has over 21 years of Teaching Experience and authored 15 books and 20 research papers in International, National Journals and Conferences. His current area of research includes Data Mining, Data Ware housing, Neural Networks and Fuzzy logic. He is a member of various societies like ISTE, International Association of Engineers, Computer Science Teachers Association, International association of Computer Science and Information Technology and Fellow in Institution of Engineers (India). Email: [email protected], [email protected]. V.Sureshkumar is Professor in the Department of Computer Science and Engineering in Velammal Engineering College, Chennai, Tamilnadu, India. He obtained his Bachelor degree in Elecronics and Communication Engg., from M.K.University and Master degree in Computer Science and Engineering from Manonmaniam Sundaranar University, Tirunelveli in the year 1991 and 2002 respectively. He is doing Ph.D in Madurai Kamaraj University, Madurai. He has over 18 years of Teaching Experience and published 12 research papers in International, National Journals and Conferences. His current area of research includes Fuzzy logic, Neural Networks and Data Mining. He is a life member of ISTE. Email: [email protected]
P.Alli received her Ph.D degree in Computer Science from Madurai Kamaraj University, Madurai, India. She obtained her B.E degree in Electronics and Communication Engineering from Madras University and M.S. in Software System from BITS, Pilani. She worked as Professor and Head, Department of Information Technology, Faculty of Engineering, Avinashilingam University, Coimbatore for 12 years. She is at present Professor and Head, Department of Computer Science and Engineering , Velammal College of Engineering and Technology, Madurai. Her research interests include statistical image processing, image segmentation, image enhancement, medical image analysis and novel image reconstruction algorithms. She is especially specializing in the development of adaptive filtering technique techniques for magnetic resonance imaging. Email: [email protected]
Velammal College of Engineering & Technology, Madurai
Page 296
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Robust Security metrics for the e-Healthcare Information Systems Said Jafari#1, Fredrick Mtenzi#2, Ronan Fitzpatrick#3, Brendan O’Shea#4 School of Computing, Kevin Street, Dublin Institute of Technology, Ireland. 1
[email protected] 2 [email protected] 3 [email protected] 4 [email protected] Abstract – The application of computer based and communication technologies in the healthcare domain has resulted in a proliferation of systems through healthcare. These systems whether in isolation or integrated, are generally referred to as e-healthcare information systems. Protection of healthcare information has become an increasingly complex issue with the widespread use of advanced technologies, complexity of healthcare domain and the need to improve healthcare services. The huge amount of information generated in the process of care, must be securely processed, stored and disseminated to ensure a balanced preservation of personal privacy versus its utility for overall improvement of healthcare. The general IS security protection approaches are available to address a wide range of security concerns. However, the healthcare domain has unique security concerns that have to be uniquely addressed. Understanding these specific security and privacy challenges shall enable selection, development, deployment and assessment of IS security approaches relevant for the healthcare domain. One of the primary tasks should be identifying the unique security concerns in the domain. This paper discusses among others the unique security and privacy concerns in the healthcare domain and proposes metrics that address the challenges presented. The main contribution of the paper is the demonstration of the development of the security metrics for the healthcare domain. Keywords – security metrics, security measurement, e-health security, security assessment, e-healthcare security.
VII.
I. INTRODUCTION
The application of computer based and communication technologies in the healthcare domain has resulted in a proliferation of systems through healthcare. These systems whether in isolation or integrated, are generally referred to as e-healthcare information systems. They range from clinical information systems such as laboratory, radiology, surgical, doctor consultation, critical care, to hospital administration routines such as stores management, facilities management, financial management, and human resource management. Among the key motivation that led to embrace IT in healthcare include the need to reduce medical errors, controlling escalating healthcare costs and providing
Velammal College of Engineering & Technology, Madurai
efficient healthcare services [1]-[3]. However, digital information has much higher security risks than its counterpart paper based. Once information is digitised, processing (copying, editing, deleting, searching for a particular data of interest, linking of data attributes to determine a particular meaningful pattern) becomes easier and timeless. Additionally, the physical presence of an invader which was once a significant factor is no longer a necessary condition for launching an attack. The magnitude of security risks in digital information rose faster. This is a general observation that cut across all elements of the Information Technology Functional Services Ecosystem (ITFSE) described in section II. Information systems that are 100% secure do not exist, at least to date. This is because “software” which is a life blood of any information system is ‘buggy’ [4]-[6]. The discussion on the topic of insecurity of software; reasons and consequences are well documented in the literature [7], [8], and are not the aim of this paper either. Also, there are human users in the security chain who are referred to as the weakest link in the chain [9], [10]. The general IS security approaches are available to address a wide range of security concerns. However, the healthcare domain offers unique security concerns that have to be addressed specifically. Understanding these specific security and privacy challenges shall enable development, selection, deployment and assessment of IS security approaches relevant for the healthcare domain. The “relevance” as used in this context refers to the “best possible options” that consider the security risks of the domain in order to provide reasonable, cost-effective and justifiable protection. One of the primary tasks should be identifying the unique security concerns in the domain. This paper discusses the unique security and privacy concerns in the healthcare domain and provides metrics that capture the challenges presented. The main contribution of the paper is to demonstrate the development of security metrics for the healthcare domain.
Page 297
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The rest of the paper is organised as follows: Section II discusses the holistic view of security as mapped from the proposed IT Functional Services Ecosystem. Section III presents a review of literature that addresses security and privacy issues in healthcare information systems. Section IV outlines and explains specific e-healthcare security and privacy challenges. Section V covers issues related to measurements and metrics. It includes examples of the metrics derived to address some of the presented challenges. Section VI provides conclusion and future work. VIII. IX. II. IT FUNCTIONAL SERVICES ECOSYSTEM (IT-FSE) AND SECURITY THINKING
Information Technology is the hub that connects nearly all human activities. This is illustrated in figure 1. The proliferation of automation and advancement in networking technologies has affected the way we look at implemented systems in different domains. Systems which were once considered different now have lots in common, particularly in the security context. This is because the same information infrastructures for data storage and data communication are used across them. In that way security problems that affect one entity in the IT-FSE can affect other entities as well. Based on this view, it is fair to have general security approaches that address security concerns as experienced by all entities. Tools were also developed in the same line of thinking. Firewalls, passwords, encryption algorithms, and even security protocols such as secure socket layer among others; all are designed to be application domain neutral. When thinking of technical controls, all available tools are generic. That is, as far as security is concerned, technically all systems should think alike.
In the contemporary security thinking, it is suggested that security is not merely a technical problem, so its problems cannot be solved by technical solutions alone [9], [11]. This understanding has broadened the security view to include personnel and organisational issues [9], [11]. The security field has progressed in terms of its scope and sophistication of security tools. The market is flooded with these tools for any organisation to use. However, despite this proliferation of tools, reports on security breaches increase [12], [13]. This trend suggests that a thoughtful effort should continuously be applied to address security issues particularly in application domains that involve sensitive information like healthcare domain. In the following section, a review of literature that addresses security and privacy issues is presented. The review shall lead to identification of healthcare unique security and privacy challenges. X. III. RELATED WORK Researches in healthcare security and privacy issues are considered and they are organised in four broad interrelated categories. These categories include policy models, design architectures for databases and communication networks, access control, and requirement elicitation with its modelling techniques. I. A. Policy Models In the first category, Anderson [14], [15] proposed a security policy model for clinical information systems. The model strongly considers consent propagation and enforcement to elevate privacy property. Threats and risks to the patient confidentiality, integrity and availability of the personal health information are also discussed. B. Database and Networks Architectures In the second category, Agrawal et al [16], [17] introduced the database architecture and prototype to address privacy problems for sensitive personal information. This technology is known as the Hippocratic Database, engineered in accordance to the ten designed principles proposed by author in the effort to address key issues in privacy management. Also, network architecture for healthcare are prototyped [3]. C. Access Control In the third category, Blobel [18] discussed specifications for privilege and access for healthcare. Gritzalis and Lambrinoudakis [19] propose a role-based access control security architecture to provide authentication and authorisation in web-based distributed system in the context of health care unit (HCU). In this work, security requirements for designing a secure architecture for interconnecting healthcare systems are discussed. Also, in Røstad’s PhD thesis [20] access control requirements are discussed and identify the dynamic nature of access control in healthcare as a result of the unpredictable, dynamic and individual care process.
Velammal College of Engineering & Technology, Madurai
Page 298
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
D. Requirements Engineering The fourth category is concerned with requirements elicitation. Blobel [21] presented a framework to facilitate security analysis and design for health information systems. Tools such as those described in unified modelling language (UML), object orientation and component architecture are defined to suite the healthcare environment. Also, the framework utilised generic security model and a layered security scheme to address the challenge imposed by security enhanced health information system. Cysneiros in [22], presents requirements elicitation techniques that are useful for compliance with the constraints imposed by several peculiarities intrinsic to this domain. The techniques discussed aim to assist in selecting a proper requirement elicitation approach. Matousek in [23] introduces the basic security and reliability considerations for developing distributed healthcare systems in healthcare. Based on the above review, there is large body of research specialised in healthcare domain in attempts to address various concerns. This indicates the domain has unique challenges. However, what is unique requirement other than wrapping on the “sensitivity of information” is not clearly featured in many literature reviewed. This paper includes literature [14]-[26] that in one way or another pointed and elaborated the nature of sensitivity and how to address the challenges. In the following section, a summary and discussion of these challenges are presented.
XI. IV. HEALTHCARE SECURITY AND PRIVACY CHALLENGES Healthcare information systems capture, store, process, and disseminate information across their components. This information is sensitive. Protection of this information and information systems is vital for successful healthcare delivery. Failure of protection mechanisms or inadequate
Velammal College of Engineering & Technology, Madurai
The healthcare domain in the information age, presents unique security and privacy challenges different from others [5]-[14]. These challenges as described in the following subsections, make protection of healthcare information systems a daunting task. Understanding them, provide a locus for enhancement of protection mechanisms as well as its assessment. Table 1 contains a summary of challenges and their associated security risks or threats organised into three main categories. The categories considered include information needs, nature of healthcare service delivery and laws, rules and regulations.
A. Information Needs In the first category, information need is considered. It is well known that the primary function of healthcare services is provision of care to patients and supporting public health. Patient care requires capturing of detailed information about the patient for identification, longitudinal and cross-section information for supporting evidence-based care delivery [24]. Addressing this demand becomes a significant challenge in the course of balancing protection versus accessibility of information. This could lead to information aggregation risks, a situation where more people have access to many records [14], [15]. B. Nature of Healthcare Service Delivery The second category is the nature of healthcare service delivery. The following are challenges and their associated risks for this category. The nature of work requires collaboration among multioccupations communities (e.g., physicians, nurses, technicians, and administrative staff) [25]. In some cases, duty of care is accomplished in teams (for example in operation theatres). The work is often non-routine, so it is difficult to pre-schedule events and activities [25].
Page 299
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Emergencies, exceptions and other contingencies in similar scenarios complicate standardisation of workflows and impose risks to accountability enforcement. Autonomy is another natural feature in healthcare environment. Service departments, clinicians and patients operate as autonomous entities with their own knowledge, ethics, beliefs and goals [26]. Harmonising conflicting security and privacy requirements in an autonomous culture is difficult and could be source of risks. Healthcare service utilises different technologies that generate heterogeneous data. Integrating these tools increase attack surface. Also, the sheer volume of people in healthcare environment imposes a threat to physical security [24]. C. Laws, Rules and Regulations The third category is concerned with laws, rules and regulations that healthcare organisations have to comply with. Among the sources of security requirements include governing laws, ethical rules and regulations applicable to the area of occupancy. Balancing the demand for multiple compliances is difficult and imposes significant risks to information protection controls. The discussed challenges and risks provide additional considerations during elicitation of security and privacy requirements and goals. As the focus of this paper is metrics development, the security and privacy requirements and goals are main inputs. Meaningful measurements and metrics for security are based on security goals [27]. For measurements to yield value, security requirements and goals must be properly formulated within the considerations of the unique challenges of the application domain. In the section V, security assessments and metrics for healthcare domain are considered. XII.
V. SECURITY ASSESSMENT AND METRICS
A. The Need for Metrics Healthcare stakeholders such as patients and clinicians need to have confidence in the systems they are interacting with in the duty of care. These systems have to be justifiably secure for users to have trust in them [28]. Security metrics provide insights to efforts engaged to accomplish security goals and illuminate anomalies or weaknesses in the security controls implemented [29]. They offer evidence of the security level to justify security claims and demonstrate compliances when required. The overall helpfulness of metrics covers a wide range of security and privacy decision making; particularly in the balance between security, privacy and access to information in supporting clinical operations, research, and public health, as the main concerns of healthcare domain. B. Security Metrics Clear and precise articulation of security and privacy challenges in the healthcare domain forms the basis for
Velammal College of Engineering & Technology, Madurai
understanding the security and privacy requirements. Requirements portray security goals. Good metrics are goal driven [27], [30]. Based on the unique challenges presented in table 1, security metrics for healthcare should also incorporate the unique peculiarities of the application domain. Table 2 provide examples of the metrics and criteria to be considered for measurements based on the observation made on table 1. The terms metric, base metric, derived metric and threshold as used in this paper are borrowed from the international measurement standard (ISO/IEC 27004) [31]. However, this new standard does not use the term ‘metric’, it uses ‘measure’ instead. We have taken the liberty to synonymise the terms based on our experience on usage of these terms in previous and similar standards [32], [33]. Therefore, the terms are defined as follows: 1. Metric – variable to which a value is assigned as the result of measurement. 2. Base metric – measure defined in terms of an attribute and the method for quantifying it. 3. Derived metric – measure that is defined as a function of two or more values of base measures. 4. Threshold – target, or patterns used to determine the need for action or further investigation, or to describe the level of confidence in a given result. The sample metrics and criteria presented in table 2, consider some of the challenges presented in table 1 such as those leading to information aggregation concerns and accountability. For details on how the challenges in table 1 are related to metrics in table 2, follow the superscripts cross-reference on table 1, 2 and 3. The metrics provided are those which can be captured automatically after accomplishment of care. The reasons for automating these metrics include eliminating human biases in data gathering and computations, ensuring reproducible experiments and comparison of results. These metrics (in table 2) calculate the “Information loss Factor”, a derived metric named InfoLossFactor (μ) designed to monitor suspicious use of patient information as used in the duty of care. Also, it can be used to provide evidence against claims of illegal information disclosure, or indicating weakness in security controls deployed for protecting information. To derive this metrics, the following assumptions are considered: 1. All transactions are captured in the audit trails and the integrity of the audit trails has no doubt. The system under monitoring considered is a clinical information system, in hospital settings where the number of patients attending for care is reasonably higher than clinicians. 2. Patients’ records are accessed in the duty of care involving patients whose records are accessed. This assumption may seem unrealistic; however it serves to demonstrate our proposed metrics. 3. Database maintenance functions such as update, backup are not included in the metrics proposed.
Page 300
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
To enable derivation of the formulas, terms were proposed by authors. Table 3 provides brief interpretation of terms used in table 2, source of data and technique for data extraction. The values of the thresholds provide reasonable computed value if no suspicious indication is spotted. For instance, InfoLossFactor (μ) (in this case) its threshold is 7, to represent a secure information accesses. However the computed value of μ can increase if individual thresholds of base metrics are exceeded. The increase indicates a suspicious information access may have occurred. A different explanation can be considered on the computed value of γ1 in equation (7); in a reasonable situation, its value should approach zero, indicating fewer service departments are involved in exchanging patients’ information. In the computation of μ, the value of γ1 is taken as zero (its magnitude has no significant contribution in the in the overall equation). While not addressed in this paper, setting minimum values for thresholds is of paramount. When computed values go below stated values, different insight can be spotted; either there are significant changes in work flows or the assumptions made on the formulas to compute metrics values are wrong. Also, it should be understood that not all of the challenges highlighted in table 1 can be properly addressed by technical metrics. The effort should be made to automate as many as possible. C. Metrics Evaluation Strategy Evaluation of metrics is important to any proposed metrics. The evaluation strategy for our metrics is based on the use of principles outlined. We borrowed some of these principles from Heinrich et al [34] and Reiter et al [35]. However, these principles were designed for evaluation of DataQuality metrics and authentication metrics respectively; they fit for our metrics too. A compiled list of principles is as follows: • Principle 1: Meaningful output – the metrics should strive to measure what they are intended to measure [34], [35]. • Principle 2: Feasibility – the metrics should exhibit practicality based on determinable inputs and ease of application [34], [35]. • Principle 3: Aggregation – individual values of metrics can be summed up to few values meaningfully [34]. • Principle 4: Precision – the extent that repeatable concise results can be demonstrated for several measurements taken under similar condition [36], [37]. • Principle 5: Cost effective – metrics data must be inexpensive to gather in terms of time and cost, preferably gathered automatically [29], [37]. Demonstration of these principles against our proposed metrics is beyond the scope of this paper. However, it is
Velammal College of Engineering & Technology, Madurai
evident that our metrics do comply to some of these principles straight away. As security controls, standards, laws and regulations for healthcare are developed specifically to address the challenges of this domain, the same attention should be engaged in developing assessment approaches and metrics to measure the degree of protection offered by the proposed approaches. XIII. VI. CONCLUSION Protection of healthcare information has become an increasingly complex issue with the widespread use of ICT, complexity of healthcare domain and the need to improve healthcare services. The huge amount of information generated in the process of care, must be securely stored, processed and disseminated to ensure a balanced preservation of personal privacy versus its utility for overall improvement of healthcare services. The general IS security approaches are available to address a wide range of security concerns. However, the healthcare domain offers unique security concerns that have to be addressed specifically. In this paper, we have discussed the unique challenges imposed on security and privacy of information as enhanced by an array of factors in the healthcare domain and proposed sample of the security metrics to assess the degree to which those challenges can be controlled. Understanding these specific security and privacy challenges shall enable selection, development, deployment and assessment of IS security policies, tools and approaches relevant for the healthcare domain. As part of future work, on the basis of the presented security and privacy challenges, the development of metrics and criteria to be used in the assessment of the information protection in healthcare domain will be extended. XIV. REFERENCES [1] A. Mukherjee and J. McGinnis, "E-healthcare: an analysis of key themes in research," International Journal of Pharmaceutical and Healthcare Marketing, vol. 1, pp. 349-363, 2007. [2] V. Mantzana and M. Themistocleous, "Identifying and Classifying Benefits of Integrated Healthcare Systems Using an Actor Oriented Approach," Journal of Computing and Information Technology, vol. 2, pp. 265--278, 2004. [3] M. Beyer, K. A. Kuhn, C. Meiler, S. Jablonski, and R. Lenz, "Towards a flexible, process-oriented IT architecture for an integrated healthcare network," 2004. [4] S. M. Bellovin, "On the Brittleness of Software and the Infeasibility of Security Metrics," IEEE Security \& Privacy, vol. 4, pp. 96--96, 2006. [5] J. Viega and G. McGraw, Building secure software: Addison-Wesley Boston, 2002. [6] E. Voas, F. Charron, G. McGraw, K. Miller, and M. Friedman, "Predicting how badly “good” software can behave," IEEE Software, vol. 14, pp. 73-83, 1997.
Page 301
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[7] A. C. Michael, "Who is liable for bugs and security flaws in software?" Commun. ACM, vol. 47, pp. 25-27, 2004. [8] D. Rice, Geekonomics: the real cost of insecure software: Addison-Wesley Professional, 2008. [9] M. N. Wybourne, M. F. Austin, and C. C. Palmer, "National cybersecurity research and development challenges related to economics, physical infrastructure and human behavior," Institute for Information Infrastructure Protection (I3P) 2009. [10] B. Schneier, "Secrets |& Lies," C. Long, Ed.: Wiley Publishing, Inc., 2004, pp. 255-269. [11] B. von Solms, "Information security—a multidimensional discipline," Computers & Security, vol. 20, pp. 504-508, 2001. [12] PwC, "Information security breaches survey 2006," PricewaterhouseCoopers 2006. [13] Berr, "2008 INFORMATION SECURITY BREACHES SURVEY," Department of Business Enterprises \& Regulatory Reform 2008. [14] R. J. Anderson, "A security policy model for clinical information systems," presented at IEEE Symposium on Security and Privacy, Oakland, CA, 1996. [15] R. J. Anderson, "Patient confidentiality-at risk from NHS-wide networking," Current perspectives in healthcare computing, pp. 687-692, 1996. [16] R. Agrawal, A. Kini, K. LeFevre, A. Wang, Y. Xu, and D. Zhou, "Managing healthcare data hippocratically," presented at Intl. Conf. on Management of Data, 2004. [17] R. Agrawal and C. Johnson, "Securing electronic health records without impeding the flow of information," International journal of medical informatics, vol. 76, pp. 471-479, 2007. [18] B. Blobel, "Authorisation and access control for electronic health record systems," International Journal of Medical Informatics, vol. 73, pp. 251-257, 2004. [19] D. Gritzalis and C. Lambrinoudakis, "A security architecture for interconnecting health information systems," International Journal of Medical Informatics, vol. 73, pp. 305-309, 2004. [20] L. Røstad, "Access Control in Healthcare Information Systems," in Computer and Information Science, vol. PhD. Trondheim: Norwegian University of Science and Technology, 2008, pp. 161. [21] B. Blobel, "Security requirements and solutions in distributed electronic health records," Computers and Security, vol. 16, pp. 208-209, 1997. [22] L. M. Cysneiros, "Requirements engineering in the health care domain," presented at Requirements Eng. (RE 02), 2002. [23] K. Matousek, "Security and reliability considerations for distributed healthcare systems," presented at 42nd ANNUAL 2008 IEEE INTERNATIONAL CARNAHAN CONFERENCE ON SECURITY
Velammal College of Engineering & Technology, Madurai
TECHNOLOGY (ICCST 2008), PRAGUE, CZECH REPUBLIC, 2008. [24] ISO/IEC-27799, "Health Informatics-Information security management in health using ISO/IEC 27002," International Organisation for Standardisation 2008. [25] Y. Xiao, "Artifacts and collaborative work in healthcare: methodological, theoretical, and technological implications of the tangible," Journal of biomedical Informatics, vol. 38, pp. 26-33, 2005. [26] A. C. Elliott, "Health care ethics: cultural relativity of autonomy," Journal of Transcultural Nursing, vol. 12, pp. 326-330, 2001. [27] V. R. Basili, G. Caldiera, and H. D. Rombach, "The goal question metric approach," Encyclopedia of software engineering, vol. 1, pp. 528-532, 1994. [28] J. Grimson, W. Grimson, and W. Hasselbring, "The SI challenge in health care," Communications of the ACM, vol. 43, pp. 48--55, 2000. [29] A. Jaquith, Security Metrics: Replacing Fear, Uncertainty, and Doubt: Addison-Wesley Professional, 2007. [30] J. S. Taylor, "Autonomy and informed consent: A much misunderstood relationship," The Journal of Value Inquiry, vol. 38, pp. 383-391, 2004. [31] BS-ISO/IEC-27004, "Information technology — Security techniques— Information security management measurements," International Organisation for Standardisation (ISO) 2009. [32] M. Swanson, N. Bartol, J. Sabato, J. Hash, and L. Graffo, Security metrics guide for information technology systems: National Institute of Standards and Technology, 2003. [33] E. Chew, M. Swanson, K. Stine, N. Bartol, A. Brown, and W. Robinson, "Performance Measurement Guide for Information Security (NIST Special Publication 800-55 Revision 1)," National Institute of Standards and Technology 2008. [34] B. Heinrich, M. Klier, and M. Kaiser, "A Procedure to Develop Metrics for Currency and its Application in CRM," J. Data and Information Quality, vol. 1, pp. 1-28, 2009. [35] M. K. Reiter and S. G. Stubblebine, "Toward acceptable metrics of authentication," presented at Proc. IEEE Symposium on Security and Privacy, 1997. [36] D. S. Herrmann and S. Herrmann, Complete guide to security and privacy metrics: Measuring regulatory compliance, operational resilience and ROI: CRC Press, 2007. [37] D. Rathbun and L. Homsher, "Gathering Security Metrics and Reaping the Rewards," SANs Intitute 2009.
Page 302
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Face Detection Using Wavelet Transform And Rbf Neural Network M.Madhu1, M.Moorthi2, S.Sathish Kumar3, Dr.R.Amutha4 1, 2, 3 Research Scholar, Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram, India. 4. Professor and Head of the Department Sri Venkateswara college of Engineering, Chennai, India. [email protected], [email protected]
Abstract— Face Recognition is an emerging field of research with many challenges such as physical appearance where faces does not share the same physical features, face acquisition geometry that is, how the images are obtained and what environment the face imaged, illumination, pose and occlusion and so on. The main objective of this paper is to recognize the given face image from the database using wavelets and neural networks. DWT is employed to extract the input features to build a face recognition system and neural network is used to identify the faces. In this paper, discrete wavelet transforms are used to reduce image information redundancy because only a subset of the transform coefficients are necessary to preserve the most important facial features such as hair outline, eyes and mouth. We demonstrate experimentally that when DWT coefficients are fed into a feed forward neural network for classification, a high recognition rate can be achieved by using a very small proportion of transform coefficients. This makes DWT-based face recognition much faster than other approaches. Key words: Face recognition, neural networks, feature extraction, discrete wavelet transform
Introduction Face recognition approaches on still images can be broadly grouped into geometric and template matching techniques. In the first case, geometric characteristics of faces to be matched, such as distances between different facial features, are compared. This technique provides limited results although it has been used extensively in the past. In the second case, face images represented as a two dimensional array of pixel intensity values are compared with a single or several templates representing the whole face. More successful template matching approaches use Principal Components Analysis (PCA) or Linear Discriminant Analysis (LDA) to perform dimensionality reduction achieving good performance at a reasonable computational complexity/time. Other template matching methods use neural network classification and deformable templates, such as Elastic Graph Matching (EGM). Recently, a set of approaches that use different techniques to correct perspective distortion are being proposed. These
Velammal College of Engineering & Technology, Madurai
techniques are sometimes referred to as view-tolerant. For a complete review on the topic of face recognition the reader is referred to [1] and [2]. Neural networks have been widely applied in pattern recognition for the reason that neural-networks-based classifiers can incorporate both statistical and structural information and achieve better performance than the simple minimum distance classifiers [2]. Multilayered networks (MLNs), usually employing the backpropagation (BP) algorithm, are widely used in face recognition [3]. Recently, RBF neural networks have been applied in many engineering and scientific applications including face recognition [4][5]. The paper is organized as follows. The DWT is presented in Section 2.1. Section 2.2 presents the NN classifier for face recognition. The simulation results are given in Section 3, and finally the conclusion is drawn in Section 4.
Proposed method
Wavelet Transform In this paper, we proposed to modify the DCT approach with the wavelets. The DCT has several drawbacks. Computation of the DCT takes an extremely long time and grows exponentially with signal size. In this case, we used the Haar wavelets, representing the simplest wavelet basis. We employed the non standard decomposition, which alternates between row and column processing, allowing a more efficient coefficients’ computation. The proposed algorithm uses these coefficients as inputs for the Neural network to be trained. But the DWT has a distinct advantage; The DWT, in essence, can be computed by performing a set of digital filters which can be done quickly. This allows us to apply the DWT on entire signals without taking a significant performance hit. By analyzing the entire signal the DWT captures more information than the DCT and can produce better results. The image is separated into four sub images. The bottom left, bottom-right and top-right show the high-frequency detail of the image. The top left quadrant contains the low frequency or lower detail portion of the image; we can see that most of the information is in this portion.
Page 303
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Gaussian activation function. The response of such function is non negative for all values of x. The function is defined as, f(x) = exp(-x2) The training algorithm
Figure 1.DWT sub band structure
Thus, the DWT separates the image’s high frequency components from the rest of the image, resizes the remaining parts and rearranges them to form a new ‘decomposed’ image. The coefficients of the decomposed image are given to the neural network as input for training and recognition.
Neural Networks A Neural Network is used in order to significantly reduce the computation time, false recognition rate of the system. The architecture of RBF Network consist of 3 layers they are input, hidden and output layers as shown in Fig 2.2 1
1
1
2
2
2
3
Input layer
45
Hidden layer
3
Step 1: Initialize the weights Step 2: While stopping is false do step 3 to step 10. Step 3: For each input do step 4 to step 9 Step 4: Each input unit xi= i=1----n. Receives input signals to all units in the layer above (hidden unit) Step5: Calculate the radial basis function Step 6: Choose the center for the RBF, the centers of chosen from the set of input vectors. A sufficient number of centers have to be selected in order to ensure adequate sampling of the input vector space. Step 7: The output of the unit vi(xi) in the hidden layer. vi(xi) = exp (xji ^ -xji) 2 / σi2 Where xji^ = center of the RBF unit for input variables σi2 = width of the ith RBF unit Xji = j th variable of input pattern Step 8: Initialize the weights in the output layers of network to some random value. Step 9: Calculate the output of the network Ynet= wim vi (xi)+ wo Where H= number of hidden layer Ynet= output value of mth node in output layer for the nth incoming pattern. Wim = weight between ith RBF unit and mth output node Wo = biasing term at mth output node Step 10: Calculate error and test the stopping condition. The following flow diagram shows the new approach for face detection using wavelet decomposition and Neural networks,where DWT is used for the extraction of features from facial images and recognition is done using neural network.
Output layer LOAD DATABASE
Figure 2.Architecture of RBF Neural Network
It is a multilayer feed forward network. There exists n number of input neurons and m number of output neurons with hidden layers existing between the input and output layers. The interconnection between the input and hidden layers forms hypothetical connection and between the hidden layer and output layers forms weighted connections. The training algorithm is used for the update of weights in all the interconnection.
LOAD QUERY IMAGE
WAVELET DECOMPOSITIO FEATURE TRAIN THE FACE DETECTED
The important aspect of the RBFN is the usages of activation function for computing the output. RBF uses
Velammal College of Engineering & Technology, Madurai
Page 304
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Figure 3. Flow Diagram
Conclusion
Thus for both the database and query images Haar wavelet is applied for decomposition and the wavelet coefficients are taken for calculating the features and these are taken as input to the NN for training and recognization. Finally, the recognized image is obtained.
Results The following shows the input image loaded for detection
In this paper, a new approach for face detection has been proposed by using wavelet decomposition and Neural networks,where DWT is used for the extraction of features from facial images and recognition experiments were carried out by using neural network. The ORL database images are used for conducting all the experiments. Facial features are first extracted by the DWT which greatly reduces dimensionality of the original face image as well as maintains the main facial features. Compared with the well-known DCT approach, the DWT has the advantages of data independency and fast computational speed. The presented Neural Network model for face recognition with a focus on improving the face recognition performance by reducing the input features. Simulation results on ORL face database also show that the proposed method achieves high training and recognition speed. References [1] Zhao, W.,
Figure 4. Input Image
The following shows the wavelet decomposed image of the input image
Figure 5. Wavelet decomposed face image
The following shows the simulation results of the detected face
Figure 6 Simulation result of the detected face
Velammal College of Engineering & Technology, Madurai
Chellappa R., Rosenfeld A. and Phillips P.J.: Face Recognition: A literature survey. Technical Report CART-TR-948. University of Maryland, Aug. 2002. [2] Chellappa R., Wilson, C.L., and Sirohey, S.: Human and machine recognition of faces: A survey, Proc. IEEE, 83(5): 705C740, 1995. [3] Valentin D., Abdi H., Toole, A. J. O., and Cottrell, G. W.: Connectionist models of face processing: A survey, Pattern Recognit., 27: 1209C1230, 1994. [4] Er, M. J., Wu, S., Lu, J., and Toh, H. L.: Face recognition with radial basis function (RBF) neural networks, IEEE Trans. Neural Netw., 13(3): 697C710, 2002D [5] Yang F. and Paindavoine M.: Implementation of an RBF neural network on embedded systems: Real-time face tracking and identity verification, IEEE Trans. Neural Netw., 14(5): 1162C1175, 2003. [6] Kwak., K.-C., and Pedrycz., W.: Face Recognition Using Fuzzy Integral and Wavelet Decomposition Method, IEEE Trans. on SMC-B, 34(4): 1667-1675, 2004.
Mr.M.Madhu is pursuing his Ph.D program at Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram. He completed his B.E degree in Electronics and Communication Engineering in the year 2001 and M.E Applied Electronics in the year 2006, at Arulmigu Meenakshi Amman College of Engineering, Kanchipuram, Chennai, India. He has 7 years of teaching experience and he is currently working as Assistant Professor in the department of Electronics and Communication Engineering at Rajiv Gandhi College of Engineering, Sriperumbudur, Chennai. He has published and presented many papers in National and International Conference in the area of Image processing. His research interests are Image Segmentation, Image Compression and Image detection, speech processing.
Page 305
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Mr.M.Moorthi is pursuing his Ph.D program at Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram. He completed his B.E degree at Arulmigu Meenakshi Amman College of Engineering, Kanchipuram, in Electronics and Communication Engineering in the year 2001 and M.E Medical Electronics in the year 2007 at Anna University, Guindy campus, Chennai, India. He has 9 years of teaching experience and he is currently working as Associate Professor in the department of Electronics and Communication Engineering at Prathyusha Institute of Technology and management, Chennai. He has published and presented many papers in National and International Conference in the area of Image processing. His research interests are Image Segmentation, Image Compression and Image detection.
Velammal College of Engineering & Technology, Madurai
Mr.S.Sathish Kumar is pursuing his Ph.D program at Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram. He completed his B.E degree in Electronics and Communication Engineering in the year 2001 and M.E Applied Electronics in the year 2006 , at Arulmigu Meenakshi Amman College of Engineering, Kanchipuram, Chennai, India. He has 9 years of teaching experience and he is currently working as Associate Professor in the department of Electronics and Communication Engineering at Prathyusha Institute of Technology and management, Chennai. He has published and presented many papers in National and International Conference in the area of Image processing. His research interests are Image Segmentation, Image Compression and Image detection, speech processing.
Page 306
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
An Clustering approach based on Functionality of Genes for Microarray data to find meaningful associations Selvanayaki M#1, Bhuvaneshwari V#2 #
School of Computer Science and Engineering, Bharathiar University Coimbatore-641 046, Tamil Nadu, India 1
[email protected] [email protected] Abstract— Bioinformatics is the science of storing, extracting, Molecular Function. A functionally meaningful cluster contains many genes that are annotated to a specific GO organizing, analyzing, interpreting, and utilizing information terms. The paper aims to cluster microarray data of yeast from biological sequences and molecules. Microarrays consist based on the functionalities of genes based on GO ontology to of large numbers of molecules (often, but not always, DNA) find meaningful associations of genes from the dataset. distributed in rows in a very small space. Clustering methods provide a useful technique for exploratory analysis of microarray data since they group gene with similar functionalities based on GO Ontology. In this paper, Gene Keywords— Bioinformatics, Data Mining, Microarray, GO Ontology is used to provide external validation for the clusters to determine if the genes in a cluster belong to a Ontology, Clustering. specific Biological Process, Cellular Component and 2
I. INTRODUCTION Data mining is the non-trivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data [3]. Data mining is a larger process known as knowledge discovery in databases (KDD). Data mining is the process of discovering meaningful, new correlation patterns and trends by shifting through large amount of data store in repositories, using patterns recognition techniques as well as statistical and mathematical techniques. Bioinformatics is the application of computer technology to the management of biological information. Computers are used to gather, store, analyze and integrate biological and genetic information which can be applied to gene-based drug discovery and development. The need for Bioinformatics capabilities has been precipitated by the explosion of publicly available genomic information resulting from the Human Genome Project [4].
numbers of genes interact with each other and how a cell's regulatory networks control vast batteries of genes simultaneously. It is also called a DNA chip or a gene chip. It is the technology used to obtain a genomic profile [15]. GO (Gene Ontology) is a controlled vocabulary used to describe the biology of a gene product in any organism. There are three independent sets of vocabularies or ontologies that describe the molecular function of a gene product, the biological process in which the gene product participates, and the cellular component where the gene product can be found. The ontology is represented as a network, a directed acyclic graph (DAG), in which terms may have multiple parents and multiple relationships to their parents. In addition, each term inherits all the relationships of its parent(s).
Bioinformatics and data mining provide exciting and challenging researches in several application areas especially in computer science. Bioinformatics is the science of managing, mining and interpreting information from biological sequences and structures [6]. Data are collected from genome analysis, protein analysis, microarray data and probes of gene function by genetic methods. Microarray is a new way of studying how large
Various data mining techniques can be used in microarray clustering analysis such as hierarchical clustering, classification, k-means algorithms, genetic algorithm, association rules and self organization map [20]. The objective of the paper is to use the data mining technique, k-means clustering for microarray data to group genes with similar functionalities based on GO Ontology. The proposed idea is to compare clustering of genes based on original microarray data and clusters based on functionality of genes derived using GO ontology measures.
Velammal College of Engineering & Technology, Madurai
Page 307
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The paper is organized as follows. Section 2 provides the literature study of the various clustering algorithms using microarray. Section 3 explores the k- means algorithm methodology for clustering. In section 4 the implemented results for the yeast dataset are analyzed and validated within the go ontology functionalities. The final section draws the conclusion of the paper. II. REVIEW OF LITERATURE This section provides the details of the literature study from various papers related to microarray cluster analysis methods and techniques. Microarray technology has become one of the indispensable tools that many biologists use to monitor genome wide expression levels of genes in a given organism [7]. The data from microarray experiments is usually in the form of large matrices of expression levels of genes (row) under different experimental conditions (columns). Different methods have been proposed for the analysis of gene expression data including hierarchical clustering, self-organizing maps, and k-means approaches. In [5], Cheremushkin E.S. et al. describe partitioned clustering techniques used to find a close optimal solution. This work was aimed at the developed and implementation of a clustering algorithm based on fuzzy c-means in association with genetic algorithm. In [2], Gruzdz A. et al. proposed association rule is used to discover mutations in colon cancer diseases. In [10], Sukjoon Yoon et al. identification of the change of gene expression in multifactorial diseases, such as breast cancer is a major goal of DNA microarray experiments. In [15], Bolshakova et al. cluster validity methods to estimate the number of clusters in cancer tumor datasets. This estimation approach support biological and biomedical knowledge discovery. In [12], John Sikorski et al. proposed microarray technology to determine the relationships between genes, those that are differentially or coordinately expressed under specific conditions. In [13], LI Guo-qi et al. had developed classifier method to predict unknown function of genes in the microarray dataset on different levels. In [15], Robert S.H. et al. describes a new clustering method based on Linear Predictive Coding to provide enhanced microarray data analysis. In this approach, spectral analysis of microarray data is performed to classify samples according to their distortion values. The technique was validated for a standard data set. In [8], Eisen et al. proposed hierarchical clustering is probably the most extensively used microarray data mining techniques, using one of several techniques to iteratively, starting with one gene, combine genes with their nearest neighbor, gradually building clusters and associations of clusters, resulting in a hierarchical tree. Distance between clusters is defined by the distance between their average
Velammal College of Engineering & Technology, Madurai
expression patterns. In [20], Ka Yee Yeung et al. cluster analysis was used to identify genes that show similar expression patterns over a wide range of experimental conditions in yeast. Such genes are typically involved in related functions and are frequently co-regulated. In [19], Taraskina et al. presented a clustering method used to indentifying biologically relevant groups of genes. This paper developed an algorithm of fuzzy c-means family, designed for clustering of microarray data and distance matrices with genetic algorithm as optimization. Genetic algorithm use to find a close optimal solution to the problem of clustering. Euclidian distance method is used to calculate distance between genes. In [18], Sterhan Symons et al. proposed machine learning techniques are applied to microarray data, for diagnostic purposes. Especially in cancer diagnostics microarray classification tools are used cancer subtype discrimination and outcome prediction. In [9], Gao Cong et al. has developed two new efficient algorithm, RERII and REPT, to explore the row enumeration space to discover frequent closed patterns. In [14], Manoranjan Dash et al. proposed novel feature selection method used for large number of genes. Feature selection is a preprocessing technique that finds an optimal subset of informative features from the original set. In [1], Ben-Dor et al. discussed approaches to identify patterns in expression data that distinguish two subclasses of tissues on the basis of a supporting set of genes that results in high classification accuracy. III. METHODOLOGY Gene ontology is a collection of controlled vocabularies that describes the biology of a gene product. It consists of approximately 20,000 terms arranged in three independent ontology: Biological Process, Cellular Component, and Molecular Function [17]. Gene ontology is used to provide external validation for the clusters to determine if the genes in a cluster belong to a specific Cellular Component, Molecular Function or Biological Process. A functionally meaningful cluster contains many genes that are annotated to a specific GO terms. The main idea of the paper is clustering microarray data based on functionalities. The following information is downloaded from the website. The Yeast Saccharomyces Cerevisiae Dataset is downloaded from National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO) website. The Yeast microarray data contains about 6400 Genes (e.g. YALO51W, YALO54C) with their corresponding Yeast values (0.1650, 0.2720). The file “yeastgenes.sgd” and GO ontology was obtained from the GO annotation website.
Page 308
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The proposed methodology as given in Fig. 1 is used for analyzing microarray data to infer association of genes based on the clustered results using k means algorithm. The framework consists of three phases. In the first phase is preprocessing of microarray data is done, in the second phase the genes are mapped based on functionality using GO ontology. The clustering is done in the third phase using k means algorithm for the proposed approaches on the original dataset and the dataset based on the functionality of genes. The clustering results are analyzed for finding the association between genes in the microarray dataset.
TABLE I. EXPRESSED GENES (3677)
The yeast microarray data gene is mapped with SGDgenes structure to find the genes that are expressed in microarray yeast data to map with the SGDgo. After mapping, 3677 genes were found to be expressed in the yeast dataset. B) Mapping genes based on functionality SGDann is a master structure of Yeast microarray data is shown in Fig 2. It contains the parameters namely SGDaspect, SGDgenes, SGDgo. The entire microarray data are mapped with this structure (SGDgenes) and the microarray data is splitted based on the functionality of gene using GO ontology. For each gene the corresponding functionality of genes are found (Biological Process, Cellular Component, and Molecular Function). All the genes with same functionality are grouped for the corresponding Yeast values from the microarray dataset. The three dataset are constructed with 1079 records for biological process, 1445 records for cellular component and 1153 records for molecular function from the original microarray data are shown in Fig 3, Fig 4 and Fig 5.
Fig 1. Framework to infer association of genes
A) Preprocessing microarray data The preprocessing phase consists of two main processes that are removal of noisy data and mapping of microarray data and SGD go. 1) Removal of noisy data The yeast microarray data contains 6400 genes, with null values and empty genes. In the preprocessing step the empty spots and null values are removed using the knnimpute method. After removal of empty genes from the dataset then there are about 6314 genes used for further analysis.
Fig 2. SGDann Structure
2) Mapping of microarray data and SGD go SGDann is the master structure of the yeast genes containing all the information about the yeast data with 37794 yeast related genes. The SGDgenes column is alone extracted from the structure to map with the microarray data proposed in the work. The snapshot of microarray genes with SGDgenes is shown in Table 1.
Velammal College of Engineering & Technology, Madurai
Page 309
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig 3. Biological process (1079)
C) Clustering using k-means algorithm 1) Vector Space Model Vector space model (or term vector model) is a standard algebraic model for representing documents (and any objects, in general) as vectors of identifiers, such as, for example, index terms. It is used in information filtering, information retrieval, indexing and relevancy rankings. Its first use was in the SMART information retrieval system. A document is represented as a vector. Each dimension corresponds to a separate term [10]. The definition of term depends on the application. Typically terms are single words, keywords, or longer phrases. Several different ways of computing these values, also known as (term) weights, have been developed. One of the best known schemes is tfidf weighting (tf-term frequency, idf-inverse document frequency). The vector space model similarity measure is computed as the Euclidean distance from the query to each document [11]. The main idea of subspace (vector space model) clustering is to find a subset of genes and a subset of conditions under which these genes exhibit a similar trend [16]. 2) Input matrix The microarray data after preprocessing has 3677 expressed genes with yeast values. The expressed genes values are taken as input for vector space model. The snapshot of input matrix is shown in Table 2. TABLE II. INPUT MATRIX
Fig 4. Cellular Component (1445)
The above matrix is used to find the similarity distance using k-means algorithm. The distance is calculated using the formula given below
Fig 5. Molecular Function (1153)
The k-means algorithm is implemented in third phase on these datasets to determine the association between genes.
Velammal College of Engineering & Technology, Madurai
In this equation distance examines the root of square difference between coordinates of a pair of objects x and y. 3) Distance matrix
Page 310
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The corresponding distance matrix generated for the given input matrix is shown in Table 3. TABLE III. DISTANCE MATRIX
The distance matrix is given as input for the k-means algorithm. The basic k-means algorithm is defined below. 4) K-means algorithm Pseudo code
Fig 6. Yeast Microarray Dataset
The entire microarray dataset is mapping based on the functionality. After mapping to get 1079 for biological process, 1445 for cellular component and 1153 for molecular function are shown in Fig 7.
Fig 7. Mapping gene based on functionality
VI. EXPERIMENTAL RESULTS A) Results The implemented results for the yeast dataset are analyzed and validated within the go ontology functionalities to find association between genes. The original microarray data has only genes. These genes are mapped with SGDgenes to get the expressed genes for the original microarray dataset. The number of genes in the dataset is shown in the Table 4. TABLE VI. DATASET
Dataset
Number of genes
SGDgenes (master structure) Yeast microarray data After removal of empty genes Expressed genes
37794 6400 6413 3677
Biological Process
1079
Cellular Component
1445
Molecular Function
1153
B) Analysis of K-Means Algorithm 1) Clustering on original microarray data The 20 cluster were generated on the microarray dataset and after preprocessing of microarray dataset using kmeans algorithm. The snapshot of the gene generated in the two different dataset is shown in Table 5. The count of genes was different in both the cluster generated for microarray data with and without preprocessing of microarray data.
The yeast microarray dataset contain 6400 genes. After removal of empty genes from the dataset then there are about 6314 genes are shown in Fig 6.
Velammal College of Engineering & Technology, Madurai
Page 311
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE V. CLUSTERING ON ORIGINAL MICROARRAY DATA
2) Clustering microarray data based on functionality
The same sets of microarray gene YAL054C, YAL034C, YBL043W were grouped in different clusters before and after preprocessing. The cluster 6 contains the above said genes grouped with 670 genes before preprocessing the cluster 14 contains the same genes grouped with 350 genes.
The genes are clustered based on the ontology functionality that is Biological Process, Cellular Component and Molecular Function. These clusters were generated based on GO Ontology using k-means algorithm as shown in Table 6.
The same set of gene (YAR068W, YGR207C) was clustered in Biological Process, Cellular Component and Molecular Function but with different count value. The gene YAR068W, YGR207C was clustered in the count value of 140 genes for Biological Process, 80 genes for Cellular Component and 130 genes for Molecular Function. Some of the genes are clustered only in two
functionality. For example, the gene YEL048C, YMR259C was clustered in the count value of 65, 120 genes for Biological Process and Cellular Component alone. The pictorial representations of Table 6 are shown in Fig 8.
Velammal College of Engineering & Technology, Madurai
Page 312
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig 8. Number of cluster based on different functionality
The clustering method proposed help to find association of genes across various functionality. For each functionality (Biological Process, Cellular Component, Molecular Function) GOid are used to find association between genes. Each functionality having more then one genes are shown in Table 7. TABLE 7. DIFFERENT FUNCTIONALITY GENES
The three functionalities show that different sets of clusters can be generated from the dataset using different set of genes. The first two functionalities (Biological Process and Molecular Function) are taken and compared with genes to find out the meaningful associations of genes from the dataset. By clustering and analyzing the clusters based on the two functionalities the following genes YAR029W, YBR074W, YBR096W, YBR138C, YBR235W, YBR239C, YBR241C, YBR255W, and YCR015C are found to be have association. The clusters generated for the two functionalities are shown in Fig 9.
Fig 9.Association of genes – Biological and Molecular Functionality
The association of genes clustered using two functionalities the biological and molecular with their goid is shown in Table 8. Velammal College of Engineering & Technology, Madurai
TABLE 8. ASSOCIATION BETWEEN GENES GOID
Page 313
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
transporter or pore. The goid 15075 is a molecular function. It catalysis of the transfer of an ion from one side of a membrance to the other. The goid 22891 is also a molecular function it enables the transfer of a specific substance or group of related substances from one side of a membrance to the other. The biologically meaningful unknown function is grouped based on functionalities. In our analysis of the clustered results we have found that the set of genes grouped based on the goid 8150, 3674 always occur together and they have a unknown biological and molecular function. The cluster approach based on fucntionalty helps to finds the meaningful clusters rather than grouping the original dataset. The genes were grouped based on the goid 8150 and 3674, where the first belongs to biological process and the second go belobgs to molecular function. The genes grouped with goid 8150 belong to unknown function of biological process. Which describes any process specifically pertinent to the functioning of integrated living units: cells, tissues, organs, and organisms. The gene clustered with goid 3674 is an unknown molecular function and it describes the actions of a gene product at the molecular level. The following goid 6810, 15075 and 22891 have a different functionality. The goid 6810 is a Biological Process; it describes the directed movement of substances (such as macromolecules, small molecules, ions) into, out of, within or between cells, or within a multicellular organism by means of some external agent such as a
V. CONCLUSION Grouping of Microarray data with similar functionalities in genes are grouped based on Go ontology and is implemented using the data mining clustering technique the K-means algorithm. In the proposed work the microarray data was extracted for functionality for Biological Process and Molecular Function and implemented and tested the clusters for analyzing meaningful associations. The work proposed helped to find the association of genes using the clustering algorithm based on GO Ontology and from the study we conclude that the clustering based on functionality of genes have meaningful associations than directly clustering the dataset.
[8]
REFERENCES [1] A. Ben-Dor, R. Shamir, and Z. Yakhini, “Clustering gene expression patterns” in J Comput Biol 6(34):281-97. [2] A.Gruzdz, A.Ihnatowicz, J.Siddiqi, and B.Akhgar, “Mining Genes Relations in Microarray Data Combined with Ontology in Colon Cancer Automated Diagnosis System”, Vol.16, Nov 2006, ISSN 13076884. [3] Arun. K. Pujari, “Data Mining Techniques”, Universities press (India) Limited 2001, ISBN-817371-3804. [4] Bradley Coe and Christine Antler, “ S po t yo ur g en es – an ov erview of th e micr oarr ay”, August 2004. [5] Cheremushkin E.S, “The Modified Fuzzy C-Means Method for Clustering Of Microarray Data”. [6] Daxin Jiang, Chun Tang, and Aidong Zhang, “Cluster Analysis for Gene Expression Data: A Survey, IEEE Transactions on knowledge AND Data Engineering”, Vol 16, No. 11, November 2004. [7] D. Jiang, Ch. Tang, A. Zhang, “Cluster Analysis for Gene Expression Data: A Survey,” IEEE, Vol. 16, no. 11, November 2004.
Velammal College of Engineering & Technology, Madurai
Eisen MB, Spellman PT, Brown PO, Botstein D: Cluster analysis and display of genome- wide expression patterns. PNAS 1998, 95(25): 1486314868. [9] Gao Cong, Kian-Lee Tan, Anthony K.H. Tung, Feng Pan, “Minning Frequent Closed Patterns in Microarray Data”. In Proc. ACM SIGKDD Int,l Conf. On Knoweldge Discovery and Data mining (KDD), 2003. [10] Georage Tsastaronis and Vicky Panagiotopoulo, “A Generalized Vector Space Model for Text Retrieval Based on SemanticRelatedness” Proceedings of the EACL 2009 Student Research Workshop, Pages 70-78, 2 April 2009 © 2009 Association for Computational Linguistics. [11] G. Salton, A. Wong, and C.S. Yang, “A Vector Space Model for Automatic Indexing”, Communications of the ACM, Vol. 18, no. 11, Pages 613-620. [12] John Sikorski, “Mining Association Rules in Microarray Data”, Bioinformatics, 19, 2003. [13] LI Guo-qi, Sheng Huan-ye,”Classification analysis of microarray data based on Ontological engineering”, journal of Zhejiang Univ Sci A 2007 8(4): 638-643. [14] Manoranjan Dash and Vivekanand Gopalkrishnan, “Distance Based Feature Selection for Clustering Microarray Data”, PP. 512-519, 2008 ©SpringerVerlag Berlin Heidelberg 2008.
Page 314
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[15] Nadia Bolshakova, Francisco Azuaje, Padraig Cunningham, “An integrated tool for Microarray data clustering and cluster validity assessment”, 2008. [16] Pankaj Chopra, Jaewoo Kang, Jiong Yang, Hyung Jun Cho, Heenam, Stancey and Min-GOO Lee, “Microarray data mining using landmark gene-guided clustering”, BMC Bioinformatics, 2008, 9:92, February 2009. [17] Pasquier. C. The : Ontology – driven analysis of microarray data. Bioinformatics 20(16), 2636-2643, 2004. (www.Wisegeek.com/what is microarray.htm.) [18] Stephan Symons and Kay Nieselt, “Data Mining Microarray Data- Comprehensive Benchmarking of Feature Selection and Classification Methods”, 2007. [19] Taraskina A.S “The Modified Fuzzy C-Means Method for Clustering Of Microarray Data”. [20] Yeung K, Medvedovic M, Bumgarner R: Clustering gene-expression data with repeated Measurements Genome Biology 2003.
Velammal College of Engineering & Technology, Madurai
Page 315
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
An Energy Efficient Adavanced Data Compression And Decompression Schemes For Wsn G.Mohanbabu#1, Dr.P.Renuga#2 Lecturer, Department of Electronics and Communication Engineering, PSNA CET, Anna University #2 Assistant Professor, Department of Electrical and Electronics Engineering, TCE, Anna University
#1
1
[email protected] 2 [email protected]
Abstract - Efficient utilization of energy has been a core area of research in wireless sensor networks. Sensor nodes deployed in a network are battery operated. As batteries cannot be recharged frequently in the field setting, energy optimization becomes paramount in prolonging the batterylife and, consequently, the network lifetime. The Communication module utilizes a major part of the energy expenditure of a sensor node. Hence data compression methods to reduce the number of bits to be transmitted by the communication module will significantly reduce the energy requirement and increase the lifetime of the sensor node. The present objective of the study contracted with the designing of efficient data compression algorithm, specifically suited to wireless sensor network. Keywords – Wireless sensor network, Data compression, Huffman algorithm.
I. INTRODUCTION Wireless Sensor Network (WSN) comprises of several autonomous sensor nodes communicating with each other to perform a common task. A wireless sensor node consists of a processor, sensor, communication module powered by a battery. Power efficiency is considered to be a major issue in WSN, because efficient use of energy extends the network lifetime. Energy is consumed by the sensor node during sensing, processing and transmission. But almost 80% of the energy is spent in the communication module for data transmission in sensor network[1]. Sensor networks have a wide range of application in temperature monitoring, surveillance, bio medical, precision agriculture. Failure of sensor node causes a partition of the WSN resulting in critical information loss. Hence there is great interest shown by the many researchers in extending the lifetime of sensor nodes by reducing the energy required for transmission. Several algorithms have been proposed for energy efficient wireless sensor network in literature. The spatio-temporal correlations among sensor observations are a significant and unique characteristic of the WSN which can be exploited to drastically increase the overall network performance. The existence of the above mentioned correlation in sensor data is exploited for the development of energy efficient communication protocols well suited to WSN. Recently there is a major interest in
Velammal College of Engineering and Technology, Madurai
the Distributed Source Compression (DSC) algorithm which utilizes the spatial correlation in a sensor network for data compression. WSN application requires dense sensor deployment[1] and as a result of this, multiple sensors record information about a single event. Therefore it is unnecessary for every sensor node to send redundant information to the sink node due to the existence of high spatial correlation. Instead a smaller number of sensor measurements might be adequate to communicate the information to the sink with certain reliability. In[2] a distributed way of continuously exploiting existing correlations in sensor data based on adaptive signal processing and distributed source coding principles is discussed. In[6,8] hardware architecture for data compression using adaptive Huffman algorithm for data compression is proposed. In[3,4,7] also the spatial correlation in sensor data is exploited for the data compression. Though research has progressed in the area of data compression, not many have worked in lossless algorithms for wireless sensor networks. Compression algorithms are mainly for data storage and therefore simple and application specific algorithms are required for resource constrained sensor nodes. In[5] a simple compression algorithm based on static Huffman coding particularly suited for memory and computational resource constrained wireless sensor node is proposed. This algorithm exploits the temporal correlation in sensor data and the algorithm computes a compressed version using a small dictionary. In[5], to design the dictionary, the statistics of the data are required, unlike in real time, statistics of the data will change depending on the event. Therefore the objective of the study is to design a simple algorithm to compress sensor data which does not require prior knowledge of the statistics of sensor data and the data compression is performed adaptively based on the temporal correlation in sensor data. II. MATERIALS AND METHODS To establish a modified algorithm for data compression in wireless sensor network, the following simple and adaptive algorithms were simulated and compared.
Page 316
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Simple Huffman algorithm In the simple Huffman algorithm proposed in[5], each sensor node measure mi is converted by an ADC to binary representation ri using R bits, where R is the resolution of the ADC. For each new measure mi, the compression algorithm computes the difference di = ri - ri-1, which is input to an entropy encoder. The encoder performs compression losslessly by encoding differences di more compactly based on their statistical characteristics. Each di is represented as a bit sequence bsi composed of two parts si and ai, where si gives the number of bits required to represent di and ai is the representation of di. Code si is a variable length code generated by using Huffman coding. The basic idea of Huffman coding is that symbols that occur frequently have a smaller representation than those that occur rarely. The ai part of the bit sequence bsi is a variable length integer code generated as follows: If di>0, ai corresponds to the ni lower-order bits of the direct representation of di If di<0, ai corresponds to the ni lower-order bits of the two’s complement representation of (di-1). If di=0, si is coded as 00 and ai is not represented. The procedure used to generate ai guarantees that all the possible values have different codes. Once bsi is generated, it is appended to the bit stream which forms the compressed version of the sequence of measures mi. Here <<si,ai>> denotes the concatenation of si and ai. Since transmission of a bit needs energy comparable to the execution of thousand instructions, just saving only a bit by compressing original data corresponds to reduce power consumption[5]. We observe that the proposed algorithm is simple but has an important drawback. In order to assign probability, the statistics of sensor data must be known already. Hence this algorithm may not be suitable for real time sensor data. III. ADAPTIVE HUFFMAN ALGORITHM Huffman coding requires prior knowledge of the probabilities of the source sequence. If this knowledge is not available, Huffman coding becomes a two pass procedure: the statistics are collected in the first pass and the source is encoded in the second pass. In the Adaptive Huffman coding procedure, neither transmitter nor receiver knows anything about the statistics of the source sequence at the start of transmission. The tree at both the transmitter and the receiver consists of a single node that corresponds to all symbols Not Yet Transmitted (NYT) and has a weight of 0. As transmission progresses, nodes corresponding to symbols transmitted will be added to the tree and the tree is reconfigured using an update procedure. Considering a simple 4 bit ADC representation for each data, then before the beginning of transmission, a fixed 4 or
Velammal College of Engineering and Technology, Madurai
5 bit code depending whether the symbol is positive or negative is agreed upon between the transmitter and receiver. The actual code consists of two parts: The prefix corresponding to the code obtained by traversing the tree and the suffix corresponding to 4 or 5 bit binary representation corresponding to positive or negative data respectively. In the process of coding, the probability for the incoming source sequence is assigned as the elements get in to the tree formed. This gives rise to a problem where the elements which come in the initial stages of tree formation having lesser probability hold smaller codes. Thereby, the compression ratio obtained is lesser. The pseudo code for adaptive Huffman algorithm is shown in Fig. 1. module adaptivehuffman_encode (diff, btree) createroot() search (diff, btree) if(node present) incrementnode->wt code = traverse(node) else prefix = traverse (nyt) if(diff>=0) suffix = diff4 if (diff<0) suffix = diff5 code =<<prefix, suffix>> nyt->lchild = createnyt() nyt->rchild = createnode (diff) nyt = nyt>lchild endif update (btree) balance (btree) endmodule Fig.1 : Pseudo code of adaptive Huffman encoding algorithm.
Using Adaptive Huffman algorithm, we derived probabilities which dynamically changed with the incoming data, through Binary tree construction. Thus the Adaptive Huffman algorithm provides effective compression by just transmitting the node position in the tree without transmitting the entire code. Unlike static Huffman algorithm the statistics of the sensor data need not be known for encoding the data. But the disadvantage in this algorithm is its complexity in constructing the binary tree which makes it unsuitable for sensor nodes. Modified adaptive algorithm In static and dynamic algorithms, the ultimate objective of compression was achieved with fixed and dynamic probabilities respectively. The main disadvantage of Static Huffman algorithm is that, we do not have the prior knowledge of the incoming source sequence. Also the statistics of sensor data may be varying. In Adaptive
Page 317
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Huffman algorithm, though the probabilities are assigned dynamically, because of the increased number of data available in the source sequence, the number of levels and hence the number of bits transmitted increases and is found to be effective only for very frequently and first occurring data. The binary tree construction is based on the order of arrival of incoming data. Hence, both static and adaptive Huffman algorithms are observed to have some drawbacks. The proposed Modified Adaptive Huffman algorithm overcomes the disadvantages of both static and dynamic Huffman algorithm by combining the advantages of the two algorithms and ultimately increasing the compression ratio. The algorithm uses a tree with leaves that represent sets of symbols with the same frequency, rather than individual symbols. The code for each symbol is therefore composed of a prefix (specifying the set, or the leaf of the tree) and a suffix (specifying the symbol within the set of same-frequency symbols). The Binary Tree is constructed with the incoming elements and the codes framed by traversing the tree as in Adaptive Huffman algorithm. Elements are grouped as nodes and codes are specifically assigned as in Static Huffman algorithm. At every update the weights at every level is checked and updated such that the higher weights will be occupying the initial stages of the tree. {0} NYT 1
12001000800600400200-
{-0.3, -0.2, 0.2, 0.3} {-0.1, 0.1}
0
This enhances the performance in two ways: It reduces the number of levels in the tree. It brings the maximum possible elements to the top level of the tree. Binary tree construction
50454035302520151050 Adaptive Huffman
Binary tree construction used in the algorithm is shown in Fig. 2, where each node consists of set of elements. The temporal correlation in sensor data only the difference of the sensor data di = ri-ri-1 is encoded. Consider for example the set {-0.2, 0.1, 0.3} is transmitted. Initially the encoder will have only the NYT node. Since -0.2 occurs for the first time, it is transmitted as 10010 and a new node
Fig.3 : Compression analysis in terms of No. of bits transmitted.
Velammal College of Engineering and Technology, Madurai
Modified Adaptive Huffman
Fig.2 : Binary Tree
Static Huffman
NYT
Modified Adaptive Huffman
1
Original Data
1
Adaptive Huffman
0
1
Static Huffman
NYT
{-0.3, -0.2, 0.2, 0.3}
which contains this data is inserted. When the next data 0.1 arrives, the tree is traversed in search of the node1 which contains the data 0.1 in the binary tree. Since the node is not available, the code corresponding to NYT node i.e., 0 is transmitted, followed by 1 corresponding to the data 0.1.. So the code transmitted is 01. For the next data is 0.3, the tree is traversed in search of the node 2 containing the data 0.3. Since node 2 is already available in the tree, the prefix corresponding to node traversal 1 is transmitted. The binary representation of the array index containing the data is transmitted as suffix i.e., 11 corresponding to 3 which is the array index containing the data 0.3 is transmitted. So the code is transmitted is 111. Initially the algorithm starts with subsets each having different probability. But this does not require to transmit the tree prior to decompression. In the proposed algorithm, once we reach a leaf while decoding, we know how many bits are there still to be read and that allows us to make a single operation to get a bit string which is readily converted to an integer.
Page 318
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 4 : Compression Ratio
Compression ratio Compression ratio is calculated as defined in[5]. The formula used for compression ratio analysis is given as follows: Compression ratio = 100(1- compressed size/original size) Compressed size is the number of bits obtained after compression and original size will be the total number of bits required without using compression algorithm. The performance of the modified adaptive Huffman, static and adaptive Huffman compression algorithm were analyzed in terms of compression in number of bits required for transmission is shown in fig. 3 and compression ratio of each algorithm is shown in fig. 4. IV.CONCLUSION
[2] Chou, J. and K. Ramachandran, 2003. A distributed and adaptive signal processing approach to reducing energy consumption inn sensor networks. Proceeding of the INFOCOM 22nd Annual Joint Conference of the IEEE Computer and Communications Societies, Mar. 30-Apr. 3, IEEE Xplore Press, USA., pp: 1054-1062. http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1208942 [3] Tang, C., C.S. Raghavendra and K.V. Prasanna, 2003. An energy efficient adaptive distributed source coding scheme in wireless sensor networks. Proceeding of the IEEE International Conference on Communications, May 11-15, IEEE Xplore Press, USA., pp: 732-737. DOI: 10.1109/ICC.2003.1204270. [4] Ye, Q., Y. Liu and L. Zhang, 2006. An extended DISCUS scheme for distributed source coding in wireless sensor networks. Proceeding of the International Conference on Wireless Communications, Networking and Mobile Computing, Sept. 22-24, IEEE Xplore Press, Wuhan, pp: 1-4. DOI: 10.1109/WiCOM.2006.286 [5] Francesco Marcelloni, 2008. A simple algorithm for data compression in wireless sensor networks. IEEE. Commun. Lett., 12: 411413. DOI: 10.1109/LCOMM.2008.080300 [6] Lin, M.B., J.F. Lee and G.E. Jan, 2006. A lossless data compression and decompression algorithm and its hardware architecture. IEEE Trans. Very Large Scale Integrat. Syst., 14: 925-936. DOI: 10.1109/TVLSI.2006.884045 [7] Pradhan, S.S., Julius Kusuma and K. Ramachandran, 2002. Distributed compression for dense microsensor networks. IEEE. Signal Proc. Mag.,19: 15-60. DOI: 10.1109/79.985684 [8] Cesare, A., Romolo Camplani and C. Galperti, 2007. Lossless compression techniques in wireless sensor networks: Monitoring microacoustic emissions. Proceeding of the IEEE International Workshop on Robotic and Sensor Environments, Oct. 12-13, IEEE Xplore Press, Ottawa, Canada, pp: 1-5. DOI: 10.1109/ROSE.2007.4373963 [9] http://www.mass.gov/dep/water/resources/
A simple algorithm namely modified adaptive Huffman algorithm which does not require prior knowledge of the statistics of sensor data is proposed for data compression in the sensor network. The data compression is performed adaptively based on the temporal correlation in the sensor data. This modified adaptive Huffman encoding algorithm effectively combines the advantages of static and adaptive Huffman algorithms to provide effective compression by reducing the number of levels in the binary tree. The implementation of this algorithm shows better compression ratio than adaptive Huffman algorithm. Since the number of levels is restricted, this algorithm requires less computation than adaptive Huffman which is an important requirement for wireless sensor nodes. This algorithm outperforms the static and adaptive Huffman algorithm in providing better performance in terms of number of bits transmitted. Future Work The proposed work can be implemented and tested using Beagle board advanced multi core DSP processor and the hardware results can be verified and compared with the software simulation results. V.REFERENCES [1] Akylidiz, I.F., W. Su, Y. Sankarasubramaniam and E. Cayirici, 2002. Wireless sensor networks: A survey. Comput. Networks, 38: 393422. DOI: 10.1016/S1389-1286(01)00302-4
Velammal College of Engineering and Technology, Madurai
Page 319
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Active Noise Control: A Simulation Study #
Sivadasan Kottayi#1, Narayanan N.K.*2 Electrical & Electronics Engineering Technology Department, Yanbu Industrial College, Yanbu Al-Sinaiyah, Yanbu, KSA 1
[email protected]
* School of Information Science & Technology, Kannur University, Kerala, India. 2
[email protected]
AbstractIn the last decade active noise control has emerged as a viable technology for cancellation of low frequency noise. This paper explains the basic principle of active noise control and a detailed account of broad band feed forward active noise control system together with the necessity of nullifying the secondary path as well as acoustic feedback problems. A simulation study of active noise control system for a passenger train cabin, an example of broad band ANC has been carried out using the noise signal collected from an air conditioned train cabin. Though it is a multichannel ANC problem having different types of complexities, it has been simplified with certain assumptions and simulated for a single channel using different adaptive algorithms. Among the five algorithms tried, normalized LMS algorithm is found working well in this environment.
KeywordsActive noise control, Adaptive noise cancellation, Adaptive systems, Digital Signal Processing applications I. INTRODUCTION Active noise control (ANC) is an emerging research area in engineering & technology having numerous applications. It promises viability of numerous quality products. One of the most successful ANC products in the market is active noise cancelling headset which reduces undesired noise going to human ears. This increases the work efficiency of factory & aerospace employees. The advent of high speed cost effective Digital Signal Processing chips enhances the feasibility of many ANC products [1,2,3] which in turn increases the comfort of human beings while they are at home, travelling or in office. Sound frequencies audible to human ears lie in the frequency range of 20 Hz to 20 KHz. However, only a certain band of frequencies is pleasurable to human ears. The hustle and bustle of market places, noise in buses and trains, rumble of air- conditioners, hissing of fans - all make human beings exhausted. Such disturbing noises need to be eliminated. II. SCHEMES FOR ELIMINATING NOISE There are two types of noises: high frequency noise and low frequency noise. High frequency noise can be reduced through conventional passive method [4], which employs heavy barriers to block the transmission of sound and also use certain acoustic materials to absorb sound energy. However, these techniques are found less
Velammal College of Engineering and Technology, Madurai
effective at low frequency (< 500Hz) as to stop such kind of noise one need to have barriers with large width, which are expensive and often difficult to implement. Heavy and bulky barriers are prohibited in aerospace applications. In the case of mobile telephony, it is arduous to attend the incoming telephone calls in a noisy environment. Moreover, when you make a telephone call in such an environment, the surrounding noise also propagates along with the voice signal to the destination, which is also not desirable. In all these situations conventional passive techniques cannot be applied, but noise can be eliminated electronically via active noise control. III. BASIC PRINCIPLE OF ACTIVE NOISE CONTROL Active noise control (ANC) is based on destructive interference between two sound fields: one is the original or primary sound source, which has to be eliminated and the other is secondary sound source or control signal generated by the Electronic Controller, having the same frequency of primary sound but 180° out of phase, propagated through a loudspeaker to interfere with the primary sound so that both signals get cancelled by themselves. Since sound propagation is considered to be linear and the principle of superposition is applicable to it, their acoustic pressures cancel themselves. However, how much an Active Noise Control is effective depends upon the location of the sensing microphones and the secondary signal source. It also depends upon the acoustic environment as to whether the pressure wave propagates freely in air or is enclosed within a confined space. Destructive interference is most effective when two sound fields are accurately aligned in space over an acoustic wavelength. IV ACTIVE NOISE CONTROL TECHNIQUES In ANC two types of control techniques are in practice – one is Feedback ANC and the other is feed forward ANC. Both the techniques are explained below. A. Feedback ANC Feedback control approach was first tried for active noise cancelation by Olson and May [5] in the year 1953. A physical illustration of such a system in a ventilation duct is shown in Fig. 1 and its equivalent electrical block diagram is shown in Fig. 2. In Fig. 2 ‘e’ represents the signal derived from the monitor (error) microphone, which is the combined effect of the
Page 320
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
disturbance ‘d’ and the output of feedback loop. ‘W’ was a simple gain and phase inversion circuit described by Olson and May. The electrical transfer function from loudspeaker input to microphone output ‘S’ is called secondary or error path. It contains the electro – acoustic response of the loudspeaker, electro acoustic response of error microphone and the acoustic characteristics of the path between loudspeaker and error microphone. The transfer function between the disturbance or primary source and measured error is thus
Secondary source
Monitor microphone
Primary Source
Duct
Loudspeaker
Fig. 1
Electronic controller
A single channel feedback ANC System in a duct
microphone. Noise cancelling head set, which is one of the most successful ANC products in the market works on this principle as the proximity of eardrum is close to error microphone. However, most of the other applications except purely periodic disturbances, Feedback ANC is not successful because of the restriction of placing error microphone close to secondary source. This is another major disadvantage of Feedback ANC. Therefore, Feed forward ANC is widely used for active noise cancellation. B. Feed-forward ANC Feed forward ANC can be further divided into Broad-band feed forward ANC and Narrow band feed forward ANC. In this paper we concentrate on broad band feed forward ANC as the paper study about the noise inside a passenger train cabin, which is considered as broad band. The noise we hear inside an air-conditioned train cabin is a composite noise consists of engine sound, body vibration, track sound, sound of air-conditioner and the noise made by passengers. A broad band ANC system that has a single reference microphone, single secondary signal source and single error microphone is shown in Fig. 3. Signal processing time of the controller must be equal to the time taken by the acoustic signal to reach from the reference sensor to the secondary source; or else the ANC system may not be effective in reducing the noise.
d Loudspeaker
+ S
+
W
Fig. 2
Σ
e
Equivalent electrical block diagram could be increased in the feedback path without limit causing the overall transfer function of the feedback loop to become small. The feedback loop forcing ‘e’ to be small compare to ‘d’ will be to cancel acoustic pressure at the error microphone as required for active noise control. However, the frequency response of the secondary path can never be made perfectly flat and free from phase shift. The loudspeaker introduces considerable phase shift. The acoustic path from loudspeaker to microphone involves some delay due to the acoustic propagation time and this will also introduce an increasing phase shift in the secondary path. If the phase shift in the secondary path approaches 180°, the negative feedback described above becomes positive feedback and the control system become unstable is a disadvantage of Feedback ANC. If the system is digital as it is imperative for implementing adaptive control, delay in secondary path will be much more due to the process delay of A/D converter, anti-aliasing filter, D/A converter and reconstruction filter. Due to this reason feedback ANC is normally implemented using analog circuit and secondary source (loudspeaker) should be placed very close to error
Velammal College of Engineering and Technology, Madurai
Primary Source
Secondary source
Error microphone
Reference microphone Anti-aliasing filter and A/D t
D/A converter & reconstructtion filter
Anti-aliasing filter & A/D converter
Digital filter
Fig. 3
A Feed forward Adaptive ANC Algorithm
The broad band feed-forward ANC system shown in Fig. 3 is illustrated in an adaptive system identification configuration as in Fig. 4, in which an adaptive filter W(z) is used to estimate an unknown plant P(z). The primary path P(z) consists of an acoustic response from the reference sensor to the error sensor where the noise attenuation is to be realized. If the plant is dynamic, adaptive algorithm has to continuously track time variations of the plant dynamic. In Fig. 4, E(z) = 0 after adaptive filter W(z) converges. We then have W(z) = P(z) for X(z) ≠ 0. Which implies that the adaptive filter output y(n) is identical to the primary disturbance d(n). When d(n) and y(n) are acoustically combined, the residual error is e(n) = d(n) – y(n) = 0 which results in perfect
Page 321
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
cancellation of both sounds based on the principle of super position. More insight into the system reveals that we have to compensate secondary path transfer function S(z) from y(n) to e(n) which includes D/A converter, reconstruction filter, power amplifier, loudspeaker, acoustic path from loudspeaker to error microphone, error microphone, preamplifier, anti-aliasing filter and A/D converter. All these components add noise and delay in generating cancelling signal y(n). A scheme suggested by Morgan [6] x(n)
Unknown Plant P(z)
d(n)
e(n) Σ
+
-
y(n) Digital filter W(z)
Fig. 4
x(n)
System identification view point of ANC
Adaptive filter W(z) (z)
e(n)
d(n)
Unknown Plant P(z)
Σ
S(z) y(n)
LMS Algorithm Fig. 5 Block diagram of ANC system using the FXLMS algorithm. to solve this problem is to place an identical filter in the reference signal path as shown in Fig. 5 to weight update of LMS algorithm which realizes the Filtered X-LMS (FXLMS) algorithm [7]. FXLMS algorithm for active noise control has been derived by Burgess [8] as In ANC applications, S(z) is unknown and must be . Therefore, the estimated by an additional filter filtered reference signal is generated by passing the reference signal through this estimate of the secondary path as
Velammal College of Engineering and Technology, Madurai
is the estimated impulse response of the where . With in the limit of slow secondary path filter adaptation, the algorithm will converge with nearly 90° of and S(z). Of-line modeling and phase error between online modeling can be used to estimate [1] S(z). Another problem in feed-forward ANC is the radiation of cancelling signal y(n) towards the reference microphone, resulting in a corrupted reference signal x(n). The coupling of the acoustic wave from the canceling loudspeaker to the reference microphone is called acoustic feedback. The feedback component of the reference microphone signal can be cancelled electronically using a feedback neutralization filter , which model the feedback path. The models and can be estimated simultane-ously by using the offline modeling technique [1] The noise field in a reasonably large enclosure, for example, a train compartment is more complicated than a narrow duct. The sound field inside a passenger train cabin is created by standing waves rather than propagating waves and depends on the superposition of a number of acoustic modes. A mode is characterized by the number of wavelengths that fit along one dimension of the enclosure. Theoretically, each mode contains fundamental frequency plus infinite number of harmonics [9]. For perfect multichannel active noise control, infinite number of microphones are essential to sense each harmonic and also infinite number of loudspeakers are required for sending control signals, which is practically impossible to implement. However, many microphones and loudspeakers, well placed throughout the enclosure are needed to make the enclosure reasonably quiet. Situations will be worsened when the characteristics of primary signal changes with time as it is very common in ANC applications. The spectrum of the sound inside a train compartment changes when it passes through a tunnel or over a bridge. Thus a good ANC system able to tackle such situations is quite challenging. To encounter such conditions, a multichannel adaptive noise control system should be devised with suitable adaptive control algorithm, which can generate secondary signals for nullifying the noise. Primary source
P +
M
+ F
Σ
Σ
S M
J
x(n)
d(n)
K
W
y(n)
e(n)
J M
Page 322
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Block diagram of a multiple channel feed forward acoustic ANC system that includes secondary path as well as feedback path is illustrated in Fig. 6. The system has J reference sensors, K cancelling signals to drive the corresponding secondary sources and M error sensors. The wide arrows represent an array of signals (acoustic or electrical) that are symbolically expressed as vectors. primary path transfer function Matrix P represents Pmj(z) from the primary source to each error sensor output secondary path em(n). The matrix S represents transfer function Smk(z) from K secondary sources to M feedback error sensors. Also the matrix F represents path transfer function Fjk(z) from K secondary sources to J reference sensors. There are possible feed forward channels, each demanding a separate adaptive filter and adaptive filters are presented by the matrix W. these Full description of a multiple channel ANC using FXLMS algorithm is given in [1]. V. SIMULATION STUDY OF ACTIVE NOISE CONTROL SYSTEM FOR A PASSENGER TRAIN CABIN Active noise control of a closed chamber like train cabin is a multichannel system problem as explained earlier. However, for study purpose only one channel has been considered and simulated using matlab. The primary noise x(n) has been collected inside an air-conditioned train compartment. Thousand samples with sampling interval of 0.001 second are used for simulation. The primary path transfer function is modeled as low pass filter which is achieved using matlab command b=fir1(n,Wn), returns row vector b containing n+1 coefficients. This is a hamming window based linear phase filter of order n with normalized cut off frequency Wn. In simulation the order of the filter is fixed as 32. In an ANC system, secondary path introduces delay as well as noise in control signal y(n) which is due to the electronic subsystems, loud speaker and error microphone explained earlier. For simplicity, instead of estimating the secondary path transfer function S(z), one of its most important features of introducing noise in y(n) is taken into consideration and hence random noise has been added with control signal y(n). Matlab command n = 0.02*randn(1000,1) generates a scaled version of normally distributed random numbers of matrix with mean = 0 , variance = 1 and standard deviation = 1, which represents electronics noise of secondary path. The electronic noise added to control signal is approximately 1% of the train noise. The delay factor is not considered in simulation as it can be compensated while practically implementing the system by proper placement of secondary sources and error microphones. The feedback path transfer
Velammal College of Engineering and Technology, Madurai
function F(z) has been omitted in this simulation study as it is assumed that secondary source is made directing towards error microphone as well as the reference sensor is assumed as directional microphone and placed in the opposite direction of loudspeaker. These assumptions made us to eliminate feedback problem. TABLE I SIMULATION RESULTS Diff. algorithms
Resultant noise inside the train cabin for different values of µ µ=0.25
µ=0.26
µ=0.27
µ=0.28
µ=0.29
µ=0.30
-23.56 µ=1.8
-23.67 µ=1.9
-23.79 µ=2.0
-24.02 µ=2.1
-23.62 µ=2.2
-17.70 µ=2.3
-26.65 µ=0.08
-27.01 µ=0.09
-27.19 µ=0.10
-27.60 µ=0.11
-27.94 µ=0.12
-27.79 µ=0.13
-23.26 µ=0.006
-23.80 µ=0.007
-23.67 µ=0.008
-24.61 µ=0.009
-25.10 µ=0.01
-25.25 µ=0.02
-15.90 µ=0.004
-16.14 µ=0.005
-16.57 µ=0.006
-16.55 µ=0.007
-16.77 µ=0.008
-15.84 µ=0.009
-16.42
-16.30
-16.11
-15.57
-14.78
-14.73
LMS NLMS SDLMS SELMS SSLMS
The ANC system in this environment has been simulated on a digital computer using different control
Primary noise (x) 0.5 0.4 0.3 0.2 Signal value
Fig. 6 Block diagram of an adaptive multiple-channel feed forward ANC system with feedback paths.
0.1 0 -0.1 -0.2 -0.3 -0.4
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Fig 7 Original noise in side the cabin Fig. 8(a) Resultant noise when NLMS algorithm is used (-27.94 dB for µ=2.2) algorithms [1,10] namely LMS adaptive algorithm, Normalized LMS adaptive algorithm, Signed data LMS adaptive algorithm, Sign error LMS adaptive algorithm, Sign- sign LMS adaptive algorithm. In each case performance of the system has been evaluated on the basis of the minimum output power of the resultant noise (error signal) after the destructive interference between primary and secondary noise. In all cases the simulation has been done by keeping the length of the adaptive filter as 32 taps. Just for an example, the matlab command ha=adaptfilt.lms (32,mu) constructs a 32nd order FIR LMS adaptive filter. The parameter mu is the step size of adaptive filter. By adjusting the value of mu the optimum performance of the filter can be achieved. In the same manner ANC systems using other algorithms are simulated and their performance
Page 323
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 8(c) Resultant noise when LMS algorithm is used (-24.02) for µ=0.28) plot of Error signal (e) 0.6
0.5
0.4
Signal value
is tabulated in Table-1 for different values of µ. . The primary signal is shown in Fig. 7. The error signals obtained while using different algorithms mentioned above are depicted in Fig. 8(a-e). Among the five control algorithms experimented, NLMS algorithm performed well in this environment. It could reduce the train cabin noise to a minimum level of -27.94 dB
plot of Error signal (e)
0.3
0.2
0.1
0.2
0 0.15
Signal value
-0.1 0.1
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Fig. 8(d) Resultant noise when SELMS algorithm is used (-16.77 dB for µ=0.01)
0.05
plot of Error signal (e)
0
0.6 -0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
0.5
plot of Error signal (e)
0.4 Signal value
0.3
0.25
Signal value
0.2
0.3
0.2
0.1
0.15 0
0.1 -0.1
0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
0
-0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Fig. 8(b) Resultant noise when SDLMS algorithm is used (-25.25 dB for µ=0.13) plot of Error signal (e) 0.3
0.25
Signal value
0.2
0.15
Fig. 8(e) Resultant noise when SSLMS algorithm is used (-16.42 dB for µ=0.004) VI. CONCLUSION A single channel ANC system for passenger train cabin has been simulated using the real signal collected from an air-conditioned train cabin. Simulation study reveals that normalized LMS algorithm works well in this environment compare to other experimented algorithms as it could reduce more cabin noise. The study can be extended by incorporating the estimated transfer functions of secondary path as well as acoustic feedback path in the system together with problems associated in a multichannel ANC system.
0.1
0.05
0
-0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Velammal College of Engineering and Technology, Madurai
Page 324
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
REFERENCE [ 1] S.M. Kuo and D.R. Morgan, Active Noise Control Systems – Algorithms and DSP implementations. New Yourk: Wiley, 1996. [ 2] S.J. Elliott and P.A. Nelson, “Active Noise Control,” IEEE Signal Processing Magazine, October 1983. [3] Colin H. Hansen, Understanding active noise cancellation, Spon Press, London, 2001 [4] L. L. Beranek and I. L. Ver, Noise and Vibration Control Engineering – Principles and applications. New York, Wiley, 1992. [5] Olson H.F & May E.G “Electronic sound absorber,” Journal of the Acoustical Society of America, 25, pp 1130-1136, 1953. [6] D.R. Morgan, “An analysis of multiple correlation cancellation loops with a filter in the auxiliary path,” IEEE Trans. Acoust., speech, Signal Processing, Vol. ASSP-28, pp. 454-467, Aug. 1980. [7] B. Widrow and S.D. Stearns, Adaptive Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1985. [8] J.C. Burgess, “Active adaptive sound control in a duct: A Computer simulation,” Journal of Acoust. Soc. of America, vol. 70, pp. 715-726, Sept. 1981. [9] Stephen J. Elliott, “Down with Noise,” IEEE Spectrum, June 1999. [10] Monson H. Hayes, Statistical Digital Signal Processing & Modeling. John Wiley & Sons, 1996.
Velammal College of Engineering and Technology, Madurai
Page 325
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Texture Segmentation Method Based On Combinatorial Of Morphological And Statistical Operations Using Wavelets V.Vijayapriya1, Prof.K.R.Krishnamoorthy2 1 M.E, CSE, Anna university,Trichy /2 HOD, Dept. of ECE, SACS MAVMM Engineering College, Madurai. [email protected]
ABSTRACT Segmentation process for texture analysis is widely used for pattern recognition, industrial automation, biomedical image processing and remote sensing. Segmentation based on combinations of morphological and statistical operation is preferred in this project on wavelet transform images.The features like shape, size, contrast or connectivity makes attractive usage of mathematical morphology and can be considered for segmented oriented features.In order to get a better performance with wavelet transform for haar, db1, db6, coif6 and sym8 the system proposes derived equation on dilation, erosion, mean and median which results in segmentation.The process divides the wavelet combinatorial segmentation algorithm in to 3 groups based on type of operation and number of operations. The method using wavelet transform is applied on Brodatz Texture which results in good segmentation. The 500 Brodatz Texture image considered for performance Analysis which is compared with derived equations and filters in order to get a segmented results. The segmented results are used in comparative study with Peak Signal to Noise Ratio (PSNR) technique. Keywords: Discrete Wavelet transform, morphological and statistical operation, filter analysis.
I.
INTRODUCTION
The aim of the present work is to explore the improved version of wavelet based texture segmentation derived by (a) the method used for the construction of the multiresolution image with morphological operations, and (b) the combinatorial coefficient selection process designed to emphasize the information that is most important in the segmentation process.
1.1 Overview of texture analysis Texture analysis is an important issue in many areas like object recognition, image retrieval study, medical imaging, robotics, and remote sensing. Despite the development of a family of techniques over the last couple of decades, there are only a few reliable methods. Multiresolution techniques seem to be attractive for many applications. In this study, present an
Velammal College of Engineering and Technology, Madurai
approach based on the discrete wavelet transform and morphological concept. We integrate the framework of Brodatz texture images with the transform coefficients to obtain a flexible method for texture segmentation. Compared to intensity (spatial domain), the wavelet coefficients appear to be more reliable with respect to noise immunity and the ease of feature formation The wavelet transform is a multi-resolution technique, which can be implemented as a pyramid or tree structure and is similar to sub-band decomposition. There are various wavelet transforms like Haar, Daubechies, Coiflet, Symlet and etc. They differ with each other in the formation and reconstruction. The wavelet transform divides the original image into four subbands and they are denoted by LL, HL, LH and HH frequency subbands. The HH subimage represents diagonal details (high frequencies in both directions ), HL gives horizontal high frequencies (vertical edges), LH gives vertical high frequencies (horizontal edges), and the image LL corresponds to the lowest frequencies. At the subsequent scale of analysis, the image LL undergoes the decomposition using the same filters, having always the lowest frequency component located in the upper left corner of the image. Each stage of the analysis produces next 4 subimages whose size is reduced twice when compared to the previous scale. i.e. for level ‘n’ it gives a total of ‘4+(n-1)*3’ subbands. The size of the wavelet representation is the same as the size of the original. The Haar wavelet is the first known wavelet and was proposed in 1909 by Alfred Haar. Haar used these functions to give an example of a countable orthonormal system for the space of square integrable functions on the real line. The Daubechies wavelets (1992) are a family of orthogonal wavelets defining a discrete wavelet transform and characterized by a maximal number of vanishing moments for some given support. With each wavelet type of this class, there is a scaling function which generates an orthogonal multiresolution analysis.
Page 326
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
II. METHODOLOGY
Input image
The figure 1.1 consists of three blocks with arrow indicating the process between the blocks. Each block represents a separate module.
DWT(FILT ERSELECTI ON
MORPHOLOGICAL
SEGMENTATIONSYST EM
Figure 1 overview of the project 2.1 Wavelet Transform for Image Analysis (a) DWT The first module of the project is concerned with the implementation of the image decomposition using Lifting wavelet Transform, where the input images are divided in to four number of coefficients like Approximation, Horizontal, Vertical and Diagonal. (b) Filter selection Selecting the appropriate wavelet filter for achieving good segmented result where used orthogonal filters such as Haar, Daubechies, Symlet and Coiflet filters. 2.2 Morphological and Statistical Operation The second module of the project is concerned with the morphological operations using Erosion and Dilation and statistical operations using Mean, Median filters with some randomly selected coefficients as alpha, beta and gamma for getting high quality segmentation. 2.3Segmentation system The third module is concerned with comparative analysis of the Brodatz texture image segmentation using DWT for 500 set of standard images. The images used in this analysis were obtained from internet. The Eight types of segmented results are evaluated by using PSNR and Distortion Measure technique to identify the quality of segmented images among haar, daubechies, coiflet and symlet wavelet filters. III. RELATED WORK
Velammal College of Engineering and Technology, Madurai
The concept of morphology proposed many segmentation algorithms based on the orientation and type of Structuring Element (SE). But this paper advocates a new method of segmentation of images, based on combinatorial approach of mathematical morphology and primitive statistical operations. A gray level image typically consists of both bright and dark object features with respect to size or scale. The basic objective of the present segmentation algorithm is to isolate or sketch out the most optimal contours of these bright and dark features. [1] One of the most popular filter pairs is the 9/7 biorthogonal pair of Cohen, Daubechies, and Feauveau, which is adopted in the FBI finger-print compression standard. Present a technique to rationalize the coefficients of wavelet filters that will preserve biorthogonality and perfect reconstruction. Furthermore, most of the zeros at z = -1 will also be preserved. These zeros are important for achieving regularity. The rationalized coefficients filters have characteristics that are close to the original irrational coefficients filters. Three popular pairs of filters, which include the 9/7 pair, will be considered [2]. IV. PROPOSED WORK The present project advocates a new method of segmentation of images, based on combinatorial approach of mathematical morphology and primitive statistical operations. The segmentation results depend upon the combinatorial coefficients of α, β and γ. This dependency is restricted to 95% by keeping the condition α > β and α > γ. S[i,j]=α*D[i,j]-β*E[i,j]- γ *(D[i,j]-E[i,j])
….(3.1)
S[i,j]=α*D[i,j]-β*E[i,j]- γ *(D[i,j]+E[i,j]) ….(3.2) D [i, j], E [i, j] represents Dilated and Eroded values of the 3×3 mask respectively, S [i, j] represents segmented grey level value of the mask. The group two combinatorial segmentation is represented by equations (3.3) and (3.4). S[i, j] =|α *D[i, j] − β * E[i, j] −γ * Avg | …(3.3) S[i, j] =|α *D[i, j] − β * E[i, j] −γ * Med | …(3.4) Where α > β and α > γ , ‘Avg’ represents average value of the 3×3 mask and ‘Med’ represents median value of the 3×3 mask. The segmentation of group three forms from the two operations of morphology and/or statistics. The mathematical representation for group three combinatorial segmentation is given in equations (3.5), (3.6), (3.7) and (3.8). S[i, j] = [α *D[i, j]− β *Avg]
…(3.5)
Page 327
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
S[i, j] = [α *Avg − β *E[i, j]]
…(3.6)
25 Eq3
20
S[i, j] =[α *D[i, j]− β *Med]
…(3.7)
S[i, j] = [α *Med − β *E[i, j]]
…(3.8)
10
Where α ≥ β. This is needed to keep the resultant value of equations (3.5), (3.6), (3.7) and (3.8) as positive.
V. DISCUSSION
EXPERIMENTAL
RESULTS
AND
The comparative analysis of various wavelet filters using wavelet transforms has been performed for Approximated value and detailed value with the combinatorial coefficients 431 is analyzed.The PSNRvalue for one level decomposition has been performed.
5
Eq1
Eq7
Eq5
Eq6
Eq8
Approxim ation
Eq 4
Eq3
25
Eq4
20 15 10 5
Eq 1
Eq7
Eq6
Fig‐3 Approximation and detailed coefficients segmentation of coif6 filter using alpha=4, peta=3, gama=1
Eq 3
Eq1 Eq1
Detailed
Eq8
Eq 6
EQ5
Eq7 Eq3 Eq5 Eq6 Eq8 Eq2 Eq4
Eq6
Eq7 Eq8
0
EQ5 Eq 7
Eq 1 Eq 3 Eq 5 Eq7 Eq 2 Eq 4 Eq 6 Eq 8
EQ5
Eq8
Detailed
15 10
Eq4
-5
20
Eq2
30
25
Eq3
Eq1
30
5 0
Eq4
15
Approximation
0
Detailed
Approxim ation
Fig‐1 Approximation and detailed coefficients segmentation of sym8 filter using alpha=4, peta=3, gama=1
Fig‐4 Approximation and detailed coefficients segmentation of Db6 filter using alpha=4, peta=3, gama=1.
Inf
5 4
Eq3
25
3
Eq4
1.458
2
20
1.722 -1.883
1
15
10
0 EQ5 Eq1
Eq3 Eq2
Eq4
Eq6
Eq8
-1
Eq7
Eq1
Eq7
Eq5
Eq6
Eq8
-2
5
Db1 Detailed
Db6
Sym 8
Coif6
0 Approximation
Fig‐5 Segmentation2 result PSNR value comparision of Haar with Db1, Db6, sym8 and coif6 using alpha=4, peta=3,gama=1
Fig‐2 Approximation and detailed coefficients segmentation of Db1 filter using alpha=4, peta=3, gama=1
Velammal College of Engineering and Technology, Madurai
Page 328
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
conference. Acoust, Speech, Signal Processing, Volume. 6, page no. VI–513. April. 2003
VI. CONCLUSION AND FUTURE WORK The comparative analysis of various wavelet filters using wavelet transforms has been performed for Approximated value and detailed value with the combinatorial coefficient 431 is analyzed. A graph was drawn for PSNR vs filter, from the graph obtained that the db1 filter provides better performance over the remaining filter. The comparative analysis of various wavelet filters for Detailed value and approximated value with various combinatorial coefficients has to be performed with PSNR and Distortion measure techniques.
[10]
Michael Weeks,” Digital Signal Processing Using MATLAB and Wavelets”, 2007 Edition. Infinity Science Press LLC. (An imprint of Laxmi Publications Pvt. Ltd.)
[11]
JuHan, Kai-Kuang Ma, “Rotation invariant and scale-invariant Gabor features for texture image retrival”, science direct image and vision computing 25 page no. 1474-1481, 2007.
[12]
Dogu Baran Aydogan, Markus Hannula, Tuukka Arola, Jari Hyttinen. “Texture based Classification and Segmentation of Tissues using DT-CWT feature extraction methods”, 21st IEEE International Symposium on Computer-Based Medical Systems. page no. 614-619, 2008.
REFERENCES
[1]
Vijayakumar.V, Raju.U.S.N and Narasimha Rao.A. “Wavelet based texture segmentation based on combinatorial of Morphological and Statistical operations”. IJCSNS International Journal of Computer Science and Network Security , volume Page no 176-180 August 2008
[2]
Jianyu Lin and Mark J. T. Smith “New Perspectives and Improvements on the Symmetric Extension Filter Bank for Subband/Wavelet Image Compression” -IEEE Transactions on signal processing, volume. 17, NO. 2, page no 177 - 189 February 2008.
[3]
Balasingham.I and Ramstad, “Are the Wavelet T.A Transforms the Best Filter Banks for Image Compression?” research article, Hindawi Publishing Corporation EURASIP Journal on Image and Video Processing, page no 1-7, Article ID 287197, 2008.
[4]
Sweldens.W, The lifting scheme: “a coustom-design construction of biorthogonal wavelets” appl. Comut.harmon.anal.volume no. 3,page no.186-200, 1996
[5]
sweldens.w, “the lifting scheme: a construction of second generation wavelets,” SIAM J. Math.Anal,page no .511-546,1997.
[6]
Torsten Palfner and Erika Muller, “effects of symmetric periodic extension for multiwavelet filter banks on image coding”. IEEE Institute of telecommunications and information technology university of Rostock, Germany,1999.
[7]
Bryan E. Usevitch, “A tutorial on modern lossy wavelet image compression: foundations of JPEG 2000” IEEE signal processing magazine, September 2001.
[8]
[9]
Soman K.P and Ramachandran K.I,”Insight into wavelets from theory to practice, 2006 edition. Prentice hall of India private limited. Adiga A, Ramakrishnan K. R, and Adiga B. S “A Design and implementation of orthonoramal symmetric wavelet transform using PRCC Filter Banks”-in proceeding IEEE International
Velammal College of Engineering and Technology, Madurai
Page 329
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fpga Design Of Routing Algorithms For Network On Chip R.Anitha#1, Dr.P.Renuga#2 Lecturer, Department of Electronics and Communication Engineering, PSNA CET, Anna University, India #2 Assistant Professor, Department of Electrical and Electronics Engineering, TCE, Anna University, India
#1
1
[email protected] 2 [email protected]
Abstract: In this paper, we present a methodology to develop efficient and deadlock-free routing algorithms for Networkon-Chip (NoC) platforms that are specialized for an application or a set of concurrent applications. The proposed methodology, called the Application- Specific Routing Algorithm (APSRA), exploits the application-specific information regarding pairs of cores that communicate and other pairs that never communicate in the NoC platform to maximize communication adaptivity and performance. Based on the selected network topology, outgoing packets are switched either directly to the East/West router outputs or to the second crossbar for the North/South router outputs. Incoming packets that arrive through the North/South ports that are connected to the second crossbar are switched directly to the attached node element, but packets arriving through the East/West ports must be switched from the first crossbar to the second crossbar in order to reach the attached node element. In this paper we store the Routing table in compressed form in the router. Hence the number of gates required is reduced. In future also we perform the Routers data compression by using Modified Run length encoding which makes the fast data transfer.
1. INTRODUCTION Network on chip has emerged as a dominant paradigm for the synthesis of multicore SoCs they are generally viewed as the ultimate solution for the design of modular and scalable communication architectures and provide inherent support for the integration of heterogeneous core through the standardization of the network boundary. The network topology and routing algorithm used in the underlying onchip communication network are the two important aspects. A highly adaptive routing algorithm has a potential of providing high performance, fault tolerance and uniform utilization of network resources. Generally routing algorithms providing high adaptiveness guarantee freedom from deadlock by means of virtual channels (VCs).For arbitary network topologies, a relatively large number of VCs could be required to provide deadlock freedom, high adaptivity, and shortest paths. However the use of the VCs introduces some overhead in terms of both additional resources and mechanisms for their management. To limit such overhead, some proposals try to minimize the number of VCs required to obtain a high degree of adaptivity. For these reasons, we believe that reducing as much as possible
Velammal College of Engineering and Technology, Madurai
the number of VCs required to provide deadlock-free routing and, at the same, guaranteeing a high degree of adaptiveness must be considered one of the primary concerns in the NoC domain. In this paper, we present a methodology to design highly adaptive routing algorithms for NoC architectures that do not require the use of any VCs. In order to develop efficient, highly adaptive, deadlock-free and topologyagnostic routing algorithms. We call algorithms developed using this information as Application-Specific Routing Algorithms. The APSRA methodology is general, in the sense that it is not designed for a given/fixed network topology. 2. APSRA DESIGN METHODOLOGY Communicatio T
Applic ation
T Task T
T
Ta sk
T 1 T 2
T
Mappi ng Topology Graph T l
I P P 7
M
APSR
Compres sion
C C
P6
Cb
dS C
Rout i
Compre ssed
Configure/Rec
Fig. 1. APSRA Design Methodology
The APSRA design methodology is depicted in fig.1. It gets as inputs the application modeled as a task graph and the network topology modeled as a topology graph. Using Task Mapping information, APSRA generates a set of routing tables that not only guarantee both the reachability and the deadlock freeness of communication among tasks
Page 330
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
but also try to maximize routing adaptivity. A compression technique can be used to compress the generated routing tables. Finally, the compressed routing tables are uploaded to the physical chip. 2.1. ARCHITECTURAL IMPLICATIONS OF APSRA (TL,BR,Colo InReg ion Top
cfg setup dst
InReg ion Top
2.2. ROUTING TABLE COMPRESSION
hit ao
Admissible Outputs
M
Dem
dst
addr
ao
cfg setup dst
cfg
additional degree of freedom in routing algorithms as they can be reconfigured at runtime to adapt to the current traffic condition or offline to manage network topology variation due to manufacturing defects. A disadvantage is that a table can take a large space if many destination addresses should be stored. That is, in such an implementation, the cost (silicon area) of the router will be proportional to the size of the routing table.
hit
.
We proposed a method to compress the routing table to reduce its size such that the resulting routing algorithm remains deadlock free and has high adaptivity. The basic idea of the compressing method is shown in Fig.3 Let us suppose that the routing table associated to the west input port of node X is that shown in Fig. 3a.
setup ao
cfg setup
InReg ion Top
dst
hit
Enco der M:lg
Data En Cfg=(TL,BR,Col setup dat
Re TL
Data En
Re
BR color
InReg ion
ao hit
dat En
Re
Fig. 2. Block diagram of the Architectural Implications of APSRA
One way to implement the routing function is to design it in hardware logic (also known as the algorithmic approach). In this case, there can, e.g., be a state machine that takes the current node address, the destination address, and some status information stored in the router as inputs and outputs the admissible out ports. Another way to implement the routing function is to use a routing table. The index to the table, where the admissible outputs are stored, is the destination address or a function of the destination address. The values of the table are dependent on which router it is I Implemented. The algorithmic approach comes at the cost of a loss of generality with respect to the table-based approach. In fact, the algorithm is only specific to one topology and to one routing strategy on that topology. On the other hand, this mechanism turns out to be more efficient (area wise and performance wise) than table-based routing. However, routing tables introduce an
Dest.
Admissible Outputs
A
North, East
B
East
C
South, East
D
South, East
E
South, East
F
South, East
G
South
H
South
I
East
Fig. 3. (a). Routing Table before Compression
A B
A B
X
R 1
X
C D F G H I
C D E
F E G H I
R 2 Fig. 3. (b). Color based Clustering
Velammal College of Engineering and Technology, Madurai
Page 331
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
asserted, the address input is used to select the region to be configured. Configuring a
Dest.
Admissible Outputs
A
North, East
B
East
R1
South, East
R2
South
I
East
Fig. 3. (c). Compressed Routing Table
Since we are considering minimal routing, the destination node for a packet received from the west port will be in the right part of the NoC. Now, let us suppose to associate a color for each one of the five possible sets of admissible output directions. For instance, we mark with color Red (Green/Blue) all the destination nodes that, to be safely reached, needs that a packet in the west input port must be forwarded through the North (South/East) output port. Similarly, we mark with color Purple (Yellow) all the destination nodes that, to be safely reached, needs that a packet in the west input port must be forwarded through the North or East (South or East) output channels. The left side of Fig. 2b shows the set of destinations colored on the basis of the information stored in the routing table. The next step is to perform a color-based clustering of the destination nodes by means of rectangular regions. The right side of Fig. 2b shows an example of such a clustering. The basic idea of the compression technique is that it is no longer necessary to store the set of all the destinations, but only the set of Clusters, as shown in Fig. 2c. If a small loss in adaptivity is allowed, a cluster merging Procedure can be used to increase the compression ratio. For instance, in Fig. 13c, clusters R1 (yellow) and R2 (green) can be merged to form a single cluster R3 (yellow) whose destinations can be reached by using only the east output port. As reported i[55], in practical cases, using from four to eight clusters is enough to cover all practical traffic scenarios without any loss in adaptivity. Using from two to three clusters gives a small reduction in adaptivity, which translates to less than 3 percent performance degradation as compared to uncompressed routing tables. From the architectural viewpoint, the logic to manage the compressed routing table is very simple. Fig. 3 shows the block diagram of the architecture implementing the routing function that uses the compressed routing table. For a given destination, the block extracts from the compressed routing table the set of admissible outputs. The input port setup allows us to configure the router. When the setup is
Velammal College of Engineering and Technology, Madurai
region means storing attributes characterizing the region (top right, bottom left, and color) into the compressed routing table. The block In Region checks if a destination dst belongs to a region identified by its top-left corner (TL) and its bottom-right corner (BR). If this condition is satisfied, the output assumes the value of the color input, and output hit is set. 2.3. Table-Based Implementation A. Area Overhead Analysis To evaluate the overhead in silicon area due to the use of routing table with respect to routing logic, we designed in VHDL and synthesized by using Synopsys Design Compiler for a UMC 0:13-_m technology library the following blocks: Routing function: It is the block that gives the set of admissible outputs for the current node and a given destination. We designed three different blocks that implement the XY routing function, the Odd-Even routing function, and the routing-table-based routing function. The first two are pure combinatorial logic. The latter one, whose block diagram is depicted in Fig. 14, contains memory elements also. With regard to the routing-table-based routing block, we considered two different instances able to manage four regions (RT4) and eight regions (RT8), respectively. Input FIFO_5. They are the FIFO buffers at the input of each router. For a mesh-based network topology, there are five FIFO buffers in total. Crossbar: It is a general 5 _ 5 crossbar block that allows us to simultaneously route non conflicting packets. Arbiter: It is a general arbiter that manages the situation where several packets simultaneously want to use the same output. In this case, arbitration between these packets has to be performed. We used a round-robin policy. Table 3 reports the silicon area for each of the main blocks of a router for three different routers Implementing XY routing, Odd-Even routing, and table-based routing, respectively. With regard to the latter, two different implementations able to manage compressed routing tables with a budget of four (RT4) and eight (RT8) table entries have been considered. The cost overhead of a routing table implementation based on the proposed compression technique and architecture represents only a small fraction of the overall router cost.
Page 332
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3. APSRA SIMULATED OUTPUT Simulation Results For South‐west Port
Simulation Results For East Port
4. CONCLUSION In this paper, we have proposed a topology –agnostic methodology to develop efficient APSRA for NoCs. Our methodology not only uses the information about the topology of communicating cores but also exploits information about the concurrency of communication transactions Routing algorithms generated by the APSRA method has even higher performance and adaptivity advantage over other deadlock-free routing algorithm. The APSRA methodology uses a heuristic to remove the minimum number of edges in the channel dependency graph to ensure deadlock-free routing. A straightforward table implementation may not be practically useful, since the table size grows with the number of nodes in the network. Therefore, we have also proposed a technique to compress tables such that the effect on the original routing algorithm is minimal.
[1] A. Ivanov and G.D. Micheli, “The Network-on-Chip Paradigm in Practice and Research,” IEEE Design and Test of Computers, vol. 22, no. 5, pp. 399-403, Sept.-Oct. 2005. [2] S. Kumar, A. Jantsch, J.-P. Soininen, M. Forsell, M. Millberg, J. Oberg, K. Tiensyrja, and A. Hemani, “A Network on Chip Architecture and Design Methodology,” Proc. IEEE CS Ann. Symp. VLSI, p. 117, 2002. [3] W.J. Dally and B. Towles, “Route Packets, Not Wires: On-Chip Interconnection Networks,” Proc. 38th Design Automation Conf. (DAC ’01), pp. 684-689, 2001. [4] F. Karim, A. Nguyen, and S. Dey, “An Interconnect Architecture for Networking Systems on Chips,” IEEE Micro, vol. 22, no. 5, pp. 36-45, Sept.-Oct. 2002. [5] P.P. Pande, C. Grecu, A. Ivanov, and R. Saleh, “Design of a Switch for Network on Chip Applications,” Proc. IEEE Int’l Symp. Circuits and Systems (ISCAS ’03), vol. 5, pp. 217-220, May 2003. [6] T. Bjerregaard and S. Mahadevan, “A Survey of Research and Practices of Network-on-Chip,” ACM Computing Surveys, vol. 38, no. 1, pp. 1-51, 2006. [7] P.P. Pande, C. Grecu, M. Jones, A. Ivanov, and R. Saleh, “Performance Evaluation and Design Trade-Offs for Networkon- Chip Interconnect Architectures,” IEEE Trans. Computers, vol. 54, no. 8, pp. 1025-1040, Aug. 2005. [8] D. Linder and J. Harden, “An Adaptive and FaultTolerant Wormhole Routing Strategy for k-Ary n-Cubes,” IEEE Trans. Computers, vol. 40, no. 1, pp. 2-12, Jan. 1991. [9] C.J. Glass and L.M. Ni, “The Turn Model for Adaptive Routing,” J. Assoc. for Computing Machinery, vol. 41, no. 5, pp. 874-902, Sept. 1994. [10] A.A. Chien and J.H. Kim, “Planar-Adaptive Routing: Low-Cost Adaptive Networks for Multiprocessors,” J. ACM, vol. 42, no. 1, pp. 91-123, Jan. 1995. [11] J. Upadhyay, V. Varavithya, and P. Mohapatra, “A Traffic- Balanced Adaptive Wormhole Routing Scheme for Two-Dimensional Meshes,” IEEE Trans. Computers, vol. 46, no. 2,pp. 190-197, Feb. 1997. [12] G.-M. Chiu, “The Odd-Even Turn Model for Adaptive Routing,” IEEE Trans. Parallel and Distributed Systems, vol. 11, no. 7, pp. 729-738, July 2000.
5. REFERENCES
Velammal College of Engineering and Technology, Madurai
Page 333
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Creating Actionable Knowledge within the Organization using Rough set computing Mr.R..Rameshkumar #1 , Dr. A. Arunagiri *2 , Dr.V .Khanaa #3 , Mr.C.Poornachandran
=4
# Information System Department,Bharat University, Selaiyur,Chennai – 600073, T.N,India. 1
[email protected] [email protected]
3
* EEET Department,Yanbu Industrial College, Yanbu Al- Sinaya , Yanbu, KSA. 2
[email protected]
= Department of Computer Science,Government Arts College, Nandanam,Chennai, India. 4
[email protected]
• Scalable from 2 to many experts.
Abstract For a management consultant in order to successfully assist an organization in creating new actionable knowledge (knowledge that is used to create value), the consultant must be aware of a knowledge dimension called Coalescent knowledge. The process for creating actionable knowledge in this dimension is a dialogue process. For example, product development is guided by several expert knowledge including critical process relationships which are dynamic, derived from experience and are often nonlinear. A new inductive learning intelligent technique called Rough set theory, which is to be used along with coalescent knowledge is described and proposed. The frame work of the classic management has been discussed and is modified using the Rough set theory method.
Keywords : Actionable Knowledge, knowledge, Rough Set , Critical Process
Coalescent
Morgan etal updated the knowledge creation theory documented by Nonaka and Takkeuchi, [2] and is expressed in Table 1: Changes to the Nonaka & Takkeuchi (NK) knowledge Creation theory Process/Mode
Velammal College of Engineering and Technology, Madurai
New Knowledge
Transition
Transition Form
Form Socialization
Tacit-to-Tacit
Tacit to Coalescent
Externalization
Tacit to Explicit
Coalescent to Explicit
Combination 1.0 Introduction For a management consultant to successfully assist an organization in creating new actionable knowledge (knowledge that is used to create value), the consultant must be aware of a new knowledge dimension called Coalescent knowledge (Morgan, Morabito, Merino,Reilly, 2001). Knowledge in the Coalescent dimension has the following attributes: • Created via a dialogue process involved group of experts • The knowledge is visible, expressible, shared and virtual • Can be private or public knowledge • Can be used to create a sustainable competitive advantage • Facilitates the opportunity for group of experts to act as if they have one mind (decide by rough set ) to accomplish organizational objectives
NK Knowledge
Internalization
Explicit to
Explicit to
Explicit
Coalescent
Explicit to Tacit
Coalescent to Tacit
Table 1 2.0 Rough Set Theory The concept of rough sets has been introduced by Zdzislaw Pawlak in the early 1980’s as a new concept of set used in many application and research. The Rough Set theory is an important part of soft computing. The methodology is concerned with the classificatory analysis of imprecise, uncertain or incomplete information or knowledge expressed in terms of data acquired from experience. It generates output as a list of certain and possible rules. It is
Page 334
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
formalized through the simple concepts of lower and upper approximation, which are in turn defined on the basis of set. The Rough set theory is enhancements of Neural networks, genetic algorithms, clustering ,support vector machines, regression models, etc.
Combination
Socialization
Internalization Expert No.1
Rough Set
Explicit
Coalescent . Tacit
3.0 Updated Knowledge Creation Process According to Morgan etal The first process in creating knowledge is the socialization. In this process, an individual shares his or her tacit knowledge with another individual or a group via some form of dialogue and/or observation [1], [2], In any dialogue and/or observation, each individual brings his or her tacit knowledge and references/links to explicit knowledge. For this analysis, it is assumed that the exchange of knowledge will be via dialogue. During the dialogue process, the first individual tries to define his/her tacit knowledge for the second person(s). This process requires the use of fields of interaction. The second person(s) then links their knowledge base to the knowledge being communicated. This is a repetitive action until the first and second person(s) agree on a common set of constructs, which defines the knowledge being communicated. The process has now created a shared virtual knowledge, which only exists between the individuals involved in the dialogue. This knowledge is shared and not codified, so it does not fall within the definition of tacit or explicit knowledge. The consultant can use the knowledge creation process flow that includes the interaction of people in the organization at a high level [1] to analyze the organization potential for creating knowledge in the coalescent dimension
.. 4.0 Refined Knowledge Creation Process Figure 2 shows the Refined knowledge creation process that includes the interaction of people or (experts) in the organization.
Velammal College of Engineering and Technology, Madurai
. . .
Expert No. N
Socialization
Externalization
Internalization
Refined Knowledge Creation Process Flow Figure 2
In the refined knowledge creation process[2], the creation of new tacit knowledge does not require the inclusion of explicit knowledge. As Coalescent knowledge matures it can be externalized via codification to become explicit knowledge which may be private or public. If there is only one Expert in the tacit dimension, then that expert would have both Tacit and Coalescent knowledge. If it requires to internalize some explicit knowledge, then the explicit knowledge would be converted to Coalescent knowledge. The Coalescent knowledge is shared between the expert and the creators of the explicit knowledge. Although the creators are not actively participating in the dialogue, the expert doing the internalization assigns them a virtual role. Knowledge that is shared by two or more expert’s can be considered communal knowledge. The formation of Coalescent knowledge from tacit dimension will be certain when the dimension includes minimum number of experts. Whereas the uncertainty arises in the knowledge creation process within the organization when the tacit dimension involves many number of experts. Since the Rough set deals with the uncertainties, it can be effectively used to refine the Coalescent dimension and to make the actionable knowledge more certain. The Rough set can make use of different experts tacit knowledge derived from similar organization in order to refine the uncertainty problems.
5.0 Classic Management Process The classic management process is comprised of four subprocesses: Planning, Organizing, Leading, and Controlling (PLOC)[5]. The controlling and planning sub-processes are connected through a feedback loop to ensure that the objectives of the plan are being met. The controlling function compares performance to the standards set in the planning process. If deviations are detected, the information is feed back to the planning process for changes in the plan that will cause the standards to be met. During the controlling and/or planning processes, an
Page 335
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
evaluation should be performed to determine why there was a deviation from standards.
C l a s s i c M a n a g e m e n t P r o c e s s
Planning
Leading
Organizing
Controlling
Feedback Loop (P L O C) Figure 3
6.0 Modified Classic Management process by Rough Set theory By adopting the critical reflective practice principals as defined by Van Aswegen (1998)[6] to an open group dialogue format, management consultant could make use of the Coalescent knowledge derived from the tacit dimension. Based on the formation of the uncertainty problems, the Coalescent knowledge can be refined by implementing the rough set concept in order to eliminate the uncertainty and to minimize the deviation from the standards.
Knowledge Creation Theory, Paper presented at the First International Conference on Knowledge Management, co-sponsored by Academy of Management, ISEOR, University [2] Nonaka, Ikujiro and Takeuchi, 1995, The Knowledge Creating Company, Oxford University Press, New York, NY. [3] Pawlak,Zdzislaw (1991). Rough sets: Theoretical Aspects of Reasoning About Data. Dordrecht: Kluwer Academic Publishing. ISBN 0-7923-1472-7. [4] Morabito, Joseph, Sack, Ira, and Bhate, Anilkumar, 1999, Organization Modeling – Innovate Architectures for the 21st Century. Prentice-Hall, Inc., Upper Saddle River, NJ. [5] Freeman, R.E., and Stoner, J.A.F., 1989, Management. Prentice Hall, Englewood Cliffs, NJ, 4th Edition. [6] Van Aswegen, Elsie Johanna, 1998, Critical Reflective Practice: Conceptual Exploration and Model construction (Emancipatory Learning, Naturalistic Inquiry), Dissertation, University of South Africa, South Africa [7] Komorowski, J. et al., 1999. Rough sets: A tutorial. In S.K. Pal and A. Skowron (eds.), Rough Fuzzy Hybridization: A New Trend in Decision-Making. Springer, pp. 3-98.
Modified Classic Management Process
Critical Reflective Practice Figure 4
7.0 Conclusion : This paper describes the concept of tacit knowledge dimension and the formation of Coalescent knowledge from tacit dimension. As the tacit dimension increases, the uncertainty problem is eliminated through Rough set method. The concept of Classical Management Process is explained and is modified by the refined Coalescent knowledge through Rough set in order to minimize the deviation between the reference and the actual outcome. References: [1] Morgan, Morabito, Merino, and Reilly, 2001, Defining Coalescent Knowledge: A Revision of
Velammal College of Engineering and Technology, Madurai
Page 336
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
COMPVAL – A system to mitigate SQLIA S. Fouzul Hidhaya1 , Dr. Angelina Geetha2 Department of Computer Science and Engineering, B.S. Abdur Rahman University, Vandalur, Chennai -600048, Tamil Nadu, India. 1
[email protected] 2 [email protected]
Abstract - In this Internet Era, Web Applications are a boon to the business to customer data exchange. The web applications are directly linked to the database and so the web applications are used as a interface to carry out the SQL Injection Attacks on the databases. This paper discusses a novel method to mitigate the SQL Injection Attacks. This system has two processes running in parallel. The actual process of retrieving the data from the database is done in parallel with the process of syntactically analyzing the query statement. The intended output and the actual output of the data are compared and validated and the corresponding HTTP response is framed. This system provides a double check for the SQL injection attack and from them the experimental results it has been proved effective and has imposed a average overhead of 15 milliseconds when tested on the web applications. Keywords - SQL Injection, Vulnerabilities, Web application security, Security threats, Web application vulnerability.
I. INTRODUCTION Business to customer (B2C) data exchanges like the online shopping, banking. gaming, ticket booking, travel booking etc. has been made feasible by the web applications. Web application work with a back-end database to store or retrieve data. These data are accepted from the user or retrieved from the database to be given to the user dynamically by embedding them in a HTML code. The input supplied by the user could be benign or malicious. If the data is malicious, undesired data could be extracted from the database. This type of attack is called the SQL injection attack (SQLIA). An example of SQL Injection is given here. A SQL injection attack occurs when an attacker causes the web application to generate SQL queries that are functionally different from what the user interface programmer intended. For example the application contains the following code, Query = “SELECT * FROM members WHERE login ='”+ request.getParameter(“login”)+” ‘ And password=’ “ + request.getParameter(“password”)+” ‘ “; In this code the web application retrieves user inputs from login and password, and concatenates these two user inputs into the query. This is used to authenticate the user. If the attacker enters admin into the login filed
Velammal College of Engineering and Technology, Madurai
and xyz’ OR ‘1=1 into the password field. Now the query string will be, SELECT * from members WHERE name=’admin’ AND password=’xyz’ OR ‘1=1’ In this case, the password field, which should have only one string, is replaced with sub strings, which will finally lead to the retrieval of the data of the administrator which is undesirable. An effective and easy method, for detecting and protecting existing web applications from SQL Injection Attack, is need of the moment, which will help the organizations to be secured against these attacks. Detection or prevention of SQLIA is a topic of active research in industry and academia. The most commonly injection mechanisms [13] are 1. Injection through user input. 2. Injection through cookies 3. Injection through server variables 4. Second order injection The proposed attack works on injection through user input. The rest of the paper is organized as follows. Section 2 discusses the related work carried out in this area, Section 3 illustrates the system design. Section 4 evaluates the effectiveness of the proposed approach and Section 5 concludes the paper. II. RELATED WORK Researchers have proposed a wide range of techniques to address the problem of SQL Injection. These techniques range from development best practices to fully automated frameworks for detecting and preventing SQLIAs[12]. One technique being used to mitigate SQLIA is the defensive coding practice. Since the root cause for the SQL injection vulnerabilities is insufficient input validation, the straight forward solution for eliminating these vulnerabilities is to apply suitable defensive coding practices, like Input type checking, encoding of inputs, pattern matching and identification of all input sources. But defensive coding is prone to human error and is not as rigorously and completely applied as automated techniques. While most developers do make an effort to code safety, it is difficult to apply defensive coding practices correctly to all sources of input.
Page 337
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
functionalities of the application, and stores them in the repository named the ‘Prototype document’. In the normal phase(Section 3.2), the HTTP request is interpreted by the ‘HTTP Interceptor’ and then sent to the web application which retrieves the data from the database and with this data the web application frames the HTTP response. In the validator phase (Section 3.3), the HTTP Interceptor sends the Intercepted HTTP request to the ‘Query Interceptor’. Here the query is extracted and send to the ‘parse tree generator’ and the ‘syntax analyzer’. The parse tree generator generates the parse tree for the query and sends to the syntax analyzer. Syntax analyzer compares the query from the query interceptor and the query database and extracts the user input and applies this user input to the comparator phase (Section 3.4). If the request is marked as benign the HTTP response is immediately send to the user. If the request is marked as malicious by the syntax analyzer then the comparator compares the intended output for the query and the actual output that is retrieved from the database. If they are the same then requested data is send in the HTTP response. Else a ‘data unavailable response’ is send in the HTTP response.
Researchers have proposed a range of techniques to assist developers and compensate for the shortcomings in the application of defensive coding. The black box testing methodology used in WAVES[6], which uses a web crawler to identify all points in web application that can be used to inject SQLIAs. It uses machine learning approaches to guide its testing. Static code checkers like the JDBCchecker[4] is a techniques for statically checking the type correctness of dynamically generated SQL queries. This will be able to detect only one type of SQL vulnerability caused by improper type checking of input. Combined static and dynamic analysis like the AMNESIA [5] is a model based technique that combines static analysis and runtime monitoring. AMNESIA interprets all queries before they are sent to the database and checks each query against the statically built models. SQLGuard [3] and SQLCheck [10] also check queries at runtime to see if they conform to a model of expected queries. Taint based approaches like the WebSSARI[8]detects input-validation-related errors using information flow analysis. In this approach, static analysis is used to check taint flows against preconditions for sensitive functions. Livshits and Lam [7] use information flow technique to detect when tainted input has been used to construct a SQL query. Security gateway [9] is a proxy filtering system that enforces input validation rules on the data flowing to a web application. SQLRand [2] is an approach based on instruction-set randomization. It allows developers to create queries using randomized instruction instead of normal SQL keywords. A proxy-filter intercepts queries to the database and derandomizes the keywords. III. I.
Web application
Prototype document
Figure 2: Query Extractor
SYSTEM DESIGN
System Architecture
Figure 1 : COMPVAL System Architecture
The main system architecture of COMPVAL is illustrated in Figure 1. This system could be divided into four phases, Query extractor, normal phase, validator phase and the comparator phase. The query extractor phase extracts the query from the web application code to cover all the
Velammal College of Engineering and Technology, Madurai
Query Extractor
II.
Normal Phase The normal phase is similar to the general operation of the web application without any SQLIA detecting mechanisms. The HTTP interceptor intercepts the HTTP request sent by the user and sends the request to the web application. The web application then sends the request to the database and retrieves the data. This data is then framed as a HTTP response and instead of sending it to the user the response here is send to the comparator for evaluation. C. Validator phase In this phase, the intercepted HTTP request is send to query interceptor and the query is extracted from the HTML and send to the parse tree generator and the syntax analyzer. The syntax analyzer uses XSL’s pattern matching. The syntax analyzer works as follows: 1. Accepts the query from the query interceptor and tokenizes the query.
Page 338
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2. Transforms the query into XML format. 3. Uses the XSL’s pattern matching algorithm to find the prototype model corresponding to received query. 4. From the prototype query it identifies the user input data. 5. Receives the parse tree from the parse tree generator and fits in the nonleaf nodes. 6. Validates the resulting parse tree using the validation algorithm and identifies if the query is benign or malicious. 7. Sends the prototype query and the result to the comparator. 1) XSL’s Pattern Matching : When the query is submitted by the query interceptor, the query is first analysed and tokenized as elements. The prototype document contains the query pertained to that particular application and it also has the intended output for each query model. For example, as explained in section 1, the input query is, SELECT * FROM members WHERE login=’admin’ AND password=’XYZ’ OR ‘1=1’ When this query is received this is converted into XML format using a XML schema. The resulting XML would be, <SELECT> <*> <FROM> <members> <WHERE login= ’admin’> <AND password= XYZ> <OR 1=1> </OR> </AND> </WHERE> </members> </FROM> </*> </SELECT> Using the pattern matching the elements is searched so that the nested elements is similar to query tokens. The corresponding matching XML mapping is, <SELECT> <identifier> <FROM> <identifier> <WHERE id_list= ’userip’> <AND id_list=’userip’> </AND> </WHERE>
Velammal College of Engineering and Technology, Madurai
</identifier> </FROM> </identifier> </SELECT> When the match is found, the corresponding prototype query would be, SELECT identifier FROM identifier WHERE identifier op ‘userip’ AND identifier op ‘userip’ which will be used to identify the user input data. This search is less time consuming because the search is based on text and string comparison. The time complexity is O(n). This helps in increasing the effectiveness of the program and reduces the latency time. 2) Syntax analyzer: The syntax analyzer uses the prototype query model to recognize the user inputs and the this inputs are inserted in the parse tree which is generated by the parse tree generator. Now the syntax analyzer has the parse tree and the user inputs. Figure3 represents the parse tree for the benign query in section I. Every user input string can find a non-leaf node in the parse tree, such that its sub-tree leaf nodes comprise the entire user input string. An example application of the parse tree generator is given in figure 4. This algorithm for identifying malicious output is used in SQLProb[14].This system uses the validation algorithm to validate if the request is benign or malicious. The password field is parsed into the set Uni=1 (leaf(ui)) with five leaf nodes: XYZ, OR, 1, = and 1. The analyzer algorithm is given in Table 1. Next, we do depth-first-search from these five leaf nodes. The traversed paths intersect at a non-leaf node, SQLExpression. Finally, we do breath-first-search from SQLExpression to reach all the leaf nodes of the parse tree, which is a superset of Uni=1(leaf(ui)) implying that the input string ui is malicious. The algorithm described above takes quadratic time, because step 2 and step 3 take time of n × h, where n is the number of leaf nodes parsed by ui, and h is the average number of layers from leaf nodes to nl node in the parse tree. In addition, step 4 takes time complexity for a breathfirstsearch is O(n2). Therefore, the overall time complexity is O(n2). 3) Comparator Phase: The validator phase sends the result, indicating if the request is benign or malicious depending on the analyses it carried out, to the comparator. The comparator now has the results retrieved from the database and if the analyzer had certified that the request is benign without any delay the requested data is immediately send to the user. If the request is analyzed to be malicious, the comparator extracts the intended output of the model query from the Prototype document and compares it with the actual output from the application encoded in the HTTP response is database. If the result is the same, the requested data is send to the user as HTTP response. Else HTTP response with a ‘Data not available’ reply is sent to the user.
Page 339
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
SELECT * From members WHERE login=’admin’ and password=’xyz’
SELECT
select_list
identifier
FROM
Table_list
WHERE
Where_cond
identifier
SQLExpression
SQLAndExpression *
members SQLAndExpression
SQLAndExpressio
selectItem
ID
factor
Operator
login
selectItem
ID
ID
=
factor
Operator
=
password
admin
ID
XYZ
Figure 3: Parse tree for benign query
SELECT * From members WHERE login=’admin’ and password=’xyz’
SELECT
select_list
identifier
*
FROM
Table_list
identifier
Where_cond
SQLExpression
members
SQLAndExpression
SQLAndExpression
SQLAndExpressio
selectItem factor
ID
WHERE
Operator
ID
SQLAndExpressio
factor
factor
Operator
ID
ID
ID
Operator
1 login
admin
factor
selectItem
password
=
ID
1
=
=
XYZ
OR
Figure 4: Parse tree for malicious query
Velammal College of Engineering and Technology, Madurai
Page 340
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IV. A.
EVALUATION
Experimental Setup
The test suite used to attack is AMNESIA attack test suite [5]. This suite has a list of string patterns which are both benign and malicious. Using these attacks on the web application COMPAL achieved 100% detection rate and was able to prevent the malicious string patterns from extracting data from the database. The performance metric taken into consideration is the response time per web request. For every web request, the delay varied depending upon the query and its alignment. The response time for each web application used is shown in figure 5. TABLE I VALIDATION ALGORITHM
if Uni=1(leaf(ui)) Umk=1leaf(node)k) then Return True; else Return False; end end
0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0
Actual COMPVAL
st o
re
ir ok
pd
Bo
Em
Po
rt a
l
R e s p o n s e ti m e i n S e c
6. 7. 8.
V. CONCLUSION COMPVAL, as indicated in section 1, is a novel system used to detect SQLIA’s efficiently. The proposed system does not make any changes to the source code of the application and so the source code need not be known to COMPVAL. The proposed system the detection and prevention of SQLIA is fully automated. Prototype of COMPVAL was designed to measure the performance factor and it worked successfully with 100% detection rate. In future work, the focus will be on benchmarking the system with standard algorithm and studying of alternate techniques for the detection and prevention of SQLIA using dynamic analysis. REFERENCES
Data:Parse Tree Tree(qi) and the set of user input UI(qi) Result:True if the query is an SQLIA, or False if otherwise 1. for every user input UIi,j in UI(qi) 2. do depth-first-search upward from every leaf node leaf(ui) parsed from UIi,j , according to SQL grammar G; 3. Searching stops when all the searching path intersect at a non-leaf node nl node; 4. do breath-first-search downward from nl node until reaching all m leaf nodes leaf(node)k; 5.
Figure 5: The comparative response time
Web Applications
Velammal College of Engineering and Technology, Madurai
1. C. Anley, “Advanced SQL injection In SQL Server Applications”, White paper, Next Generation Security Software Ltd., 2002. 2. S.W. Boyd and A.D. Keromytis, “SQLrand:Preventing SQL Injection Attacks”, Proc. ACNS’ 04, pp. 292-302, June 2004. 3. Gregory Buehrer, Bruce W. Weide and Paolo A. G. Sivilotti, “Using Parse Tree Validation to Prevent SQL Injection Attacks”, Proc. International Workshop on Software Engineering and Middleware, 2005. 4. C.Gould, Z.Su and P.Devanbu, JDBC Checker: A Static Analysis Tool for SQL/JDBC Application, Proc. International Conference on Software Engineering ‘04, pp.697-698, 2004. 5. W. G. Halfond and A. Orso, “AMNESIA: Analysis and Monitoring for NEutralizing SQL-Injection Attacks”, Proc. ACM International Conference on Automated Software Engineering ‘05, November 2005. 6. Y.Huang, F. Huang, T.Lin and C.Tsai, “Web Application Security Assessment by Fault Injection and Behavior Monitoring”, Proc. International World Wide Web Conference ‘03, May 2003. 7. V.B. Livshits and M.S. Lam, “Finding Security Errors in Java Programs with Static Analysis”, Proc. Usenix Security Symposium ‘05, pp. 271-286, August 2005. 8. Y.Huang, F.Yu, C. Hang,C.H. Tsai, D.T.Lee and S.Y.Kuo, “Securing Web Application Code by Static Analysis and Runtime Protection”, Proc. International World Wide Web Conference ‘04, May 2004. 9. D.Scott and R.Sharps, “Abstracting Application-level Web Security”, Proc. International Conference on the World Wide Web ‘02,pp. 396-407, 2002. 10. Zhendong Su and Gary Wassermann, “The Essence of Command Injection Attacks in Web Applications”, Proc. ACM SIGPLANSIGACT Symposium on Principles of Programming Languages ‘06, January 2006. 11. Burp proxy. [Online]. Available: http://www.portswinger.net/proxy. 12. JJTree. [online]. Available: http://javacc.dev.java.net/doc/JJTree.html. 13. W. Halfond, J. Vigeas and A.Orso, “A Classification of SQL Injection Attacks and Counter Measures”, Proc. International Symposium on Secure Software Engineering ‘06, March 2006. 14. Anyi Liu, Yi Yuan, Duminda wijesekera and Angelos Stavrou, “SQLProb: a Proxy-based Architecture Towards Preventing SQL Injection Attacks”, Proc. ACM symposium on Applied Computing ‘09, pp. 2054-2061, 2009. 15. W. Halfond, A. Orso and P. Manolios, “Using Positive Tainting and Syntax-Aware Evaluation to Counter SQL Injection Attacks”, Proc. ACM SIGSOFT Symposium on the Foundations of Software Engineering ‘06, November 2006.
Page 341
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
16. Adam Kieyzun, Philip J. Guo, Kartick Jayaraman and Micheal D. Emst, “Automatic Creation of SQL Injection and Cross-Site Scripting Attacks”, Proc. IEEE International Conference on Software Engineering ‘09, pp. 199-209, 2009. 17. Konstantinos Kemalis and Theodores Tzouramanis, “SQLIDS: A Specification-Based Approach for SQL-Injection Detection”, Proc. ACM symposium on Applied computing ‘08, pp. 2153-2158,2008. 18. Xiang Fu and Kai Quian, “SAFELI: SQL Injection Scanner using Symbolic Execution”, Proc. Workshop on Testing, Analysis and Verification of Web Services and Applications ‘08, pp. 34-39, 2008. 19. M.Johns and C. Beyerlein, “SMask: Preventing Injection Attacks in Web Application by Approximating Automatic Data/Code Separation”, Proc. ACM Symposium on Applied Computing ‘07, March 2007.
Velammal College of Engineering and Technology, Madurai
Page 342
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Integrating the Static and Dynamic Processes in Software Development V. Hepsiba Mabel#1, K. Alagarsamy*2, Justus S$3 #
Fatima College, Maryland Madurai 1
[email protected]
*
Madurai Kamaraj University Madurai
2
[email protected]
$
Velammal College of Engineering & Technology Madurai 3
[email protected]
Abstract— Software development is caught up in bunch of processes that involve project, product and people. Organizations those are into the development of software products and providing software solutions are bound by sets of processes that governs the activities of the project as well of the organizations. CMM and other standards serve as very important guidelines to companies that are in the software development. The entire processes are based on the IDEAL model, and those processes are defined static processes. Also, recent researches in Knowledge management show that knowledge based processes are deemed dynamic in nature and help the processes to improve over period of times. However, these processes are required to be integrated into one group or framework, where the practitioners will find their entire set of processes derived and defined in the CMM levels might be easy to improvement. In this work, we have integrated the static and dynamic process sets into one framework and have partially implemented them in companies who are interested in practicing the framework. Keywords— Software Processes, Process Framework, CMM.
I. INTRODUCTION The static and dynamic processes involved in the development of software products are of significant concern in research and in practice. Since most of the activities involved in software development are loosely coupled they need to be integrated to form a single framework. Since the beginning of software development and software engineering in the early 1960s, research on the activities and the processes governing the design and development of software products are active; and to till date they are active. However, it is studied from the literature [3], [7] [12] [23], [26], [30], and case studies only a limited number of publications are available for the processes that governs in software development, when compared to other areas of research. Also there is a need for a rigorous study on the process sets and key process areas that forms the capsule of
Velammal College of Engineering and Technology, Madurai
the process standards and maturity models. This notion of process and process set is a dynamic entity, because the process of software development is world wide and it crosses all continental and cultural backgrounds. Hence the process sets defined by [2] [22], are not standardized internationally. However, we need to bring the KPAs and process sets, both defined and tailor-made, into one single framework. This work deals in detail the formation of the single framework. Organizations involved in software development are concerned with delivery of quality software products within the specified timeline. Development of a software product is a time-based, distinct activity which is termed as project. A project is governed by a set of defined and tailor-made processes that controls the quality end-product and ensures that the product is delivered ontime and without overruns on the budget. To enable all the implicit processes involved in software development, project leaders and managers need to have a decision supportive intelligent system that falls into a framework for improving their software developmental processes. Motivation of the Work When in the early years of software development, there was only a limited steps and algorithms governing the developmental processes. As years and decades passed by the software development industry faced with numerous processes, processes that are anticipated, expected and unexpected, were evolved. These processes defined and regulated were not adopted by the industry for its lack in flexibility. Hence the organizations attempted to tailor-make the processes according to their organization’s need and projects’ requirements. While most of the processes defined and standardized by Software Engineering Institute [26], [30], were acceptable by software developing companies, many were deemed to be redefined and refined as years passed by. This is commended
Page 343
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
by the global, local markets, competitors, governments’ policies and other factors that arise internally and externally. Out of these changes, the software development companies and the research institutes were able to identify the two dimensions of the entire process sets. They are classified based on the nature of the processes as Static Processes and Dynamic Processes [2]. Both these process sets are formally validated by international research institutes. However the static and dynamic process sets stand alone in the practicing platform. Organizations prefer the static processes than dynamic one for their simplicity and ease of implementation and practice [23]. Although it is encouraging and result reaping, dynamic processes are recommended. This is because, dynamic processes yields good results, and provides and wider scope for process improvement. Hence we attempt in this work to integrate these two process sets and form a single framework of processes, that will help the organizations in building up their process base and leaves better opportunity for process improvement. Related Works As pointed out earlier, research works on software development processes have found a prominent position. Software Engineering Institutes [26], [30] is the major role player in proposing standards like CMM, CMMI. CMM family of standards speaks on the refinement of processes and the level of maturity with which the processes have improved. ISO9000 family is a very specialized set of standards which mainly focuses on the processes in the management of software processes and their improvement. SPICE, ISO/IEC12207, BOOTSTRAP and Rational Unified Process (RUP) are other similar standards which centers the processes and the collection of processes [12]. But organizations follow these standards just as a roadmap for their journey in the process improvement programs. Research literatures and papers on processes sets and areas cover a wide variety of topics with regard to integration and improvement [1], [2], [7], [11], [14], [24], [31]. Similar works on processes sets are still in research and study.
process becomes better defined and more consistently implemented throughout the organization. Software process capability describes the range of expected results that can be achieved by following a software process. The software process capability of an organization provides one means of predicting the most likely outcomes to be expected from the next software project the organization undertakes. Software process performance represents the actual results achieved by following a software process. Thus, software process performance focuses on the results achieved, while software process capability focuses on results expected. Software process maturity is the extent to which a specific process is explicitly defined, managed, measured, controlled, and effective. Maturity implies a potential for growth in capability and indicates both the richness of an organization's software process and the consistency with which it is applied in projects throughout the organization. As a software organization gains in software process maturity, it institutionalizes its software process via policies, standards, and organizational structures. Institutionalization entails building an infrastructure and a corporate culture that supports the methods, practices, and procedures.
A. Basics of Software Processes A software process can be defined as a set of activities, methods, practices, and transformations that people use to develop and maintain software and the associated products [23] (e.g., project plans, design documents, code, test cases, and user manuals). As an organization matures, the software
Processes and Maturity Levels The formula for success of a software development organization lies in the continual improvement of its processes and leveraging its effectiveness for business excellence [18]. As discussed previously, the learning in the organization builds knowledge and warehouses which are the help guides for process improvement. Initiating an improvement program in the organization starts from the individual; that is when the individual realizes his improvement, will he appreciate the need for the organization’s improvement. Organization’s improvement is primarily based on the improvement in the processes. Here we consider the improvement of software process. Software process is defined as a set of activities that begins with the identification of a need and concludes with the retirement of a product that satisfies the need; or more completely, as a set of activities, methods, practices, and transformations that people use to develop and maintain software and its associated products [15]. To improve the software development process, it is necessary to evaluate the success of the project and avoid repeating problems in the future. The CMM addresses these latter, less well-understood, issues in an effort to improve the software development process [10]. Each represents an improvement in the software process. An organization’s software capability can be improved by advancing through these five stages or levels. The CMM helps organizations to select improvement strategies based on current process maturity status and to identify critical issues in quality and process improvement [26]. To help small organizations, possessing only limited resources, to reach the CMM level 3 and improve certain helpful tips are formulated
Velammal College of Engineering and Technology, Madurai
Page 344
II. PROCESSES IN SOFTWARE DEVELOPMENT The process of developing software has also been under research and practice since the evolution of software. The processes and activities involved in software development is systematic and unique endeavor that use sound engineering principles in order to obtain reliable, quality software product in an economically managed timeline.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
based on the discussions given by Kathleen [15]. It is observed from surveys, the SMBs in Chennai, India, are content with what they are performing. But when they complete for global projects they fail because they have never opted for CMM standards in any of the developmental strategies of business process. If at all they attain CMM level 3, they are content in that level because of lack of action plans and further improvement initiatives. The levels of process maturity for improving organizations is given in table 1. Organizations planning for an improvement endeavor will have to identify the business and developmental process that addresses organization goals, thus identifying the key-process areas. The accepted process sets has to be tailored according
to the requirement of the project or for the implementation of activities that support specific needs, resulting in a process protocol. Then lies the real process improvement that is goal oriented and knowledge driven, leading to CMM improvement level 3. Beyond this the organization has to extend its process set by refining them, ensuing in process reusability. When the componential process sets are combined and externalized they result in the regeneration of new process sets called hybrid processes, thus leading the organization to the highest level of process optimization. All of these states of maturity are driven by the activities of knowledge management.
TABLE I LEVELS OF PROCESS MATURITY
Process Maturity Levels Process Initiation Process Development Process Improvement Process Extension Hybrid Processes
Initial State
Actions
Implementation
Static
Address Goals
Process Set
Develop Activities
Dynamic Process Improved Process Set Componential Dynamic Processes
Use dedicated resources Refine Processes and Sets Plan and apply
Engineering Practices Activities for Specific needs Use Tools for improvement Expand to other projects Derive new process sets
Result Identified KPAs Process Set CMM level improvement Process reusability Process regeneration and improvement
Static Processes The processes involved in software development are deemed for practice and improvement. Still some of the processes remain unchanged and they are termed as Static Processes. These processes mainly deal with the developmental activity of the software and the processes that govern them. There may be many factors that influence them, but whatever be the external factors these processes have to remain static throughout the software development life cycle [4]. Attempt to
change these static processes is not advised, because that may lead to unwarranted, irreversible damage to the project or product or to the organization. A process set identified in the second process maturity level is often static. This set will be the basis for all the other derived and hybrid processes that may arise during the future endeavors of the organization [8]. While this static process set is static and cannot be changed, they are deemed for improvement during maturity level 3 and the following improvement programs. Table 2 lists the set of static processes. The five set of processes identified and defined are termed as static. They are bound within the scope of the project and are subject to be refined but not changed. This set of processes is defined based on the IDEAL model [16], [23]. The IDEAL model, framed for SPI program, has led to the derivation of a knowledge driven model which focuses on the acquisition of explicit knowledge through externalization and its conversion to tacit knowledge through internalization and continuous learning [10], [17]. Though many KM tools comprise, both smaller, specific applications, and enterprise level tools, the urge to provide a compelling learning environment should be felt [19].
Velammal College of Engineering and Technology, Madurai
Page 345
III. A SINGLE PROCESS FRAMEWORK The discussion in the previous chapter on the processes and their integration with the maturity levels in software development would well be a basis for the integration of the two sets to activities, static and dynamic. The proposal and the practice of static activities by software organizations improve in their maturity level. But the proposal, practice and the promotion of dynamic activities helps the organization to improve software processes based on knowledge gained over the years of experience and the concept of sharing them [2]. This chapter deals in detail the two sets of processes and integrating them.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE II STATIC PROCESS SET
Phases of the Process Set Development of Software Maintenance of the end product Software Processes Production Processes Management Processes
Static Processes Design and Coding Processes Finish the product on time and within budget Finished product maintenance process Expires at the end of the scope Stick on to the processes defined for development and maintenance Document preparation processes Populate Repositories Plan and Control Use Decision support tools
Hence a need-based learning process is introduced in the establishing phase of KDM through packaging, which resolves the setback of knowledge sharing. In the Acting phase, the derived knowledge is executed through effective planning, based on the earlier experiences and lessons learnt from previous similar projects. B. Dynamic Processes In continuance with the definition of static processes, there arises certain processes as a result of improvement or refinement. These newly identified processes are dynamic processes as they are subject to refinement then and there whenever required and demanded by the project or by the organization. This set of processes involves and demands the required component called interaction and sharing. Exchange of intellectual property and experience gained over the years are the major driving catalysts for the dynamic process set. Transfer of knowledge and experience through tacit or explicit or even implicit will take place in this domain of processes. This set of activities is derived based on the knowledge driven model proposed in [2]. More often these set of activities are practiced and promoted in the fourth and fifth process maturity level and keeps going for betterment and refinement. The dynamic process set consists of the following processes. Table 2 mentions them. Most or even all of the dynamic activities, mentioned in the table, are specific and centered on knowledge and its associated components. TABLE III DYNAMIC PROCESS SET
Phases of the Process Set Development of Software Maintenance of the end product Software Processes
Dynamic Processes Interaction and Sharing of resources Exchangeability of gained experience Document revision & refinement Interaction within the
Velammal College of Engineering and Technology, Madurai
Production Processes
Management Processes
development team Improvement and Generation of new processes Codify data to repositories Generate feedback to the development team Interaction with project leaders Share expertise knowledge Generate hybrid processes
While static activities are derived from CMMI or other existing conventional standards, dynamic activities are based on interactions of knowledge workers and knowledge users, and knowledge activities like codification and populating the knowledge repositories, which in later phases help the management in institutionalizing and implementing the newly found hybrid knowledge for planning and controlling the process improvement program [1], [30]. The dynamic activities are always driven by the knowledge generated by the experiences and lessons learnt from earlier and similar projects. The knowledge based components in the software development organizations lay spread in their existence and there seems to be a gap among them in the organizational chart. But functionally they stay integrated in their operational platform. Integrating the components across the nature of data repositories, cultural, continental and psychological differences of the knowledge workers and users remains a great challenge of the SPI program [31], [32]. The process activities, both static and dynamic, serve as a link in integrating the components according to the procedures that organizations adopt for the software process improvement. It is inferred that the dynamic knowledge driven activities have a significant impact on the software processes, identified for improvement. This can be verified by measuring the functionality of the knowledge components that endure integration. In a wider spectrum, there are 22 process areas [10], where all of them can be designated to be the set of activities either as static or dynamic. Conceptually, the degree of impact of the dynamic components on these processes is proportional to the estimated functional effort of the integrated components. More precisely, higher the degree of interaction among the components more the dynamic processes (activities) is involved in action, causing the entire set of processes to improve. Operation and implementation of this knowledge centered SPI program is a major challenge development companies confront with in practice. C. Integrating the Processes into Framework An attempt to integrate the two divisions of Processes is an important task and is a careful endeavor which requires a better understanding of the entire process sets. 1) Development of Software: The static processes associated with this phase is design and coding and to finish the software product in time and within the budget. While design and coding are matters concerning technical talk, the later part
Page 346
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
finishing the product within time and budget is the great challenge that is prevailing since the start of the software development. Due to these two serious issues software projects are subject to failure or partial completion [13], [31]. Finishing a successful project is the requirement and demand from the project mangers. One of the measures of project success is completing the project within timeline and without budget overrun. This problem can be solved only when the dynamic set of activities; Interaction, Sharing of resources Exchanging the gained experience The design tools and techniques, the coding methodologies and documents are the resources that are subject to be shared among the team members and also among project groups. The exchange of knowledge and experiences gained over the years by senior managers and tech leads should be shared among the team members. This experience will help the juniors and the lower-level work groups to finish their development and coding in time and also within the budget. A study in software companies show that this kind of sharing through knowledge management systems help the organization improve the productivity of the project teams and of the individuals [19], [20]. 2) Software Processes: Whenever a software project has been initiated, the project planning phase does several important things. Among them a few are defining the processes and choosing appropriate process models The processes that are associated with the development of software are identified and defined at the earlier stages of software development life cycle. Then choosing an appropriate model for the software development is an important task the project mangers are confronting with. In practice we could see that out of the 13 process models available in literature [10], the practitioners often choose just two or three and customize them according to the requirement of the nature of the project. This activity of choosing the process models is termed as static activity because this is usually just declaring and defining the process models which are usually a common and a routine practice. However this alone is not sufficient for the software product to be successful. One has to be highly dynamic and active while in the process of software development. Hence interaction among the team members and across the levels of hierarchy of people involved in the same project is required. This creates newer ideas and innovative concepts of software development. The interaction among the team members and among the members involved in the project leads to the development of newer processes, which we deem as hybrid processes [28]. The existing and the hybrid processes are subject to improvement as time moves on and the interaction increases. 3) Production Processes: Production in software development is the number of lines of code generated during the construction of code and also the amount of work done in a given time frame. This includes finished product along with
Velammal College of Engineering and Technology, Madurai
the necessary designs and the document files. Documentation is an embedded process in the process of software development. Every project under progression should posses a knowledge repository. This is an additional document and the forth coming knowledge management system for the project and for the organization. The process of populating the repositories is the requirement for any project that considers itself serious and which contributes to the future of the organization. These two processes are termed as static because they are required components for the project. The repositories are populated with information which is then analyzed, processed and codified as knowledge units in the knowledge management systems (KMS). The codification of the information to be represented as knowledge unit is a dynamic process, because information that is generated during the project progress has to be identified and should be sent for codification in the KMS. This generated knowledge units in the KMS will help the system to generate feedback to the developers tuning them to produce quality softwares and documentation processes. IV. FINDINGS FROM IMPLEMENTING THE FRAMEWORK This section shows the findings from the experiments that are conducted as implementation part of the process framework. Critical Success Factors Table 4 shows the list of CSFs cited in the literature and in CSF interviews. These factors are listed in alphabetic order in the table 4 and are associated with the numbers in the figure 1. The most frequently cited factor in the literature is staff involvement, i.e. 46%. TABLE IV CRITICAL SUCCESS FACTORS - ANALYSIS
Critical Success Factors Creating process action teams Interactions Formal Methodology Experimenting SPI Process Awareness Process Ownership Hybrid Processes Staff Involvement Tailoring processes
Occurrence in Literature (n=47) Frequency % 7 26
Occurrence in Practice (n=23) Frequency % 2 9
12 -
21 -
5 8
22 35
4
15
3
13
9
24
12
52
3
37
-
-
12
26
4
11
24
46
10
54
18
28
2
9
This suggests that in involvement of staff can play a vital role in the implementation of improvement programs. Other
Page 347
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
frequently cited factors in the literature are Process ownership (37%), and tailoring processes (28%). It shows that practitioners consider their involvement, process ownership, and tailoring imperative for the successful implementation of improvement programs. The results also show that staff time and resources and creating process action teams are also important factors. A quarter of the literature cited reviews, experienced staff, clear and relevant process goals and assigning of responsibilities.
methodology Lack of resources Organizational politics Time pressure
7
50
8
35
4
29
12
52
5
36
4
17
Fig. 2 Critical Barriers in Process Framework
Fig. 1 Critical Success Factors in Process Framework
Table 4 shows that, like the literature, the most frequently cited factors in the interviews are staff involvement and processes awareness, i.e. cited 54% and 52% respectively. Two new methods––awareness and formal methodology–– have been identified in our empirical study which has not been identified in the literature. Other frequently cited factors are resources, staff involvement and experienced staff. Other factors are less frequently cited in the interviews. Critical Barriers Our aim of identifying critical barriers is to understand the nature of issues that undermine the process framework implementation programs. Table 5 shows the list of critical barriers cited in the literature and an empirical study. The results show that most of the practitioners in literature consider lack of resources a major critical barrier for the implementation of SPI. The results also suggest that in practitioners’ opinion time pressure and inexperienced staff can undermine the success of SPI implementation programs. It shows that practitioners would prefer to avoid organizational politics during the implementation of process framework. TABLE V CRITICAL BARRIERS - ANALYSIS
Critical Barriers Inexperienced staff Lack of awareness Lack of formal
Occurrence in Literature (n=14) Frequency % 5 36
Occurrence in Practice (n=23) Frequency 4
% 17
-
-
10
43
-
-
12
52
Velammal College of Engineering and Technology, Madurai
Staff involvement is ranked highest in CSF interviews, i.e. 54%. Two new critical barriers––lack of formal methodology and lack of awareness––have been identified in our empirical study which has not been identified in the literature. The second most cited critical barrier in CSF interviews is lack of support. The critical barrier ‘‘lack of resources’’ is cited 35% in the CSF interviews. Comparison of the critical barriers provides evidence that there are some clear similarities and differences between the findings of the two sets. There are barriers in common, i.e. inexperienced staff, lack of resources, lack of support, organizational politics, and time pressure. There are also a number of differences between the findings. For example, ‘‘changing the mindset of management and technical staff’’ and ‘‘staff turnover’’ have not been cited in our empirical study but these barriers are present in the literature. Similarly, lack of awareness of SPI and lack of formal methodology are critical in our empirical study but have not been identified through the literature. This shows that practitioners, who took part in our study, are more concerned about SPI awareness activities and implementation methodology. Implications of the Process Framework In the previous sub-section we dealt with three subdivisions that commend the static and dynamic processes. This framework we have derived out of integrating the two process sets are left for discussion and validation, which is presented in the following sections. 1) Comment on the Framework Now we put the key process area in the process framework up for discussion. The first question that needs to be answered is whether a software developers’ success relies on static or dynamic processes. In the commercial cases encountered and presented in this work 50%-70% of their effort in software development rely on the existing, well-defined static processes. However to keep their production in the uprising, they need to work on the dynamic
Page 348
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
processes and help the development processes improve over time and experience. In the case of open source products, where many of the users of the product are also developers, testers, and quality assurance team members, the same premise on the process framework holds. Since framework is a process focused one, the improvement of these processes will lead to better productivity. The second thought is whether this process framework could be followed with relevant to software development. 2) Validation of the Process Framework Usability is formally defined as the “The extent to which a product can be used by specified users to achieve specified goals with effectiveness, efficiency and satisfaction in a specified context of use’’ (ISO/DIS-9241-11, 1994). The goal of a usability analysis is to pinpoint areas of confusion and ambiguity for the user which, when improved will increase the efficiency and quality of a users’ experience (ISO/DIS-9241-11, 1994). In order to measure the usability of our framework implementation maturity model we will be using the following three goals or criteria: 1) ease of learning, 2) ease of use and 3) satisfaction. In order to validate our process framework we are planning an expert panel review process. Similar process has been used by Beecham and Hall (2003) for the validation of requirements process improvement model. An expert panel questionnaire will be sent along with the description of the process framework. This feedback will be analyzed and enhanced version of the framework will be produced based on the evaluation feedback. After the expert panel review process we are also planning to conduct a case study in a software development company in order to validate and test the framework in the real world environment.
We opt to draw a roadmap by integrating the processes that are deemed for improvement and continuous monitoring. The SPI implementation framework is in its initial stage of development and needs further improvement and evaluation. A case study will be conducted in order to test and evaluate this framework and to highlight areas where this framework can be improved.
V. CONCLUSIONS AND FUTURE WORKS In this work a Process Framework for Software Development is presented that has the potential to help companies assess and improve their process improvement program. This framework is extracted from CMMI and is based on CSFs and critical barriers identified through literature and an empirical study. For the development of this framework we have analyzed the experiences, opinions and views of practitioners in order to identify factors that have a positive or negative impact on the implementation of a improvement programs. However, this model needs further improvement and evaluation. An expert panel review will be conducted in order to get feedback from different practitioners. Based on the evaluation feedback enhanced version of the maturity model will be produced. Future Works: A case study will also be conducted in order to validate and test this process framework in practice. Our ultimate aim of conducting empirical study and developing maturity model was to develop a SPI implementation framework. We have used our empirical findings and designed a framework for guiding the design of effective implementation strategies for software process improvement.
REFERENCES [1] Alagarsamy. K., Justus S and K. Iyakutti, “Implementation Specification of a SPI supportive Knowledge Management Tool”, IET Software, Vol. 2, No.2, pp. 123-133, April 2008. [2] Alagarsamy. K., Justus S and K. Iyakutti, “Knowledge based Software Process Improvement”, Proc. of IEEE International Conference on Software Engineering Advances, France, 2007. [3] Anne Walker, and Kathleen Millington “Business intelligence and knowledge management: tools to personalize knowledge delivery” Information Outlook, August, 2003. http://www.findarticles.com/p/articles/mi_m0FWE/is_8_7/ai_ 106863496 [4] Baccarini. D, “The logical framework method for defining project success”, Project Management Journal, pp.25-32 1999. [5] Baddoo, N., Hall, T., “De-motivators of software process improvement: an analysis of practitioner’s views”, Journal of Systems and Software, Vol. 66, No. 1, pp. 23–33, 2003 [6] Baddoo, N., Hall, T., Wilson, D., “Implementing a people focused SPI programme”, Proc. in 11th European Software Control and Metrics Conference and the Third SCOPE Conference on Software Product Quality, 2000 [7] Beecham, S., Hall, T., Rainer, A., Building a requirements process improvement model, Department of Computer Science, University of Hertfordshire, Technical Report No. 378, 2003. [8] Belout. A, “Effects of human resource management on project effectiveness and success: toward a new conceptual framework”, International Journal of Project Management, Vol.16, pp. 21–26, 1998. [9] Coolican, H., Research Methods and Statistics in Psychology. Hodder and Stoughton, London, 1999. [10] CMMI of Development, ver 1.2, SEI, Pitsburgh, 2006, PA 15213-390, CMU/SEI-2006-TR-008, 2006. [11] David N Card, “Research Directions in Software Process Improvement”, Proc. in 28th Annual International Computer Software and Application Conference, 2004. [12] Emam. K.E., J.N. Drouin, and W. Melo, “SPICE:The Theory and Practice of Software Process Improvement and Capability Determination”, IEEE Computer Society, Los Alamitos, 1998. [13] Haris Papoutsakis, and Ramon Salvador Vallès, “Linking Knowledge Management and Information Technology to Business Performance: A Literature Review and a Proposed
Velammal College of Engineering and Technology, Madurai
Page 349
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Model”, Journal of Knowledge Management Practice, Vol. 7, No. 1, March, 2006. [14] Jesper Arent, and Jacob Norbjerg., “Software Process Improvement as Organizational Knowledge Creation: A Multiple Case Analysis”, Proc. Of the 33rd Hawaii International Conference on System Sciences, 2000. [15] Kathleen Coleman Dangle, Patricia Larsen, and Michele Shaw, “Software Process Improvement in Small Organizations: A Case Study”, IEEE Software, pp 68-75, 2005. [16] Kautz. K, W.H. Henrik and Kim Thaysen, “Applying and Adjusting a Software Process Improvement Model in Practice: The Use of the IDEAL Model in a Small Software Enterprise”, Proc. International Conference on Software Engineering, Limerick, Ireland, SACM 23000 1-58113-2069/00/06, 2000. [17] Kellyann Jeletic, Rose Pajerski and Cindy Brown, Software Process Improvement Guidebook, NASA, Greenbelt, Maryland, March 1996. [18] Knowledge Management, APQC Online Discussion, 2005. http://www.apqc.org/portal/apqc/site/generic?path=/site/km/o verview.jhtml [19] Knowledge Management and e-Learning Integration, 2004. “http://knowledgemanagement.ittoolbox.com/documents/popu lar-q-anda/knowledge-management-and-elearning-integration2846”. [20] Koren. E.F. and S. Elisabeth Hove, “Using a Knowledge Survey to plan Software Process Improvement Activities – a Case Study”, 2002, http://heim.ifi.uio.no/~isu/INCO/Papers/Using_efk_eurospi20 02_paper.pdf [21] Krippendorf, K., Content Analysis: An Introduction to its Methodologies, Sage, London, 1980. [22] Marwick. A.D., “Knowledge Management Technology”, IBM System Journal, Vol. 40, No 4, 2001. [23] McFeeley. B., IDEAL: A User's Guide for Software Process Improvement, SEI, Pittsburgh, Handbook CMU/SEI96-HB-001, 1996. [24] Medha Umarji and Carolyn Seaman, “Predicting Acceptance of Software Process Improvement”, Human and Social Factors of Software Engineering, St. Louis, Missouri, USA, 2005. [25] Michael, S., Lewis, B., Research Practice: International Handbooks of Quantitative Application in Social Sciences, Sage, 1994. [26] Paulk, Mark C. & Chrissis, Mary Beth, The 2001 High Maturity Workshop (CMU/SEI-2001-SR-014). Pittsburgh, PA: Software Engineering Institute, Carnegie Mellon University, January 2002. http://www.sei.cmu.edu/publications/documents/01.reports/01 sr014.html. [27] PMI Standards Committee, William R Duncan, A Guide to the Project Management Body of Knowledge, Project management Institute, 1998.
[28] Pourkomeylian. P., Software Process improvement, Doctoral Dissertation, Goteborg University, ISSN 1400-741X, 2002. [29] SEI, 2002. Capability Maturity Model Integration, (CMMISM), Version 1.1. SEI, CMU/SEI-2002-TR-029. [30] Stelzer, D., Werner, M., “Success factors of organizational change in software process improvement”, Software Process Improvement and Practice, Vol. 4, No. 4, 1999. Wesley Vestal., Integrating Knowledge Management and Organizational Learning, American Productivity & Quality Center, 2005. http://www.apqc.org/portal/apqc/site/content?docid=11978
Velammal College of Engineering and Technology, Madurai
Page 350
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Exploiting Parallelism in Bidirectional Dijkstra for Shortest-Path Computation R. Kalpana#1, Dr. P.Thambidurai#2, R. Arvind kumar #3, S. Parthasarathi #4, Praful Ravi #5 #
Department of Computer Science & Engineering , Pondicherry Engineering College Puducherry, India 1
[email protected]
Abstract— The problem of finding a shortest path between two given nodes in a directed graph is a common task which is traditionally solved using Dijkstra’s algorithm [1]. There are many techniques available to speedup the algorithm while guaranteeing the optimality of the solution. Almost all of these speed-up techniques have a substantial amount of parallelism that can be exploited to decrease the running time of the algorithm. The rapidly evolving field of multiprocessing systems and multi-core processors provide opportunities for such improvements. The main focus of this paper is to identify features of the Bidirectional search speed-up technique suitable for parallelization and thereby reduce the running time of the algorithm. Keywords— Dijkstra’s Algorithm, algorithm, running time, speed-up
Graph
Theory,
parallel
VI. INTRODUCTION Computing shortest paths is a basic operation which finds application in various fields. The most common applications[2] for such a computation includes route planning systems for bikes, cars and hikers, and timetable based information systems for scheduled vehicles such as buses and trains. In addition to these, shortest path computation also finds its application in spatial databases and even in web searching. The core algorithm that serves as the base for the above applications is a special case of single-source shortest-path problem on a given directed graph with non-negative edge weights called Dijkstra’s algorithm[1]. The particular graphs[3], [4], [5] considered for the above applications are huge and the number of shortest path queries to be computed within a short time is huge as well. This is the main reason for the use of speed-up techniques for shortest-path computations. The focus of these techniques is to optimize the response time for the queries, that is, a speed-up technique is considered as a technique that reduces the search space of Dijkstra’s algorithm by using pre-computed values or by using problem specific information. Often the underlying data contains geographic information, that is, a layout of the graph is easily obtained. In
Velammal College of Engineering and Technology, Madurai
many applications the given graph can be considered static which enables preprocessing. Another possibility for improving the runtime of speed-up techniques is to identify the regions, if any, that are suitable for parallel processing. It is observed that the various speed-up techniques posses, to varying extents, operations that are possible candidates for parallel execution. Coupled with the improvement of processing capabilities of today’s modern systems the time required for computing shortest path can be effectively reduced. Also the speed-up techniques are such that, combination of two or more techniques is possible and the effective speed-up achieved by the combination is increased. In this paper, a study of the various speed-up techniques is made and the possible parallel regions present in speed-up techniques are analyzed. Experiments [6] were carried out to analyze the performance of a particular speed-up technique with the parallelism present in it utilized in different ways. The Parallelization adapted in Highway node routing [6] for solving shortest path problem gives a very good speedup compared to that of sequential algorithm. The bidirectional search is considered as one of the basic speedup techniques, which improves the speedup by searching the graph both forward and reverse. Using the OpenMp constructs with LEDA library the speedup of the algorithm is improved. VII. PRELIMINARIES The formal definitions and a brief description of Dijkstra’s Algorithm are given in this section. A. Definitions A graph G is a pair (V, E), where V is the set of nodes / vertices and E ⊆ V × V is a set of edges, where an edge is an ordered pair of nodes of the form (u, v) such that u , v ∈ V . Usually the number of nodes V is denoted by n and the number of edges E is denoted by m. For a node u ∈ V , the number of outgoing edges {( u , v ) ∈ E } is called the degree of the node.
Page 351
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A path in graph G is a sequence of nodes (u 1 ,..., u k ) so that (u i , u i +1 ) ∈ E for all 1 ≤ i < k . A path in which u 1 = u k is called a cycle or cyclic path. Given the edge weights l : E → R , the length of the path P = ( u 1 ,..., u k ) is the sum of the lengths of its edges l ( P ) :=
∑
1≤ i < k
l ( u i , u i +1 ) .
For two nodes s , t ∈ V , a shortest s-t path is a path of minimal length with u 1 = s and u k = t . The distance d ( s , t ) between s and t is the length of the shortest path s-t. A layout of a graph G = (V , E ) is a function L : V → R 2 , that assigns each node a position in R 2 . A graph is called sparse if m = O (n ) . B. Shortest Path Problem Consider a directed graph G = (V , E ) whose edge weights are given by a function l : E → R . The single-source singletarget shortest path problem is to find a shortest s-t path from a source s ∈ V to a specified target t ∈ V . The problem is said to be well defined for all pairs, if G doesn’t contain cycles with negative lengths. If negative weights are present but not negative cycles, using Johnson’s algorithm, it is possible to convert the original edge weights l : E → R to non-negative weights l ' : E → R 0+ that results in the same shortest paths. So it is safe to consider that edge weights are non-negative. The Dijkstra’s algorithm[1] maintains, for each node u ∈ V , a label dist(u) with the current distance values. A priority queue Q contains the nodes that form the search around source s. Nodes can be either unvisited, in the priority queue or already visited. In the computation of shortest path it has to be remembered that u is the predecessor of v if a shorter path v has been found. Dijkstra’s algorithm is used for computing shortest paths from a single source to all other nodes in the graph. If only one shortest path is required to a given target t ∈ V , the algorithm can halt when the target is eliminated from the priority queue. The time complexity of the algorithm is decided by the priority queue chosen for implementation. For usual graphs, Fibonacci heaps, provide the best theoretical worst-case time of O ( m + n log n ) . Binary heaps also result in the same time complexity for sparse graphs. It is to be noted that binary heaps are easier to implement and also performs better in many practical situations. It is arguable that a better speed-up technique reduces the search front and the priority queue will be of lesser importance in such a case.
A. Bidirectional Search In bidirectional search[2] a normal or forward variant of the algorithm starting at the source and a reverse or backward variant of the algorithm starting at the destination are run simultaneously. The backward variant of the algorithm is applied to the reverse graph, that is, a graph having the same node set V as the given graph and the reverse edge set E = {(u , v ) | ( v , u ) ∈ E }. The algorithm can be terminated when a node has been labeled permanent by both the forward and backward searches. The shortest path can be obtained by combining the path obtained in the forward search with the one obtained in backward search. B. Goal-Directed Search This technique [2], [8] alters the priority of the active nodes to change the order in which the nodes are processed. More precisely a potential p t (v ) depending on a target ‘t’ is added to the priority dist (v ) . So the modified priority of a node v ∈ V is given by dist ( v ) + p t ( v ) . With an appropriate potential the search can be directed to the target thereby reducing the running time and at the same time finding a shortest path. An alternate approach to modify the priority, is to change the edge lengths so that the search is directed towards the target ‘t’. In this case, the length of an edge ( u , v ) ∈ E is replaced by l ' (u , v ) = l (u , v ) − p t ( v ) + p t (u ) . It can be easily verified that a path from s to t is a shortest s-t path according to l ' , if and only if it is a shortest s-t path according to l. C. Multilevel Approach This speed-up technique has a preprocessing step in which the input Graph G is decomposed into l + 1 levels ( l ≥ 1) and enriched with extra edges which represent shortest path between certain nodes[2],[9],[10]. The decomposition of the graph depends on the selected nodes Si, at level i such that S 0 := V ⊇ S 1 ⊇ K ⊇ S l . The decomposition of the node sets can be based on various criteria. Selecting desired number of nodes with the highest degree in the graph, works out to be an appropriate criterion. Three different types of edges are added to the graph: upward edges, going from a node that is not selected at a particular level to a node selected at that level, downward edges, connecting selected to non-selected nodes and level edges, connecting selected nodes at one level. For finding a shortest path between two nodes, it is sufficient to consider a smaller subgraph of the “multilevel graph”.
VIII. RELATED WORK There is a high volume of ongoing research on shortest paths. The basic speedup techniques and the combined speedup techniques [2], [4], [7] for Dijkstra’s algorithm that usually reduce the number of nodes visited and running time of the algorithm in practice which are discussed here.
D. Shortest-Path Containers This speed-up technique requires a preprocessing computation of all shortest-path trees. For each edge, e ∈ E , a node set, S (e ) , to which the shortest path begins with e is computed[2],[11]. By using a layout, for each edge e ∈ E , the bounding box of S ( e ) is stored in an associative array C with
Velammal College of Engineering and Technology, Madurai
Page 352
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
index set E. The preprocessed information is used for determining if a particular edge will be present on a shortest path to a given target. E. Combination of speed-up Techniques The different speed-up techniques mentioned above can be combined in with each other, which further improve the efficiency during shortest path combination. Some of the combinations were implemented and analyzed[2]. Reach based routing[12] is also combined with many of the basic techniques. IX. OPENMP This section contains parallel programming constructs used to achieve parallelism. OpenMP is an Application Programming Interface (API)[13] for portable shared memory parallel programming. The API allows an incremental approach to parallelize an existing code, so that portions of a program can be parallelized in successive steps. There is a marked difference between OpenMP and other parallel programming paradigms where the entire program must be converted in one step. A. Fork and Join Model of OpenMP Multithreaded programs can be written in different ways that allows complex interaction between threads. OpenMP provides a simple structured approach to multithreaded programming. OpenMP supports a fork-join programming model as shown in Fig. 1.
to assign work to them according to the strategy specified by the programmer. B. Features provided by OpenMP OpenMP provides the following features that can be utilized by a user (i) create teams of threads for parallel execution (ii) specify how work must be shared among the threads (iii) scope for variables (iv) synchronization constructs. Creating a team of threads is achieved using parallel construct. #pragma omp parallel [clause [[,] clause]…] Structured block Although this construct ensures that computations in the associated parallel region are performed in parallel, it does not distribute the work among the threads in a team. There is an implied barrier at the end of the parallel region that forces all threads to wait until the work inside the region has been completed. There are three work sharing constructs available in OpenMP. 1) #pragma omp for Used for sharing iterations in a loop 2) #pragma omp sections Specify different work for each thread individually 3) #pragma omp single Allow only one thread to execute the enclosed region OpenMP provides many constructs for synchronization. Locks are also supported. The shortest path problem is solved in parallel with OpenMP in [6]. Although there is improvement in speedup, there are many speedup techniques which can be parallelized and the performance can be improved. X. SPEEDUP TECHNIQUES The modules in bidirectional search are segregated in such a way that the phases where the parallelism can be applied are identified. Those portions are executed in parallel using OpenMP based on the Multicore architecture.
Fig. 1 Fork-Join programming model of OpenMP
In this approach, the program starts as a single thread of execution. Whenever an OpenMP parallel construct is encountered by a thread while it is executing the program, it creates a team of threads (this is the fork), becomes the master of the team, and collaborates with the other members of the team to execute the code enclosed by the construct. At the end of the construct, only the original thread continues; all others terminate (this is the join). Each portion of code enclosed by a parallel construct is called a parallel region. OpenMP expects the application developer to give a high-level specification of the parallelism in the program and the method for exploiting that parallelism. The OpenMP implementation will take care of the low-level details of actually creating independent threads to execute the code and
Velammal College of Engineering and Technology, Madurai
A. Bidirectional Search The bidirectional search[4] algorithm has a considerable portion of its operations suitable for parallel execution. Initialization (Forward Search) [IFS] Initialization (Backward Search) [IBS] Node Selection (Forward Search) [SFS] Distance updation (Forward Search) [UFS] Node Selection (Backward Search) [SBS] Distance updation (Backward Search) [UBS]
Fig. 2, Bidirectional Search. The inner box with dash-dot line indicates a loop.
Page 353
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The various modules for bidirectional search is outlined in Fig. 2, and detailed in Algorithm 1. In bidirectional search two searches are carried out simultaneously: a forward variant of the algorithm, starting from the source, and a reverse variant of the algorithm, starting from the target. The reverse variant of the algorithm is applied to the reverse graph. The search terminates when a node is marked permanent by both searches. In this scenario, the forward and reverse searches are mutually independent to certain extent, except for the set “perm” which contains permanent nodes marked by both.
search process. Figure 3, describes bidirectional search in which only the updation of neighbor distance values is parallelized. IFS
IBS SFS SBS
//Initialization phase for all nodes u ∈ V set dist ( u ) := ∞
UFS
UBS
initialize priority queue Q with source s and
dist ( s ) := 0 for all nodes u ∈ V set rdist (u ) := ∞ initialize priority queue rQ with target t and
Initialization parallelized using OpenMP Sections
rdist ( t ) := 0 initialize a perm to empty set // Node selection phase begins while priority queue Q and rQ are not empty { get node u with smallest tentative distance dist(u) in Q if( u = t or u ∈ perm ) halt. add u to perm get node ru with smallest distance rdist(ru) in rQ if( ru = s or ru ∈ perm ) halt. add ru to perm // update phase forward search for all neighbor nodes v of u set new-dist := dist(u) + w(u, v) update dist(v) of forward priority queue Q // update phase reverse search for all neighbor nodes rv of ru set new-dist := rdist(ru) + w(ru, rv) update dist(rv) of reverse priority queue rQ } Algorithm 1: Bidirectional Dijkstra’s Algorithm
B. Bidirectional Search with parallelized updates There are two regions where parallelism can be exploited. The initialization of required data structures for both forward and reverse searches can be performed in parallel, that is, data level parallelism can be adopted. The parallel bidirectional algorithm is given in Algorithm 2. The initialization of both forward and reverse search can be parallelized. Another operation that lends itself to parallelization is the updation of distance values, for the neighbors, of a node which is marked permanent. In this region also data parallelism is possible. It is to be noted that another possibility for parallelism is to run both the forward and reverse variants of the algorithm simultaneously as independent threads of the
Velammal College of Engineering and Technology, Madurai
Updates parallelized using OpenMP Sections
Fig. 3, Bidirectional Search with parallelized updates
//Initialization phase # pragma omp section { initialize forward priority queue Q with dist(u)=∞ and dist(s)=0 } # pragma omp section { initialize reverse priority queue rQ with rdist(u)=∞ and rdist(t)=0 } initialize perm array with empty set //Node selection phase begins While (priority queue Q and rQ are not empty) { get node with min dist(u) in Q and assign it to u if (u==t or u € perm) then exit else add u to perm get node with min rdist(u) in rQ and assign it to ru if (ru==s or ru € perm) then exit else add ru to perm //Update phase forward search # pragma omp section { for all neighbor nodes v of u set new-dist := dist(u) + w(u, v) update dist(v) of forward priority queue Q } //Update phase reverse search # pragma omp section { for all neighbor nodes rv of ru set new-dist := rdist(ru) + w(ru,rv) update dist(rv) of reverse priority queue rQ } } Algorithm 2: Bidirectional Search with parallelized updates
XI. EXPERIMENTAL RESULTS
Page 354
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Dijkstra’s algorithm, the bidirectional search and the bidirectional search technique with parallelized updates were implemented and the results were analyzed. The different algorithms mentioned were implemented in C++ with the help of LEDA library. The graph and priority queue data structures as well as other utilities such precise time measurements provided by LEDA were used in the implementation. For parallelizing the regions of code in bidirectional search the OpenMP API was utilized. The use of OpenMP helped to maintain a single source for both the sequential and parallel versions of the algorithm. The code was compiled using GNU C++ Compiler (g++ version 4.2) and the experiments were performed on an Intel Core2Duo machine (2.20 GHz) with 1 GB RAM running on Ubuntu GNU/Linux 8.04 (kernel version 2.6) The main result of this work is shown in Fig. 5. The running time of the different algorithms is plotted (along y-axis) against the number of nodes (along x-axis). The graphs generated were random graphs with the number of edges in O(n), that is, the graphs generated were sparse. The experiments were conducted for 30 runs, and the results of running time of the algorithm are tabulated in Table 1. From Fig. 5, it is observed that the bidirectional search reduces the running time of Dijkstra’s algorithm by a factor of 2.2 while the parallelized bidirectional search reduces the runtime by a factor of 3.1. Run No. 1
Nodes 1000
0.001
0.0015
Parallel Bi+Dij 0.003
5
5000
0.0115
0.006
0.0055
10
10000
0.0295
0.014
0.013
15
15000
0.0495
0.0225
0.0175
20
20000
0.0685
0.036
0.023
25
25000
0.103
0.043
0.0305
30
30000
0.126
0.05
0.0355
1.692
0.7695
0.546
0.0564
0.0257
0.0182
2.1988304 09
3.098901
Total Runtime (30 runs) Average Runtime Speed-Up
Dij
Bi+Dij
TABLE V RUNNING TIME IN DIJKSTRA AND BIDIRECTIONAL DIJKSTRA WITH PARALLELIZED UPDATES
Velammal College of Engineering and Technology, Madurai
Fig. 5. Runtime comparison of the different speed-up techniques Alg 1 – Traditional Dijkstra Alg 2 – Bidirectional Dijkstra Alg 3 – Bidirectional Dijkstra with updates parallelized
RRun No.
Nodes
Dij
Bi+Dij
1
1000
576
55
5
5000
2103
134
10
10000
4190
170
15
15000
7263
236
20
20000
11633
271
25
25000
11932
202
30
30000
14692
275
238203
6020
7940.1
200.6666667
Total Vertex Visit Count (30 runs) Average Vertex Visit Count Speed-Up
39.56860465
TABLE IVI NUMBER OF NODES VISITED IN DIJKSTRA AND BIDIRECTIONAL DIJKSTRA
The experiments were conducted for 30 runs, and the results of Number of nodes visited of the algorithm are tabulated in Table II. Fig 6. plots the number of nodes visited during shortest path computation, using Dijkstra’s algorithm, against the number of nodes available in the graph. Similarly the graph in Fig. 6 takes into account the number of nodes visited during shortest path calculation using bidirectional search in Dijkstra’s algorithm.
Page 355
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
No.of Nodes visited during shortest path computation
[2] Martin Holzer, Frank Schulz, Dorothea Wagner, and Thomas Willhalm, Combining Speed-up Techniques for Shortest-Path Computations, ACM Journal of Experimental Algorithmics, Vol. 10, Article No. 2.5, 2005. [3] Johnson, D. B, Efficient algorithms for shortest paths in sparse networks, Journal of the Association for Computing Machinery,1959, 1–13.
16000
No.of Nodes Visited
14000
[4] Frederikson, G. N, Fast algorithms for shortest paths in planar graphs with applications, SIAM Journal on Computing, 16(6), 1987, 1004–1022.
12000 10000 8000
Dij Bi+Dij
6000
[5] Agarwal. R. and Jagadish, H. V, Algorithms for searching massive graphs, IEEE Transaction son Knowledge and Data Engineering, 6(2),1994, 225–238. [6] , Dominik Schultes, Johannes Singler, Peter Sanders, Parallel Highway Node Routing, A Technical Report, Feb,2008. algo2.iti.kit.edu/schultes/hwy/parallelHNR.pdf
4000 2000 0 1000
5000
10000
15000
20000
25000
30000
No.of Nodes
[7] Willhalm, T, Engineering shortest paths and layout algorithms for large graphs., Ph.D.thesis, Faculty of Informatics, University of Karlsruhe, 2005 [8 Andrew V.Goldberg, Chris Harrelson, Computing the Shortest Path: A* Search Meets Graph Theory, Microsoft Research, 1065 La Avenida, Mountain View, CA 94062. 2005
Fig. 6. Number of nodes visited during shortest path computation Dij – Dijkstra’s Algorithm Bi+Dij – Bidirectional Dijkstra
The result will be the same for Bidirectional Dijkstra with updates parallelized. It is observed that the number of nodes visited during shortest path computation is comparatively less in bidirectional search. The Number of nodes visited in Bidirectional Dijkstra is reduced by a factor of 40. XII. CONCLUSION The speed-up technique for Dijkstra’s algorithm reduces the runtime of the algorithm while retaining optimality. Even though the number of nodes visited during bidirectional search and bidirectional search with updates parallelized are the same. The running time is still reduced in the case of bidirectional search with updates parallelized. Also the speedup techniques can be combined to increase the speed-up considerably. These techniques lend themselves for parallelization which can be effectively utilized in large applications like route planning systems, time table information systems, etc., where multi-core processors are utilized so that the system efficiency can be improved. Locks can be applied during parallelizing the bidirectional Dijkstra for special case of graphs.
[9] Peter Sanders and Dominik Schultes , Highway Hierarchies Hasten Exact Shortest Path Queries, , ESA 2005, LNCS 3669, Springer-Verlag Berlin Heidelberg 2005, 568–579.. [10] Holzer, Hierarchical speedup techniques for shortest path algorithms, M, Tech. report, Dept of Informatics, University of Konstanz, Germany, 2003. [11] Dorothea Wagner and Thomas Willhalm, Geometric Containers for Efficient Shortest-Path Computation, ACM Journal of Experimental Algorithmics, 10(1.3), 2005. [12] Gutman. R.J, Reach-based routing: A new approach to shortest path algorithms optimized for road networks, In Proceedings of the Sixth Workshop on Algorithm Engineering and Experiments and the First Workshop on Analytic Algorithmics and Combinatorics, 2004. [13] OpenMP Website (http://www.openmp.org)
REFERENCES [1] Dijkstra. E. W , A note on two problems in connection with graphs, In Numerische Mathematik. Vol. 1. Mathematisch Centrum, Amsterdam, The Netherlands, 1959, 269–271.
Velammal College of Engineering and Technology, Madurai
Page 356
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hiding Sensitive Frequent Item Set by Database Extension B. Mullaikodi, M.E#1,Dr. S.Sujatha M.E.*2, #
Computer Science and Engineering Department, Anna University, Trichy 1
[email protected]
Abstract -Sharing data among organizations often leads to mutual benefit. Recent technology in data mining has enabled efficient extraction of knowledge from large databases. This, however, increases risks of disclosing the sensitive knowledge when the database is released to other parties. To address this privacy issue, one may sanitize the original database so that the sensitive knowledge is hidden. Sensitive knowledge hiding in large transactional databases is one of the major goals of privacy preserving data mining. A novel approach that performs sensitive frequent item set hiding by applying an extension to the original database was proposed. The extended portion of data set contains transactions that lower the importance of the sensitive patterns, while minimally affecting the importance of the nonsensitive ones. We present the border revision to identify the revised border for the sanitized database and then we compute the minimal size for the extension. The hiding process involves the construction of a Constraints Satisfaction Problem, by using item sets of revised borders and its solution through Binary Integer Programming. Keywords – Privacy preserving data mining, knowledge hiding, association rule mining, binary integer programming.
I. INTRODUCTION A subfield of privacy preserving data mining is “knowledge hiding”. The paper presents a novel approach that strategically performs sensitive frequent itemset hiding based on a new notion of hybrid database generation. This approach broadens the regular process of data sanitization by applying an extension to the original database instead of either modifying existing transactions, or rebuilding the dataset from scratch. The extended portion of the dataset contains a set of carefully crafted transactions that achieve to lower the importance of the sensitive patterns to a degree that they become uninteresting from the perspective of the data mining algorithm, while minimally affecting the importance of the nonsensitive ones. The hiding process is
Velammal College of Engineering and Technology, Madurai
guided by the need to maximize the data utility of the sanitized database by introducing the least possible amount of side-effects, such as (i) the hiding of non-sensitive patterns, or (ii) the production of frequent patterns that were not existent in the initial dataset. The released database, which consists of the initial part (original database) and the extended part (database extension), can guarantee the protection of the sensitive knowledge, when mined at the same or higher support as the one used in the original database. The approach introduced in this paper is exact in nature. On the contrary, when an exact solution is impossible, the algorithm identifies an approximate solution that is close to the optimal one. To accomplish the hiding task, the proposed approach administers the sanitization part by formulating a Constraint Satisfaction Problem (CSP) and by solving it through Binary Integer Programming (BIP). II. KNOWLEDGE HIDING FORMULATION This section provides the necessary background regarding sensitive itemset hiding and sets out the problem at hand, as well as the proposed hiding methodology.
A. Frequent Itemset: Let I = {i1, i2, …………., iM} be a finite set of literals, called items, where M denotes the cardinality of the set. Any subset I ⊆ I is an item set over I. A transaction T over I is a pair T = (tid, I), where I is the item set and tid is a unique identifier, used to distinguish among transactions that correspond to the same item set. A transaction database D = {T1, T2, ....TN} over I is an N x M table consisting of N transactions over I carrying different identifiers, where entry Tnm = 1 if and only if the mth item (m ∈ [1, M]) appears in the nth transaction (n ∈ [1, N]). Otherwise, Tnm = 0. A transaction T = (tid, J) supports an item set I over I, if I ⊆ J. Let S be a set of items; notation p(S) denotes the powerset of S, which is the set of all subsets of S.
Page 357
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Given an item set I over I in D, sup (I, D) denotes the number of transactions T ∈ D that support I and freq(I, D) denotes the fraction of transactions in D that support I. An item set I is called large or frequent in database D if and only if its frequency in D is at least equal to a minimum threshold mfreq. A hybrid of Apriori and FP-Tree algorithms are proposed to be used to find the frequent item set. A hybrid of Apriori and FP-Tree algorithms are proposed to be used to find an optimized set of frequent item set. 1) The Apriori Algorithm - Finding Frequent Itemsets Using Candidate Generation:
Apriori is an influential algorithm for mining frequent itemsets for Boolean association rules. Apriori employs an iterative approach known as a level-wise search, where k-itemsets are used to explore (K+1) itemsets. First , the set of frequent 1-itemsets is found. This set is denoted L1. L1 is used to find L2, the set of frequent 2itemsets , which is used to find L3 , and so on, until no more frequent k-itemsets can be found. The finding of each Lk requires one full scan of the database. To improve the efficiency of the level-wise generation of frequent itemsets, an important property called the Apriori property, i.e., all nonempty subsets of a frequent itemset must also be frequent, is used to reduce the search space. A two step process join and prune actions are used to find Lk from Lk-1.
• If the conditional FP-tree contains a single path, simply enumerate all the patterns
B.Hiding methodology To properly introduce the hiding methodology, one needs to consider the existence of three databases, all depicted in binary format. They are defined as follows: ¾ Database Do, is the original transaction database that, when mined at a certain support threshold msup, and leads to the disclosure of some sensitive knowledge in the form of sensitive frequent patterns. This sensitive knowledge needs to be protected. ¾ Database Dx ,is a minimal extension of Do that is created by the hiding algorithm during the sanitization process, in order to facilitate knowledge hiding. ¾ Database D ,is the union of database Do and the applied extension Dx and corresponds to the sanitized outcome that can be safely released. TABLE 1 Sample : Sanitized Database D as a Mixture of the Original Database DO and the Applied Extension DX
2) FP-Growth Algorithm-Mining Frequent patterns without candidate generation : Frequent pattern growth or simply FP-growth adopts a divide-and – conquer strategy as follows: compress the database representing frequent items into a frequent pattern tree, or FP tree, but retain the itemset association information and then divide such a compressed database into a set of conditional databases, each associate with one frequent item, and mine each such database separately. TABLE 2
Major steps to mine FP-tree
Sample : Frequent Item Sets for DO and DX at msup = 3 (for table 1)
1) Construct conditional pattern base for each node in the FP-tree 2) Construct conditional conditional pattern-base
FP-tree
from
each
3) Recursively mine conditional FP-trees and grow frequent patterns obtained so far
Velammal College of Engineering and Technology, Madurai
Page 358
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Given the sanitized database D, its original version DO, and the produced extension DX , the quality of database D is measured both in the size of DX and in the number of binary variables set to “1” in the transactions of DX (i.e., the distance metric). In both cases, lower values correspond to better solutions.
III. HYBRID SOLUTION METHODOLOGY
Definition 3 (ideal solution). A solution to the hiding of the sensitive item sets is considered as ideal if it has the minimum distance among all the existing exact solutions and is obtained through the minimum expansion of DX. In that sense, ideal is a solution that is both minimal (with respect to distance and size of extension) and exact.
A. Computation of size of database extension
C. Border Revision
Database DO is extended by DX to construct database D. An initial and very important step in the hiding process is the computation of the size of DX. A lower bound on this value can be established based on the sensitive item set in S, which has the highest support. The rationale here is given as follows: by identifying the sensitive item set with the highest support, one can safely decide upon the minimum number of transactions that must not support this item set in DX, so that it becomes infrequent in D.
The rationale behind this process is that hiding of a set of item sets corresponds to a movement of the original borderline in the lattice that separates the frequent item sets from their infrequent counterparts , such that the sensitive item sets lie below the revised borderline. There are four
Lower bound Q is
⎢ sup( I N , DO ) ⎥ Q=⎢ − N⎥ +1 ⎣ mfreq ⎦
………….1
Equation (1) provides the absolute minimum number of transactions that need to be added in DX, to allow for the proper hiding of the sensitive item sets of DO. However, this
lower bound can, under certain circumstances, be insufficient to allow for the identification of an exact solution, even if one exists. To circumvent this problem, one needs to expand the size Q of DX as determined by (1), by a certain number of transactions. A threshold, called safety margin (denoted
hereon as SM), is incorporated for this purpose. Safety margins can be either predefined or be computed dynamically, based on particular properties of database DO and / or other parameters regarding the hiding process.
B. Exact and Ideal Solutions Definition 1 (feasible/exact/approximate solution). A solution to the hiding of the sensitive knowledge in DO is considered as feasible if it achieves to hide the sensitive patterns. Any feasible solution, introducing no side effects in the hiding process, is called exact. Finally, any non exact feasible solution is called approximate.
possible scenarios involving the status of each item set I prior and after the application of border revision:
C1 : Item set I was frequent in DO and remains frequent in D. C2 : Item set I was infrequent in DO and is infrequent in D. C3 : Item set I was frequent in DO and became infrequent in D. C4 : Item set I was infrequent in DO and became frequent in D. Since the borders are revised to accommodate for an exact solution, the revised hyper plane is designed to be ideal in the sense that it excludes only the sensitive item sets and their supersets from the set of frequent patterns in D, leaving the rest of the item sets in their previous status as in database DO. The first step in the hiding methodology rests on the identification of the revised borders for D. The hiding algorithm relies on both the revised positive and the negative borders, denoted as Bd+ (F1D) and Bd– (F1D), respectively. After identifying the new (ideal) borders, the hiding process has to perform all the required minimal adjustments of the transactions in Dx to enforce the existence of the new borderline in the result database.
Definition 2 (database quality).
Velammal College of Engineering and Technology, Madurai
Page 359
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
more, by carefully selecting the item sets of set C, the hiding algorithm ensures that the exact same solution to the one of solving the entire system of inequalities is attained. This is accomplished by exploiting cover relations existing among the item sets in the lattice. Set C is chosen appropriately to consist of all the item sets of the revised border. The proposed hiding algorithm is capable of ensuring that if (3) and (4) are satisfied for all the item sets in C, then the produced solution is exact and is identical to the solution involving the whole system of the 2M – 1 inequalities. Cover relations governing the various item sets in the lattice of DO ensure that the formulated set of item sets C has an identical solution to the one of solving the system of all 2M – 1 inequalities for D. Fig. 1 An sample item set lattic demonstration (a) the original border and the sensitive item sets, and (b) the revised border for Table1
D. Problem Size Reduction To enforce the computed revised border and identify the exact hiding solution, a mechanism is needed to regulate the status (frequent versus infrequent) of all the item sets in D. Let C be the minimal set of border item sets used to regulate the values of the various uqm variables in DX. Moreover, suppose that I ∈ C is an item set, whose behavior we want to regulate in D. Then, item set I will be frequent in D if and only if sup (I, Do) + Sup(I, Dx) ≥ m freq x (N+Q), or equivalently if Q
Sup (I,DO) +
∑∏
uqm
≥
mfreq x (N + Q) …(3)
q =1 i M ∈I
equivalently if Q
sup (I,DO) +
∑∏
uqm < mfreq x (N + Q) ...(4)
q =1 i M ∈I
Inequality (3) corresponds to the minimum number of times that an item set I has to appear in the extension DX to remain frequent in D. On the other hand, (4) provides the maximum number of times that an item set I has to appear in DX to be infrequent in database D. To identify an exact solution to the hiding problem, every possible item set in P, according to its position in the lattice—with respect to the revised border—must satisfy either (3) or (4). However, the complexity of solving the entire system of the 2M _ 1 inequalities is well known to be NP-hard . Therefore, one should restrict the problem to capture only a small subset of these inequalities, thus leading to a problem size that is computationally manageable. The proposed problem formulation achieves this by reducing the number of the participating inequalities that need to be satisfied. Even
Velammal College of Engineering and Technology, Madurai
The cover relations that exist between the item sets of Bd+ (F`D) and those of F`D. In the same manner, the item sets of Bd– (FD) are generalized covers for all the item sets of P \ (Bd+ (F`D) ∪ {∅}). Therefore, the item sets of the positive and the negative borders cover all item sets in P. Optimal solution set C: The exact hiding solution, which is identical to the solution of the entire system of the 2M – 1 inequalities, can be attained based on the item sets of set C = Bd+ (F`D) ∪ Bd– (F`D) Based on (7), the item sets of the revised borders Bd+ (F`D) and Bd– (F`D) can be used to produce the corresponding inequalities, which will allow for an exact hiding solution for DO.
D. Handling of Suboptimality Since an exact solution may not always be feasible, the hiding algorithm should be capable of identifying good approximate solutions. There are two possible scenarios that may lead to nonexistence of an exact solution. Under the first scenario, DO itself does not allow for an optimal solution due to the various supports of the participating item sets. Under the second scenario, database DO is capable of providing an exact solution, but the size of the database extension is insufficient to satisfy all the required inequalities of this solution. To tackle the first case, the hiding algorithm assigns different degrees of importance to different inequalities. To be more precise, while it is crucial to ensure that (4) holds for all sensitive item sets in D, thus they are properly protected from disclosure, satisfaction of (3) for an item set rests in the discretion of ensuring the minimal possible impact of the sanitization process to DO. This inherent difference in the significance of the two inequalities, along with the fact that solving the system of all inequalities of the form (4) always leads to a feasible solution (i.e., for any database DO), allows the relaxation of
Page 360
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the problem, when needed, and the identification of a good approximate solution. To overcome the second issue, the hiding algorithm incorporates the use of a safety margin threshold, which produces a further expansion of DX by a certain number of transactions. These transactions must be added to the ones computed by using (1). The introduction of a safety margin can be justified as follows: Since (1) provides the lower bound on the size of database DX , it is possible that the artificially created transactions are too few to accommodate for the proper hiding of knowledge. This situation may occur due to conflicting constraints imposed by the various item sets regarding their status in D. These constraints require more transactions (or to be more precise, more item modifications) in order to be met. Thus, a proper safety margin will allow the algorithm to identify an exact solution if such a solution exists. The additional extension of DX, due to the incorporation of the safety margin, can be restricted to the necessary degree. A portion of transactions in DX is selected and removed at a later point, thus reducing its size and allowing an exact solution. Therefore, the only side effect of the use of the safety margin in the hiding process is an inflation in the number of constraints and associated binary variables in the problem formulation, leading to a minuscule overhead in the runtime of the hiding algorithm.
E. Formulation and Solution of the CSP A CSP is defined by a set of variables and a set of constraints, where each variable has a nonempty domain of potential values. The constraints involve a subset of the variables and specify the allowable combinations of values that these variables can attain. Since in this work all variables involved are binary in nature, the produced CSP is solved by using a technique called BIP that transforms it to an optimization problem. To avoid the high degree of constraints, the application of a Constraints Degree Reduction(CDR) approach is essential. On the other hand, the resulting inequalities are simple in nature and allow for fast solutions, thus adhere for an efficient solution of the entire CSP. The proposed CSP formulation is
F. Validity of Transactions The incorporation of the safety margin threshold in the hiding process may lead to an unnecessary extension of DX.
Velammal College of Engineering and Technology, Madurai
it is possible to identify and remove the extra portion of DX that is not needed, thus minimize the size of database D to the necessary limit. To achieve that, one needs to rely on the notion of null transactions, appearing in database DX. A transaction Tq is defined as null or empty if it does not support any valid item set in the lattice. Null transactions do not support any pattern from P \ {φ}. Apart from the lack of providing any useful information, null transactions are easily identifiable, thus produce a privacy breach in the hiding methodology. They may exist due to two reasons: 1) an unnecessarily large safety margin or 2) a large value of Q essential for proper hiding. In the first case, these transactions need to be removed from DX, while in the second case the null transactions need to be validated, since Q denotes the lower bound in the number of transactions to ensure proper hiding. After solving the CSP, all the null transactions appearing in DX are identified. Suppose that Qinv such transactions exist. The size of database DX will then equal the value of Q plus the safety margin SM. This means that the valid transactions in DX will be equal to υ = Q + SM – Qinv. To ensure minimum size of DX , the hiding algorithm keeps only k null transactions, such that k = max (Q - υ, 0) ⇒ k = max (Qinv – SM, 0). As a second step, the hiding algorithm needs to ensure that the k empty transactions that remain in DX become valid prior to releasing database D to public. A heuristic is applied for this purpose, which effectively replaces null transactions of DX with transactions supporting item sets of the revised positive border. After solving the CSP in Fig. 2, the outcome is examined to identify null transactions. Then, the null transactions are replaced with valid ones, supporting item sets of Bd+ (F`D).
IV. EXPERIMENTAL EVALUATION
The algorithm was tested on different datasets using different parameters such as minimum support threshold and number/size of sensitive itemsets to hide. The thresholds of minimum support were properly set to ensure an adequate amount of frequent itemsets, among which a set of sensitive itemsets were randomly selected. We compare the solutions of the hybrid algorithm against three state-of-the-art BBAs: the BBA, the Max-Min 2 algorithm, and inline algorithm at terms of side effects introduced by the hiding process. The hybrid algorithm consistently outperforms the three other schemes, with the inline approach being the second best. An interesting insight from the conducted experiments is the fact that the hybrid approach, when compared to the inline algorithm and the heuristic approaches can better preserve the quality of the
Page 361
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
border and produce superior solutions. Indeed, the hybrid approach introduces the least amount of side effects among the four tested algorithms.
V. CONCLUSION
A novel, exact border-based hybrid approach to sensitive frequent item set hiding, through the introduction of a minimal extension to the original database was presented. A hybrid approach of combining Apriori and FP_Tree was done to find optimized frequent item set. This methodology is capable of identifying an ideal solution whenever one exists, or approximate the exact solution and the solution of this approach is of higher quality than previous approaches.
REFERENCES [1] Aris Gkoulalas – Divanis, Vassilios S. Verykois, (May 2009) “Exact Knowledge Hiding through Database Extension” IEEE Transaction on Knowledge and Data Engineering, Vol. 21, No. 5, [2] Gkoulalas-Divanis. A and Verykios. V.S, (Nov. 2006) “An Integer Programming Approach for Frequent Itemset Hiding,” Proc. ACM Conf. Information and Knowledge Management (CIKM ’06), pp. 748-757, [3] Oliveira. S.R.M. and Zaı¨ane. O.R.,, (2003) “Protecting Sensitive Knowledge by Data Sanitization,” Proc. Third IEEE Int’l Conf. Data Mining (ICDM ’03), pp. 211-218. [4] Sun. X and Yu. P.S, (2005) “A Border-Based Approach for Hiding Sensitive Frequent Itemsets,” Proc. Fifth IEEE Int’l Conf. Data Mining (ICDM ’05), pp. 426-433. [5] Verykios. V.S., Emagarmid. A.K., Bertino. E., Saygin. Y, and Dasseni. E, (Apr. 2004) “Association Rule Hiding,” IEEE Trans. Knowledge and Data Eng., vol. 16, no. 4, pp. 434-447,
Velammal College of Engineering and Technology, Madurai
Page 362
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Denial of service: New metrics and Their measurement
Dr.KannanBalasubramanian1, Kavithapandian.P2 Professor, Department of Computer Science & Engineering Mepco Schlenk Engineering College, Sivakasi, Tamilnadu, India.1 II M.E [CSE], Anna University, Thiruchirapall, Tamilnadu, Indiai.2 [email protected], [email protected]
Abstract - In this paper, we propose a novel, user-centric approach to DoS impact measurement. Our key insight is that DoS always causes degradation of service quality, and a metric that holistically captures a human user’s QoS perception will be applicable to all test scenarios. For each popular application, we specify its QoS requirements, consisting of relevant traffic measurements and corresponding thresholds that define good service ranges. We observe traffic as a collection of high-level tasks, called “transactions”. Each legitimate transaction is evaluated against its application’s QoS requirements; transactions that do not meet all the requirements are considered “failed.” We aggregate information about transaction failure into several intuitive qualitative and quantitative composite metrics to expose the precise interaction of the DoS attack with the legitimate traffic. KeywordsNetwork-level security and protection, communication/networking and information technology, computer systems organization, measurement techniques, performance of systems.
I. INTRODUCTION Denial of service (DoS) is a major threat. DoS severely disrupts legitimate communication by exhausting some critical limited resource via packet floods or by sending malformed packets that cause network elements to crash. The large number of devices, applications, and resources involved in communication offers a wide variety of mechanisms to deny service. Effects of DoS attacks are experienced by users as a severe slowdown, service quality degradation, or service disruption. DoS attacks have been studied through network simulation or testbed experiments. Accurately measuring the impairment of service quality perceived by human clients during an attack is essential for evaluation and comparison of potential DoS defenses, and for study of novel attacks. Researchers and developers need accurate, quantitative, and versatile DoS impact metrics whose use does not require significant changes in current simulators and experimental tools. Accurate metrics produce measures of service denial existing metrics agree with human perception of service denial. .
We demonstrate that our metrics meet the goals of being accurate, quantitative, and versatile 1) through testbed experiments with multiple DoS scenarios and rich legitimate traffic mixes, 2) through OPNET simulations and 3) through experiments involving human users. This project’s contributions are threefold: 1) We propose a novel approach to DoS impact measurement relying on application-specific QoS requirements. Although our proposed metrics combine several existing approaches, their novelty lies in a) the careful specification of traffic measurements that reflect service denial for the most popular applications and b) the definition of QoS thresholds for each measurement and each application class, based on extensive study of the QoS literature. 2) We aggregate multiple measurements into intuitive and informative DoS metrics that can be directly applied to existing testbed experiments and simulations, and to a variety of DoS scenarios. 3) We demonstrate that our metrics accurately capture human perception of service denial by conducting experiments with human users. II. WORK FLOW To build a network model the workflow centers on the Project Editor. This is used to create network models, collect statistics directly from each network object or from the network as a hole, execute a simulation and view results.
A. Data Flow Diagram Interactive Application
QOS Measurement
Introduce DOS
Calculate DOS Metrics
Analysis Traffic
UDP Bandwidth Flood
Analysis
Velammal College of Engineering and Technology, Madurai
Graph Representation
Page 363
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
III. Existing metrics Prior DoS research has focused on measuring DoS through selected legitimate traffic parameters: 1. Packet loss, 2. Traffic throughput or goodput, 3. Request/Response delay, 4. Transaction duration, and 5. Allocation of resources. A. Packet Loss Loss is defined as the number of packets or bytes lost due to the interaction of the legitimate traffic with the attack or due to collateral damage from a defense’s operation. The loss metric primarily measures the presence and extent of congestion in the network due to flooding attacks. It cannot be used for attacks that do not continually create congestion, or do not congest network resources at all. Examples of such attacks are pulsing attacks, TCP SYN floods attacks that target application resources, and vulnerability attacks that crash applications and hosts. Further, the loss metric typically does not distinguish between the types of packets lost, while some packet losses have a more profound impact than others (for example, a lost SYN versus data packet) on service quality. B. Throughput Throughput is defined as the number of bytes transferred per unit time from the source to the destination. Goodput is similar, but does not count retransmitted bytes. Both are meaningful for TCP-based traffic, which responds to congestion by lowering its sending rate.
C. Request/Response delay Request/response delay is defined as the interval between the time when a request is issued and the time when a complete response is received from the destination. It measures service denial of interactive applications (e.g., telnet) well, but fails to measure it for non-interactive applications (e.g., email), which have much larger thresholds for acceptable request/response delay. Further, it is completely inapplicable to one-way traffic (e.g., media traffic), which does not generate responses but is sensitive to one-way delay, loss and jitter. D. Transaction duration It is the time needed for an exchange of a meaningful set of messages between a source and a destination. This metric
Velammal College of Engineering and Technology, Madurai
depends heavily on the volume of data being transferred and whether the application is interactive and congestion sensitive. It accurately measures service denial for interactive applications such as Web browsing. For oneway traffic, such as media streaming that may not respond to congestion and runs over UDP,transaction duration will not be affected by the attack. Duration of many noninteractive transactions can be extended without causing service denial because humans expect that such transactions may occur with some delay. E. Allocation of resources Allocation of resources is defined as the fraction of a critical resource (usually bandwidth) allocated to legitimate traffic vs. attack traffic. This metric does not provide any insight into the user-perceived service quality. Instead it assumes that the only damage to legitimate traffic is inflicted by the lack of resources, and is only applicable to flooding attacks.
IV. PROPOSED METRICS We now introduce several definitions needed for DoS impact metrics. The client is the host that initiates communication with another party, which we call the server.
Definition 1. A conversation between a client and a server is the set of all packets exchanged between these two hosts to provide a specific service to the client, at a given time. A conversation is explicitly initiated by a client application (e.g., by opening a TCP connection or sending a UDP packet to a well-known service) and ends either explicitly (a TCP connection is closed, UDP service is rendered to the client) or after a long period of inactivity. If several channels are needed to render service to a client, such as FTP control and data channels, all related channels are part of a single conversation. Definition 2. A transaction is the part of a conversation that represents a higher-level task whose completion is perceptible and meaningful to a user. Transaction usually involves a single request-reply exchange between a client and a server, or several such exchanges that occur close in time. A conversation may contain one or several transactions. Definition 3. A transaction is successful if it meets all the QoS requirements of its corresponding application. If at least one QoS requirement is not met, a transaction has failed. Transaction success/failure is the core of our proposed metrics.
Page 364
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Interactive applications such as Web, file transfer, telnet,email (between a user and a server), DNS, and ping involve a human user requesting a service from a remote server, and waiting for a response. Their primary QoS requirement is that a response is served within a user-acceptable delay. Research on human perception of Web traffic delay shows that people can tolerate higher latencies for entire task completion if some data is served incrementally. Media applications such as conversational and streaming audio and video have strict requirements for low loss,low jitter, and low one-way delay. These applications further involve a media channel (where the audio and video traffic are sent, usually via UDP) and a control channel (for media control). Both of these channels must provide satisfactory service to the user. We treat control traffic as interactive traffic requiring a 4-second partial delay. Chat applications can be used for text and media transfer between two human users. While the request/response delays depend on human conversation dynamics,the receipt of user messages by the server must be acknowledged within a certain time. This delay requirement as a 4-second threshold on the round-trip time between the client and the server. Additionally, we apply the QoS requirements for media applications to the media channel of the chat application. DoS Metrics We aggregate the transaction success/failure measures into several intuitive composite metrics. Percentage of failed transactions (pft) This metric directly captures the impact of a DoS attack on network services by quantifying the QoS experienced by users. For each transaction that overlaps with the attack, we evaluate transaction success or failure applying Definition 3. A straightforward approach to the pft calculation is dividing the number of failed transactions by the number of all transactions during the attack. This produces biased results for clients that generate transactions serially.
because in some experiments it may be useful to produce a single number that describes the DoS impact. But we caution that DoS-level is highly dependent on the chosen application weights and thus can be biased.
QoS-ratio It is the ratio of the difference between a transaction’s traffic measurement and its corresponding threshold, divided by this threshold. The QoS metric for each successful transaction shows the user-perceived service quality, in the range (0, 1], where higher numbers indicate better quality. It is useful to evaluate service quality degradation during attacks. We compute it by averaging QoS-ratios for all traffic measurements of a given transaction that have defined thresholds.
The failure ratio It shows the percentage of live transactions in the current (1-second) interval that will fail in the future. The failure ratio is useful for evaluation of DoS defenses, to capture the speed of a defense’s response, and for time-varying attacks. Transactions that are born during the attack are considered live until they complete successfully or fail. Let A be a client that initiates some conversation with server B. A request is defined as all data packets sent from A to B, before any data packet is received from B. A reply is defined as all data packets sent from B to A, before any new request from A.
The DoS-hist metric It shows the histogram of pft measures across applications, and is helpful to understand each application’s resilience to the attack.
The DoS-level metric It is the weighted average of pft measures for all applications of interest: DoS-level =∑kpft(k).wk, where k spans all application categories, and wk is a weight associated with a category k. We introduced this metric
Velammal College of Engineering and Technology, Madurai
Figure illustrates request and reply identification, and measurement of partial delay, echo delay, and whole delay. We specify two types of delay requirements for e-mail, Web, telnet, and file transfer transactions where a user can utilize a partial response: 1) Partial delay measured between receipts of any two data packets from the server. For the first data packet, partial delay is measured from the end of a user request and 2) Whole delay measured from the end of a user request until the entire response has been received. Additionally, a
Page 365
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
telnet application serves two types of responses to a user: it echoes characters that a user types, and then generates a response to the user request. The echo generation must be faster than the rest of the response, so we define the echo delay requirement for telnet transactions. We identify the echo delay as the delay between a user’s request and the first response packet. V. EXPERIMENTAL SETUP
Topology Four legitimate networks and two attack networks are connected via four core routers. Each legitimate network has four server nodes and two client nodes, and is connected to the core via an access router. Links between the access router and the core have 100Mbps bandwidth and 10-40-ms delay, while other links have 1-Gbps bandwidth and no added delay. The location of bottlenecks is chosen to mimic high-bandwidth local networks that connect over a limited access link to an over provisioned core. Attack networks host two attackers each, and connect directly to core routers.
UDP flood Attack This attack can deny service in two ways: 1) By generating a large traffic volume that exhausts bandwidth on bottleneck links (more frequent variant) and 2) By generating a high packet rate that exhausts the CPU at a router leading to the target. We generate the first attack type: a UDP bandwidth flood. Packet sizes had range [750 bytes, 1.25 Kbytes] and total packet rate was 200 Kbps.
TCP SYN flood Attack Another popular attack with both attackers and researchers is the TCP SYN flood. It denies service by sending many TCP SYN packets that consume OS memory at the target. This attack can be largely countered if the target deploys the TCP SYN-cookie defense, which allocates memory only after the three-way handshake is completed.
VI. SYSTEM IMPLEMENTATION
1) Programming in OPNET
Figure Experimental topology shows the traffic patterns. Traffic patterns for IRC and VoIP differ because those application clients could not support multiple simultaneous connections. All attacks target the Web server in network 4 and cross its bottleneck link, so only this network’s traffic should be impacted by the attacks
Background Traffic Each client generates a mixture of Web, DNS, FTP, IRC, VoIP, ping, and telnet traffic. We used open-source servers and clients when possible to generate realistic traffic at the application, transport, and network level. For example, we used an Apache server and wget client for Web traffic, bind server and dig client for DNS traffic, etc. Telnet, IRC, and VoIP clients and the VoIP server were custom-built in Perl. Clients talk with servers in their own and adjacent networks.
Velammal College of Engineering and Technology, Madurai
OPNET is a simulator built on top of a discrete event system. It simulates the system behavior by modeling each event happening in the system and processes it by user-defined processes. It uses a hierarchical strategy to organize all the models to build a whole network. The hierarchy models entities from physical link transceivers, antennas, to CPU running processes to manage queues or running protocols, to devices modeled by nodes with process modules and transceivers, to network model that connects all different kinds of nodes together. OPNET also provides programming tools for us to define any type of packet format we want to use in our own protocols. Programming in OPNET includes the following major tasks: define protocol packet format, define the state transition machine for processes running the protocol, define process modules and transceiver modules we need in each device node, finally define the network model by connecting the device nodes together using user-defined link models.
Page 366
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Sample Node and Link configuration without attacks Failure ratio for transaction
Sample Node and Link configuration with attacks
Life diagram of failed transaction
Transaction cdf with respect to loss
QoS degrade measure for failed transaction
VII.CONCLUSION Our metrics are usable and they offer the real opportunity to compare and contrast different DoS attacks and defenses on an objective head-to-head basis. This work will advance DoS research by providing a clear measure of success for any proposed defense, and helping researchers gain insight into strengths and weaknesses of their solutions.
REFERENCES [1] B.N. Chun and D.E. Culler, “User-Centric Performance Analysis of Market-Based Cluster Batch Schedulers,” Proc. Second IEEE Int’l Symp. Cluster Computing and the GridProc. Second IEEE/ACM Int’l Conf. Cluster Computing and the Grid (CCGRID ’02), May 2002.
Velammal College of Engineering and Technology, Madurai
Page 367
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[2] CERT Advisory CA-1996-21 TCP SYN Flooding and IP Spoofing Attacks, CERT CC, http://www.cert.org/advisories/CA-1996- 21.html, 1996. [3] M. Guirguis, A. Bestavros, and I. Matta, “Exploiting the Transients of Adaptation for RoQ Attacks on Internet Resources,” Proc. 12th IEEE Int’l Conf. Network Protocols (ICNP ’04), Oct. 2004. [4] A. Kuzmanovic and E.W. Knightly, “Low-Rate TCP-Targeted Denial of Service Attacks (The Shrew versus the Mice and Elephants),” Proc. ACM SIGCOMM ’03, Aug. 2003. [5] S. Kandula, D. Katabi, M. Jacob, and A. Berger, “Botz-4-Sale: Surviving Organized DDoS Attacks that Mimic Flash Crowds,” Proc. Second Symp. Networked Systems Design and Implementation (NSDI), 2005.
Velammal College of Engineering and Technology, Madurai
Page 368
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
High Performance Evaluation of 600-1200V, 140A Silicon Carbide Schottky Barrier Diodes and Their Applications Using Mat Lab Manickavasagan K, P.G. Student, 401, Adhipraskthi Engineering College. Melmaruvathur, Tamil Nadu-603319. [email protected] Abstract - High performance evaluation of a 600-1200 V / 140 A range Sic schottky Barrier Diode and their applications is experimentally evaluated. The SiC Schottky Barrier Diode (SBD) is commercially available in the 600-1200 V / 1-40 A range. The main advantage of a high voltage SiC SBD lies in its superior dynamic performance in this respect: (a) The reverse recovery charge in the SiC SBD is extremely low (< 20 nC) and is the result of junction capacitance, not stored charge. Furthermore, unlike the Si PiN diode, it is independent of di/dt, forward current and temperature. (b) Higher junction temperature operation up to 175°C, (c) Reduction in the number of MOSFETs by 50%, (d) Faster switching up to 500 kHz to reduce the EMI Filter size and other passives, and (e) Reduction or elimination of the active or passive snubber. The performance of a 600 V, 4 A silicon carbide (SiC) Schottky diode is experimentally evaluated. A 300 W boost power factor corrector with average current mode control (PFC) is considered as a key application. Measurements of overall efficiency, switch and diode losses and conducted electromagnetic interference (EMI) are performed both with the SiC diode and with two ultra-fast, soft-recoveries, silicon power diodes, and the recently presented. The paper compares the results to quantify the impact of the recovery current reduction provided by SiC diode on these key aspects of the converter behavior. Keywords: (SBD) SiC Schottky Barrier Diode, (SiC) silicon carbide, Galium Arsenide (GaAs)
1.
Introduction
I should bring out my taughts and ideas through this paper in order to convey the topic. High performance evaluation of a 600-1200 V / 1-40 A range Sic schottky Barrier Diode and their applications is experimentally evaluated. Such that increase of power density is one of the main tasks for power electronics today: system sizes shall be reduced although in general the power output of the user applications increases. There are two ways to meet this challenge: 1. Reduction of losses by more efficient power electronic devices; 2. reduction of active and passive components’ number, weight and size, generally by increasing switching frequencies.
Velammal College of Engineering and Technology, Madurai
Unfortunately, Silicon (Si) bipolar diodes always show significant switching losses due to their reverse recovery behavior, especially at elevated temperatures as occurs in operation. Unipolar (“pure Schottky”) diodes on Si can only be made for voltages up to about 100V. To overcome this silicon limit, high band gap semiconductors have come into focus during the last few years. Galium Arsenide (GaAs) and Silicon Carbide (SiC) Schottky diodes have been made with breakdown voltages up to 600V and even several thousands of Volts respectively without (or more precisely with extremely small) reverse recovery. These devices are available for several years now, and their advantages have been shown in numerous applications. The power electronic systems operating in the 600-1200 voltage range currently utilize silicon PiN diodes, which tend to store large amounts of minority carrier charge in the forward-biased state. The stored charge has to be removed by majority carrier recombination before the diode can be turned off. This causes long storage and turn-off times. The prime benefits of the SiC SBD lie in its ability to switch fast (< 50 ns), with almost zero reverse recovery electron mobility than 6H-SiC. charge and the high junction temperature operation. The 600 V GaAs SBDs can be made but suffer from limitations with regards to the high junction temperature operation and 5x bigger foot-print for the same current rating. The comparable Silicon PiN diodes (Si SBDs are not viable in the 600 V range because of their large on-state voltage drops) have a reverse recovery charge of 100-500 nC and take at least 100 ns to turn-off. This places a tremendous burden on other switching elements in the system in terms of the required forward safe operating area and the switching losses incurred.
2. Electronic Properties Although there are about 170 known crystal structures, or polytypics of SiC, only two (4H-SiC and 6H-SiC) are available commercially. 4H-SiC is preferred over 6HSiC for most electronics applications because it has higher and more isotropic electron mobility than 6H-SiC. Table 1 compares the key electronic
Page 369
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
properties of 4H-SiC to Si and GaAs. The higher breakdown electric field strength of SiC enables the potential use of SiC SBDs in 600-2000 V range. Specific benefits of SiC electronic properties are: The 10x higher breakdown electric field strength of SiC reduces the specific on resistance compared to the Si and GaAs SBDs. This is illustrated in Fig. 1. At 600 V, a SiC SBD offers a Ron of 1.4 mΩ-cm2, which is considerably, less than 6.5 mΩ-cm2 for a GaAs SBD and 73 mΩ-cm2 for a Si SBD. This means that the SiC SBD will have a much smaller foot-print. The higher band gap results in much higher schottky metal-semiconductor barrier height as compared to GaAs and Si, resulting in extremely low leakage currents at elevated junction temperatures due to reduced thermionic electron emission over the barrier. The very high thermal conductivity of SiC reduces the thermal resistance of the die.
3.
SiC Schottky Diodes
Fig. 2 shows a typical temperature dependent forward characteristic of a 10 A / 600 V 4H-SiC SBD. The onresistance increases with temperature due to the reduction in the electron mobility at elevated temperatures. The diode shows 10 A at a VF of 1.5 V at 25˚C. The current reduces to approximately 5.7 A at the same VF at 200˚C. This negative temperature coefficient of forward current allows us to parallel more than one die in a package without any unequal current sharing issues. This behavior is unlike high voltage Si PiN diodes. Fig. 3 shows the reverse characteristics of the 10 A / 600 V SBD. The typical leakage current is less than 50 µA at 600 V at 25 C which increases to 70 µA at 200˚C – a very nominal increase for such a wide temperature range. The devices were packaged in plastic TO-220 packages with a thermal impedance of 1.1˚C/W. The current de-rating curve for a packaged part is shown in Fig. 4. These parts are rated for a maximum junction temperature of 175˚C. For a case temperature of up to 150˚C, the junction temperature remains below 175˚C. When the case temperature is above 150˚C, the current has to be appropriately de-rated to keep the junction temperature below 175˚C. The reverse capacitance vs. voltage curve is shown in Fig. 5. At 10 V reverse bias, the input capacitance is about 240 pF, which drops to 90 pF at 100 V and saturates to 50 pF above 300 V. This capacitance is comparable to low voltage Si Schottky Diodes. The turn-off characteristics of the 10 A/600 V 4H-SiC SBD are compared with a Si FRED at different temperatures (Fig. 6). The SiC diode, being a majority
Velammal College of Engineering and Technology, Madurai
carrier device, does not have any stored minority carriers. Therefore, there is no reverse recovery current associated with the turn-off transient of the SBD. However, there is a small amount of displacement current required to charge the Schottky junction capacitance (< 2 A) which is independent of the temperature, current level and di/dt. In contrast to the SiC SBD, the Si FRED exhibits a large amount of the reverse recovery charge, which increases dramatically with temperature, on-current and reverse di/dt. For example, the Qrr of the Si FRED is approximately 160 nC at room temperature and increases to about 450 nC at 150˚C. This excessive amount of Qrr increases the switching losses and places a tremendous burden on the switch and diode in typical PFC or motor control applications. In a switching application, the diode will be subjected to peak currents that are greater than the average rated current of the device. Fig. 7 shows a repetitive peak forward surge current of 50 A at 25˚C for the 10A / 600 V SiC SBD. This 60 Hz half sine wave measurement indicates a repetitive peak current of 5X the average. The quality of the SiC wafers has been continuously improving over the last 5 years. It is now possible to make larger area chips. Fig. 8 shows an example of a single SiC SBD chip rated at 600 V / 30 A. This device showed a leakage current of 70 µA at a reverse bias of 600 V. As was mentioned before, the negative temperature coefficient of the current makes it possible to easily parallel several chips in a single package without encountering current sharing problems. An example is shown in Fig. 9, where three chips of the type shown in Fig. 8 were packaged together to provide an 80 A / 600 V part. This package had a leakage current of 125 µA at 600 V and 25˚C. This demonstrates that the SiC SBD technology is scalable to higher currents. While the material is continuously improving, relatively high current parts can be obtained now by paralleling several chips in a package.
4. Reliability of the SiC SBD SiC is inherently a very robust and reliable material. Our 600 V SBDs have undergone extensive reliability testing and can be used as an example of SiC device reliability. To date the diodes have completed a total of 145,000 device hours of High Temperature Reverse Bias Testing (HTRB), 11,000 device hours of continuous current “burn in” testing, and 35,000 device hours of power cycle testing with no failures. The HTRB testing involved seven separate lots with test conditions of -600 volts DC at a temperature of 200˚C. The “burn in” was done at the rated device forward current, with the device junction temperature held at 200˚C. The power cycle consisted of 7
Page 370
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
minute on/off cycles (3.5 min. on / 3.5 min off) with the on-current set to the device rated current, a maximum junction temperature of 175˚C and a junction temperature delta of greater than 100˚C during the cycle.
5. Power Factor Correction (PFC) One of the largest applications for SiC Schottky rectifiers in the near future is in the continuous conduction mode (CCM) power factor correction (PFC) circuit. In traditional off-line AC-DC power supplies used in computer and telecom applications, the AC input sees a large inductive (transformer) load which causes the power factor to be substantially lower than 1. A PFC circuit allows the AC input line to see near-unity power factor, as required by new legal requirements. As shown in Fig. 10, chopping the full wave rectified input with a fast switch (MOSFET), and then stabilizing the resulting DC waveform using a capacitor accomplishes this function. When the MOSFET is ON, it is necessary to prevent the current to flow from the output capacitor or the load through the MOSFET. Hence, when the FET is ON, the Diode is OFF, and vice versa. During the switching transient when the Diode is turning OFF and the MOSFET is turning ON, the reverse recovery current from the Diode flows into the MOSFET, in addition to the rectified input current. This results in a large inrush current into the MOSFET, requiring a substantially large sized MOSFET, than that required if the Diode had no reverse recovery current. This large MOSFET represents a substantial cost in this circuit. These switching losses also limit the frequency of operation in the circuit, and hence its cost, size, weight and volume. A higher frequency would allow the size of the passive components to be correspondingly smaller. Many fast silicon rectifiers also show “snappy” reverse recovery, which results in a large EMI signature, which are also unacceptable to the new European requirements. A fast rectifier with smooth switching characteristics will allow for high efficiency PFC circuits, which also comply with new legal requirements. A 4H-SiC diode is such a rectifier. This near-zero reverse recovery SiC Schottky rectifier offers low switching losses while still showing comparable on-state performance of conventional silicon rectifiers. Due to the majority carrier transport properties of these rectifiers, they show only a capacitive current during their turn-off transient, which flows through the power MOSFET. In order to measure the benefit of these high performance rectifiers, a 250 Watt PFC test circuit was compared with an ultrafast Silicon Diode as well as SiC SBD. This test circuit used a 14 A, 500 V International Rectifier MOSFET (IRFP450), and a 6 A, 600 V ultrafast IR Si PiN diode (HFA08TB60). The input voltage was kept at a constant
Velammal College of Engineering and Technology, Madurai
120 V RMS, and the output voltage was 370 V DC. The operating frequency was 90 kHz, and the gate resistance at the MOSFET was 50 Ω. The current rating of the MOSFET was higher than the average rating to accommodate the reverse recovery current of the diode, and to maintain a high efficiency of the circuit. Under full load condition, a 600 Ω resistor was utilized, while at half load condition, 1200 Ω was used. Voltage and current measurements were taken on both the MOSFET as well as the diode, in order to estimate the power losses in these components. The input and output power was also measured to calculate the efficiency of the circuit. Under full load conditions, the temperature on the MOSFET case was measured with and without an external fan on the device. After all these measurements were taken using the ultrafast Si diode, they were repeated using Cree’s 4 A, 600 V SiC SBD. Fig. 11 shows the comparison of the switching energy losses per switching cycle in the MOSFET and Diode under half load and full load conditions. Further, the turnON and turn-OFF losses within each device are separated. Under half load conditions, the total switching losses decrease by about 25% from 266 µJ to 200 µJ when the Si diode is replaced by SiC SBD. The 50% decrease during Diode turn OFF losses, and 27% decrease during MOSFET turn ON are primarily responsible for this overall reduction in losses when a SiC SBD is used in the circuit as compared to when a Si diode is used. The MOSFET turn OFF losses and Diode turn ON losses are similar when Si and SiC Diodes are used in this circuit. Under full load conditions, Diode turn OFF losses decrease by 44%, MOSFET turn ON losses decrease by 39%, and Diode turn ON losses decrease by 29% when a SiC diode is used in this circuit as compared to a Si diode. The MOSFET turn OFF losses remain similar in both cases. An overall decrease of 27% in switching losses is measured when the circuit uses a SiC diode as compared to a Si diode. It is worth noting that diode turn ON losses are significantly lower as compared to Si PiN diodes under full load conditions because of a slower turn ON process in a PiN diode as compared to a SBD under higher current operation. These results also show that the dominant reduction in switching losses occurs due to the small reverse recovery losses in the SiC diode as compared to the case with Si diodes. Fig. 13 shows the comparison of the measured efficiency of the entire PFC circuit between Si and SiC diodes. At half load condition, the circuit efficiency increases from 88.4% with the Si diode to 95% with the SiC diode. At full load condition, the circuit efficiency increases from 90% with the Si diode as compared to 93% with the SiC diode. Ostensibly, the slightly higher on-state losses in the SiC SBD result in the relatively smaller gain in the overall circuit efficiency under full load operating condition.
Page 371
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 14 shows the measured MOSFET case temperature as a function of time after initial power up. Initially, the devices were in thermal equilibrium at room temperature. This measurement was done when a full load operating condition was impressed on this circuit. Two conditions were used for these measurements: the first was without a fixed position and speed fan, for case cooling (“in air” condition); and the second was with such a fan. With no fan, the temperature does not stabilize even after 15 minutes of the circuit power up. However, the temperature on the MOSFET was 41oC lower (86oC vs. 127oC) when a SiC SBD was used as compared to when a Si diode was used. When the fan was used for appropriate thermal dissipation, the MOSFET case temperature was only 40oC when a SiC SBD was used as compared to 50oC when a Si PiN diode was used. This increases the thermal “headroom” a circuit designer needs for more rugged operation of the circuit. Based on the measurements presented above, the most significant system advantages offered by SiC SBDs vis-à-vis Si PiN diodes in a PFC circuit are higher circuit efficiency and lower FET case temperature. These advantages can be very effectively harnessed for lowering the cost of the circuit. For a given efficiency, a higher frequency of operation of the circuit can result in smaller (and hence cheaper) inductors and MOSFETs, which are typically the most expensive components in the PFC circuit. For an identical case temperature, a smaller and cheaper MOSFET and heat sinks can be used in the circuit. Another simple circuit modification to lower the total circuit losses involves reducing the gate resistor of the MOSFET. A higher gate resistor is used in typical PFC circuit in order to limit the di/dt in the Si PiN diode, which might result in excessive reverse recovery current, and EMI emissions. Since SiC SBDs can operate under very high di/dt, a smaller MOSFET gate resistance can be utilized. Such a modification will result in lowering the MOSFET turn-OFF losses, which showed little change with direct replacement of SiC SBD with Si PiN diode in the PFC circuit described above.
FIGURES
Fig:1
Fig:2
6. TABLES AND FIGURES TABLE 1 KEY ELECTRONIC PROPERTIES OF Si,
Property
Silicon
Band gap,Eg(eV) 1.12 Electron mobility 1400 Hole mobility,µn 450 Intrinisic carrier 1.5*1010 concentration, Electron mobility, µp(cm2/vs) 1.0 Critical breakdown electicfield, Ecrit(MV/cm) 0.25 Thermal conductivity,(W/cm*K) 1.5
GaAs, and 4H-SiC GaAs 4H-SiC 1.5 92000 400 2.1*106 1.5
3.26 800 140 5*10-9 3.26
0.3
2.2
0.5
3.0-3.8
\
Velammal College of Engineering and Technology, Madurai
Page 372
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Velammal College of Engineering and Technology, Madurai
Page 373
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
7. Software In Mat Lab Programming MATLAB provides the foundation for making progress in vital areas of engineering and science. Over one million people around the world use MATLAB to help them develop cancer therapies, search for new sources of energy, make our cars safer and more fuel-efficient, and explore outer space. By combining a powerful numeric engine and technical programming environment with interactive exploration and visualization tools, MATLAB has become the language of technical computing. MATLAB is a high-level technical computing language and interactive environment for algorithm development, data visualization, data analysis, and numeric computation. Using the MATLAB product, I can solve technical computing problems faster than with traditional programming languages, such as C, C++, and Fortran. I can use MATLAB in a wide range of applications, including signal and image processing, communications, control design, test and measurement, financial modeling and analysis, and computational biology. Add-on toolboxes (collections of special-purpose MATLAB functions, available separately) extend the MATLAB environment to solve particular classes of problems in these application areas. MATLAB provides a number of features for documenting and sharing our work. I can integrate our MATLAB code with other languages and applications, and distribute our MATLAB algorithms and applications. For solving such researching tasks modern sensitive equipment and powerful software environment needed. I chose Mat lab for its advantages: Easy to use and to create programs· Big set of built-in functions for experimental data processing· External hardware interfaces · Good visualization facilities Note: All Program will Explain in conference presentation with simulation output. Example Programming as given below.
Velammal College of Engineering and Technology, Madurai
Page 374
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
% C-V Performance of Schottky diode V=[0:4]; %PhiB=[0.452 0.428 0.446]; %A=[1.6*10^(-6) 1.6*10^(-6) 1.6*10^(-6)]; q=1.6568e-025; Nd=10^(-3); C1=1.6*10^(-6)*sqrt((q*Nd)./((2*0.452)-V)); C2=1.6*10^(-6)*sqrt((q*Nd)./((2*0.428)-V)); C3=0.64*10^(-6)*sqrt((q*Nd./(2*0.446)-V)); %figure figure,plot(V,C1,':*r',V,C2,'--og',V,C3,'-.k'); xlabel('Diode Voltage (V)'); ylabel('Capacitor (C)'); legend('SBD1','SBD2','SBD3') C-V Performance of Schottky diode output
Switching Pulse For (Ms And Ms1)
Output Voltage
8. Hardware simulation Mat lab result
Output Current
Output Current
Ac Input Voltage
Velammal College of Engineering and Technology, Madurai
Page 375
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[4] F. Philippen, B. Burger, «A New High Voltage Schottky Diode based
OUTPUT VOLTAGE
th
th
on Silicon Carbide (SiC)» EPE 2001 Conf. Proc., Graz, August 27 - 29 , 2001.
8. CONCLUSION It may be concluded that SiC SBDs offer significant advantages over silicon PiN diodes in power electronic applications such as PFC. SiC SBDs are commercially available in 600-1200 V, 1-10 A range and can be utilized today to enhance the performance of the PFC circuit by improving the efficiency, reducing the switching losses in the diode and the MOSFET, reducing the MOSFET case temperature and reducing the number of MOSFETs. Additionally, they can be used to simplify or even eliminate the snubber circuits, reducing the heat sink size, or increase the frequency and reduce the size of the magnetic components. In a typical 250 W PFC circuit, an overall decrease of 27% in switching losses is measured when the circuit uses a SiC diode as compared to a Si diode. At full load condition, the circuit efficiency increases from 90% with the Si diode as compared to 93% with the SiC diode. Silicon Carbide is a wide band gap, high breakdown field material allowing high voltage Schottky diodes to be made. SiC Schottky diodes with 300, 600 and 1200-volt are commercially available at CREE. The 600-volt diodes are available with 1, 2, 4, 6, 10 and 20-amp current ratings. The 1200-volt diodes are available with 5, 10 and 20-amp current ratings. The main advantage of a high voltage SiC Schottky diode lies in its superior dynamic performance. Schottky diodes are majority carrier devices and thus do not store charge in their junctions. The reverse recovery charge in the SiC Schottky diode is extremely low and is only the result of junction capacitance, not stored charge. Furthermore, unlike the silicon PiN diode, the reverse recovery characteristics of SiC Schottkys are independent of di/dt, forward current and junction temperature. The maximum junction temperature of 175°C in the SiC Schottkys represents the actual operational temperature. The ultra-low junction charge in SiC Schottkys results in reduced switching losses in a typical hard switched CCM PFC boost converter application.
[5] M. Coyaud, J. P. Ferrieux, C. Schaeffer et al.: "Performances of SiC Schottky Rectifier in Power Factor Correction", Proc. of the 2001 IAS Annual Meeting., October, 2001, pp. 370-375. [6] M. Trivedi, K. Shenai, "Hard- and Soft-Switching Buck Converter Performance of High-Voltage 4H-SiC and Si P-i-N Diodes", Proc. of the 2001 IAS Annual Meeting., October, 2001, pp. 391-395. [7], S. Buso, G. Spiazzi: “Conducted EMI Issues in a 600-W Single Phase Boost PFC Design”, IEEE Transactions on Industry Applications, Vol. 36, No. 2, March/April 2000, pp. 578-585.
10. AUTHORS’ INFORMATION
K. MANICKAVASAGAN Place : 169, Irumbedu Village &Post, Vandavasi Taluk, Kildodungalur, Tiruvannamalai District; Tamil Nadu, India-604403. Date of Birth
: 30/07/1969
UG Studied in Name of college: Crescent Engg. College Vandalur, R.No: 92387,Chennai.Pin:600 048. Tmail Nadu, India.
9. References [1] I. Zverev, M. Treu, H. Kapels, O. Hellmund, R. Rupp, «SiC Schottky rectifiers: performance, reliability and key application», th
th
EPE 2001 Conf. Proc., Graz, August 27 - 29 , 2001. [2] M.E. Levinshtein et al., «High voltage SiC diodes with small recovery th
time», Electronics Letters, 6 July, 2000, vol. 36, no 14, pp. 1241-1242. [3] W. Wright et al. «Comparison of Si and SiC diodes during operation in th
three-phase inverter driving ac induction motor», Electronics Letters, 7 June, 2001, vol. 37, no 12, pp. 787-788.
Velammal College of Engineering and Technology, Madurai
Page 376
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Cascade Data Mining Approach for Network Anomaly Detection System Seelammal.C Department of Information Technology Kalasalingam University, India. [email protected]
Abstract— As the number of networked computers grows, intrusion detection is an essential component in keeping networks secure. Recently data mining methods have gained importance in addressing network security issues, including network intrusion detection a challenging task in network security. This paper, proposes “Heuristic+C4.5,” a method to cascade heuristic clustering and the C4.5 decision tree learning methods for classifying anomalous and normal activities in a computer network. The heuristic clustering method first partitions the training instances into clusters using similarity. The decision tree on each cluster refines the decisions by learning the subgroups within the cluster. To obtain a final decision on classification, the decisions of the heuristic and C4.5 methods are combined using nearest consensus rule. For the evaluation experiments on was conducted on 99KDDcup data set of computer networks. Results show that the detection accuracy of the Heuristic+C4.5 method is high at low false-positive-rate on KDD. Keywords—Anomaly Detection, Classification, Decision Trees, Heuristic Clustering
I. INTRODUCTION With rapid development in the computer based technology, new application areas for computer networks have emerged in Local Area Network and Wide Area Network .This became an attractive target for the abuse and a big vulnerability for the community. Securing this infrastructure has become the one research area. Network intrusion detection systems have become a standard component in security infrastructures. Intrusion detection is “the process of monitoring the events occurring in a computer system or network and analyzing them for signs of intrusions, defined as attempts to compromise the confidentiality, integrity, availability, or to bypass the security mechanisms of a computer or network”. There are generally two types of attacks in network intrusion detection. In misuse detection, each instance in a data set is labeled as ‘normal’ or ‘intrusion’ and a learning algorithm are trained over the labeled data. . An anomaly detection technique builds models of normal behavior, and automatically detects any deviation from it, flagging the latter as suspect. Attacks fall into four main categories[8]
Velammal College of Engineering and Technology, Madurai
they are Denial of Service (DoS) is a class of attacks where an attacker makes some computing or memory resource too busy or too full to handle legitimate requests, A remote to user (R2L) attack is class of attacks where an attacker sends packets to a machine over a network, then exploits machine’s vulnerability to illegally gain local access a user, User to root exploits is a class of attacks where an attacker starts out with access to a normal user account on the system and is able to exploit vulnerability to gain root access to the system. Probing is a class of attacks where an attacker scans a network to gather information or find known vulnerabilities. The main reason for using Data Mining Techniques for intrusion detection system is due to the enormous volume of existing and newly appearing network data that require for processing. The data accumulated each day by a network is huge. Several data mining techniques such as clustering, classification, and association rules are proving to be useful for gathering different knowledge for intrusion detection. Unsupervised Anomaly Detection (UAD) algorithms have the major advantage of being able to process unlabeled data and detect intrusions that otherwise could not be detected. The goal of data clustering, or unsupervised learning, is to discovery a “natural” grouping in a set of patterns, points, or objects, without knowledge of any class labels. II. RELATED WORKS Network-based computer systems play increasingly vital roles in modern society, security of network systems has become important than ever before. As an active defense technology, IDS (Intrusion Detection Systems) attempts to identify intrusions in secure architecture. In 1970s, the U.S Department of Defense outlined security goals for audit mechanisms, among which were allowing the discovery of attempts to bypass protection mechanism. Over the past years there are many different types of intrusions, and different detectors are needed to detect normal and anomalous activities in network. Some Clustering algorithms have recently gained attention in related literature, since they can help current intrusion detection systems in several respects.
Page 377
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Several existing supervised and unsupervised anomaly detection schemes and their variations are evaluated on the DARPA 1998 data set of network connections [3] as well as on real network data using existing standard evaluation techniques. Recently, anomaly detection has been used for identifying attacks in computer networks [1], malicious activities in computer systems and misuse in web systems [6], [7]. A more recent class of Anomaly Detection Systems developed using machine learning techniques like artificial neural-networks [8], Kohonen’s self organizing maps [9] fuzzy classifiers [10] and others [7] have become popular because of their high detection accuracies at low false positives. Data mining for security knowledge [2] employs a number of search algorithms, such as statistical analysis, deviation analysis, rule induction, neural abduction, making associations, correlations, and clustering. Another data mining approach [6] for intrusion detection is the application of clustering techniques for effective intrusion .How ever , the existing works that is based on either K-Means or K-Medoids have two shortcomings in clustering large network datasets namely number of clusters dependency and lacking of the ability of dealing with character attributes in the network transactions. Among these the number of clusters dependency suggests that the value of K is very critical to the clustering result and should be fixed before clustering. Trained a Hidden Markov Model [7] implemented on the trained datasets that he used to train instance-based learner. Fixed-width and k-nearest neighbor clustering techniques [4] to connection logs looking for outliers, which represent anomalies in the network traffic. A more recent class of IDS developed using machine learning techniques like artificial neural networks, Kohenen’s self organizing maps, fuzzy classifiers, symbolic dynamics [10], and multivariate analysis. However, the ADS related studies cited above have drawbacks: the majority of these works evaluate the performance of anomaly detection methods with single machine learning techniques like artificial neural-networks, pattern matching, etc., have high false positive rates. While recent advances in machine learning show that fusion [15], selection [15], and cascading [20] of multiple machine learning methods have a better performance yield over individual methods. This paper, presents a novel anomaly detection method, called “Heuristic+C4.5,” developed by cascading two machine learning algorithms: 1) the heuristic clustering and 2) the C4.5 decision tree learning. In the first stage, heuristic clustering is performed on training instances to obtain disjoint clusters. Each heuristic cluster represents a region of similar instances, “similar” in terms of Euclidean distances between the instances and their cluster centroids. In the second stage of Heuristic+C4.5, the heuristic method is cascaded with the C4.5 decision tree learning by building
Velammal College of Engineering and Technology, Madurai
an C4.5 decision tree using the instances in each heuristic cluster. For the evaluation experiments are performed on 99KDDcup dataset. Performance evaluation of the Heuristic+C4.5 cascading approach is conducted using measures: 1. Detection accuracy or true positive rate (TPR), 2. False positive rate (FPR), 3. Precision, and 4. Total accuracy (or accuracy), The performance of Heuristic+C4.5 is empirically compared with the performance of individual heuristic clustering and the C4.5 decision tree classification algorithms.
A. Contributions of the Paper The contributions of the paper are enumerated as follows. The paper presents a novel method to cascade the heuristic clustering and C4.5 decision tree learning methods for classifying data from normal and anomalous behaviors in a computer network. This paper evaluates the performance of Heuristic+C4.5 classifier, and compares it with the individual heuristic clustering and C4.5 decision tree methods using three performance measures. The paper presents a novel method for cascading two successful data partition methods for improving classification performance. From an anomaly detection perspective, the paper presents a high performance anomaly detection system. The rest of the paper is organized as follows: In Section 3, briefly discuss the heuristic clustering and C4.5 decision tree learning-based anomaly detection methods. In Section 4, present the Heuristic+C4.5 method for anomaly detection. In Section 5, discuss the experimental results and in the final section conclude the work and future directions. III.
ANOMALY DETECTION WITH HEURISTIC CLUSTERING AND C4.5 DECISION TREE LEARNING METHODS
In this section, briefly discuss the heuristic [2] clustering and the C4.5 decision tree classification [10] methods for anomaly detection.
A.
Anomaly Detection with Heuristic Clustering
This section, describes the heuristic algorithm, and illustrates how to apply this algorithm to generate detection models from audit data. Here audit data refers to preprocessed time-stamped audit records, each with a number of attributes.
1) Data Preprocessing: Depending on the monitoring data and the data mining algorithm, it may be necessary or
Page 378
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
beneficial to perform cleaning and filtering of the data in order to avoid the generation of misleading or inappropriate results. Disturbances can be noise, known changes such as dependencies on the time of day, missing data fields in some of the data records etc. 2) Notations of Some Terms: Some notations needed in heuristic clustering algorithm are Notation1: Let H = {H1 , H 2 ,...., H m } be a set of attribute values, the m is number of attribute values. Some attribute values in KDD are duration, srcbytes, dest-bytes, flag, etc.., Notation 2: Let H = H N UH s and
HN I HS = ϕ where HN is the subset of numerical attribute (e.g., no of bytes), and HS is the subset of character attribute. (e.g., service, Protocol). Notation 3: Let, ei = ( hi1 , hi 2 ,.....him ) , ei is a record, the m is number of attribute values and hij is the value of Hm Notation 4: E = {e1 , e2 ,....., en } , E is the set of records; n is the number of packets.
3) Heuristic Clustering Algorithm (HCA): In the HCA algorithms no need to fix the value of K (K-means) in the beginning of clustering. A novel heuristic clustering technology is used to determine the number of cluster automatically. For clustering, similarity between ei and every center of cluster is calculated. Sim(ei,Cj) and the similarity between every center of cluster Sim(C), if the minimal of Sim(ei,Cj) is more than the minimal of Sim(C) , create a new cluster, and the center of cluster is ei , otherwise insert the ei into Cj. Step 1. Confirm two initial cluster centers by algorithm search ( ). Step 2. Import a new record ei. Repeat 3 to 5 until no more records. Step 3. Compute the similarity by algorithm Similar (), and find
Min( Sim(ei , C j )), Min( Sim(C )). Step 4.If Min( Sim(ei , C j )) > Min( Sim(C )) then
Ck +1 = {e j }, C = {C1 , C2 ,...., C k +1} create a new cluster, and the ei is the center of the new cluster.
Velammal College of Engineering and Technology, Madurai
Else C j = C j U {ei } , insert ei into Cj. 4) The Method of Computing Similarity:The audit data consists of numerical attribute (e.g. the receive Bytes) and character attribute. (e.g. Protocol ).But whether k-Means or K-medoids all lack the ability of dealing with character attribute. The HCA algorithm resolves this problem by processing the character attribute using the method of attribute matching. The similarity of character attributes using equation (1) let ei and ej be two records in the E. all containing m attributes (including P character attributes), the nhik and nhjk is the number of hik and hjk respectively.
(nhik + nhjk ) *A hik * nkjk ) k =1 p
Sim P (ei,ej ) =
∑ (n
(1)
if (hik=hjk) then A= 0 else A=1. The similarity of numerical attribute (to the numerical attribute, still use the classical Euclidean distance to computer similarity) q
Sim N (ei , e j ) = sqrt ( ∑ | hik − hjk |2
(2)
k =1
The similarity of two records (including similarity of numerical attribute and similarity of character attribute)
Sim(ei, ej ) = Sim N (ei, ej ) + Sim P (ei, ej )
(3)
5) The Center of Cluster:A cluster are represented by its cluster center .In the HCA algorithm, the algorithm Count ( ) is used to compute the cluster center. The center of a cluster is composed of the center of numerical attributes and character attribute. Let P= (PN + PS), and P= (P1, P2,…., Pm) where PN is the center of numerical attribute, the Ps is the center of character attribute, n
P N i = (1 / n) ∑ hji
(4)
j =1
i = 1,2,...., p( p <= m) The hji is the numerical attribute. The Ps is a frequent character attribute set which consists of q (q = m-p) most frequent character attribute.
Page 379
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6) The Initial Center of Clustering: In the beginning of clustering, confirm two initial center of clustering by the algorithm Search ( ). Algorithm：Search_m (E,l). Import：data set E,the number of sampling l Output：initial center m1,m2. (1) Sampling E, get S1,S2,..,Sl (2)For i=1 to l do mi＝Count_m(Si)； (3)For i=1 to l do m＝Count_m(mi)； // m=center {m1,m2, m3……..ml} (4) m1=m , m2=max (Sim (m, mi))； 7) Detection Rule: A method is proposed to detect anomaly that does not either depend on the population ratio of the clusters. In our labeling method, assume that center of a normal cluster is highly close to the initial cluster centers vh which are created from the clustering. In other words, if a cluster is normal, the distance between the center of the cluster and vh will be small, otherwise it will be large. For each cluster center Cj (1 ≤ j ≤ k), calculate the maximum distance to vh. Then calculate the average distance of the maximum distances. If the maximum distance from a cluster to vh is less than the maximum average distance, finally label the cluster as normal. Otherwise, label as attack. Fig.1 shows how we can detect the anomaly instances.
Fig. 1 Assigning Scores to IDS
B. Anomaly Detection with C4.5 Decision Tree In this section, a brief introduction of the classification algorithms used in the IDS, i.e., the C4.5 algorithm for building decision trees.
Velammal College of Engineering and Technology, Madurai
C4.5 Algorithm: A classifier, which is a function (or 1) model) that assigns a class label to each data item described by a set of attributes, is often needed in these classification tasks. There are quite a few machine learning approaches for generating classification models, among which decision tree learning [2] is a typical one. As an example, C4.5 in [2] builds a decision tree where each internal node denotes a test on an attribute, each branch represents an outcome of the test, and leaf nodes represent classes or class distributions. The top-most node in a tree is the root node. The tree is a model generated by the classification algorithm. In order to classify an unknown sample, the attribute values of the sample are tested against the decision tree. A path is traced from the root to a leaf node that holds the class prediction for that sample. The C4.5 algorithm builds a decision tree, from the root node, by choosing one remaining attribute with the highest information gain as the test for the current node. In this paper, a later version of the C4.5 algorithm, will be used to construct the decision trees for classification. The specific algorithm is given below. The reader is referred to [2] for further details. Algorithm: Generate_decision_tree. Generate a decision tree from the given training data. Input: training samples, represented by discrete/continuous attributes; the set of candidate attributes, attribute-list. Output: a decision tree Method: (1) create a node N; (2) if samples are all of the same class, C, then (3) return N as a leaf node labeled with the class C; (4) if attribute-list is empty then (5) return N as a leaf node labeled with the most common class in samples; (6) select test-attribute, the attribute among attribute-list with the highest information gain; (7) label node N with test-attribute; (8) for each known value ai of test-attribute (9) grow a branch from node N for the condition test-attribute = ai; (10) let si be the set of samples in samples For which test-attribute = ai; (11) if si is empty then (12) attach a leaf labeled with the most common class in samples; (13) else attach the node returned by Generate_decision_tree (Si, attributelist).
Page 380
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2) Attribute Selection: The information gain measure used in step (6) of above C4.5 algorithm is used to select the test attribute at each node in the tree. Such a measure is referred to as an attribute selection measure or a measure of the goodness of split. The attribute with the highest information gain is chosen as the test attribute for the current node. This attribute minimizes the information needed to classify the samples in the resulting partitions. Such an information-theoretic approach minimizes the expected number of tests needed to classify an object and guarantees that a simple (but not necessarily the simplest) tree is found. Information Gain: Imagine selecting one case at 3) random from a set S of cases and announcing that it belongs to some class Cj. The probability that an arbitrary sample belongs to class Cj is estimated by
Pi =
freq(cj , s ) |s|
This represents the difference between the information needed to identify an element of T and the information needed to identify an element of T after the value of attribute X has been obtained. Thus, it is the gain in information due to attribute X. Gain Ratio Criterion:The notion of information gain 4) introduced earlier tends to favor attributes that have a large number of values. For example, if we have an attribute D that has a distinct value for each record, then Info(D,T) is 0, thus Gain(D,T) is maximal. To compensate for this, it was suggested in [2] to use the following ratio instead of gain. Split info is the information due to the split of T on the basis of the value of the categorical attribute D, which is defined by n | Ti | |T | SplitInfo( D) = − * log 2 i (9) | T | | T | i =1 And the gain ratio is then calculated by
∑
(5)
Where | S | denotes the number of samples in the set S and o the information it conveys is −log2 Pi bits. Suppose probability distribution P = {p1, p2,…, pn} is given then the information conveyed by this distribution, also called the entropy of P, is well known as If a set T of records is partitioned into disjoin n
Info( P) = ∑ − Pi log 2 Pi
(6)
i =1
If a set T of records is partitioned into disjoint exhaustive classes C1 ,C2 , ,Ck on the basis of the value of the categorical attribute, then the information needed to identify the class of an element of T is Info(T) = Info(P) , where P is the probability distribution of the partition (C1 ,C2 , ,Ck). First partition T on the basis of the value of a non-categorical attribute X into sets T1 ,T2 , …,Tn , then the information needed to identify the class of an element of T becomes the weighted average of the information needed to identify the class of an element of Ti , i.e. the weighted average of Info(Ti) , n |T | Info( X , T ) = ∑ i * Info(Ti ) |T | i =1
(7)
GainRatio( D, T ) =
Gain( D, T ) SplitInfo( D, T )
(10)
The gain ratio, expresses the proportion of useful information generated by split, i.e., that appears helpful for classification. If the split is near-trivial, split information will be small and this ratio will be unstable. To avoid this, the gain ratio criterion selects a test to maximize the ratio above, subject to the constraint that the information gain must be large, at least as great as the average gain over all tests examined. HEURISTIC +C4.5 METHOD FOR ANOMALY DETECTION Heuristic -based anomaly detection method are first applied to partition the training pace into k disjoint clusters C1, C2 . . . Ck. Then, an C4.5 decision tree is trained with the instances in ach Heuristic cluster. The Heuristic method ensures that each training instance is associated with only one cluster. However, if there are any subgroups or overlaps within a cluster, the C4.5 decision tree trained on that cluster refines the decision boundaries by partitioning the instances with a set of if-then rules over the feature space. The decisions of the Heuristic and C4.5 decision tree methods are combined to give a final Decision. The overall architecture of the cascade intrusion detection system is shown in fig 2. For combining the Heuristic and C4.5 decision tree methods, this paper proposes a rule called Nearest-consensus rule.
IV.
The information gain, Gain(X ,T ) , is then defined as
Gain( X , T ) = Info(T ) − Info( X , T )
(8)
Velammal College of Engineering and Technology, Madurai
Page 381
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1. TPR or recall is the percentage of anomaly instances correctly detected, 2. FPR is the percentage of normal instances incorrectly classified as anomaly, 3. “Precision” is the percentage of correctly detected anomaly instances over all the detected anomaly instances,
Training instances
Cluster
1
d1
<
DT 1
Cluster…… 2
d2 <
Cluster n
The experiment demonstrates the clustering can detect the unknown attacks and generate the new intrusion pattern
….. dn
DT 2……
The classification of normal and abnormal classes has been represented in the form of a matrix called Confusion matrix as given in Table I.
DT k
automatically. TABLE I
Fig. 2 Overall Architecture of Cascade Anomaly Detection System
CONFUSION MATRIX
A. Nearest-Consensus Rule For clusters C1,C2 . . .Cn formed by applying the heuristic clustering method on the training instances, r1, r2 . . . rn be the centroids of C1, C2, . . .Ck, respectively. Calculate the distances (d1, d2 ,…. dn ) among centroids. In the Nearest-consensus rule, the decision of the nearest candidate cluster in which there is consensus between the decisions of the Heuristic and the C4.5 decision tree methods. EXPERIMENTS AND RESULTS V. In this section discuss experiment with the heuristic clustering using the 10% subset of KDD-99 data. The data in the experiment was acquired from the 1998 DARPA intrusion detection evaluation .They set up an environment to acquire raw TCP/IP dump data for a local-area network(LAN) simulating a typical U.S.Air Force LAN. More than 200 instances of 58 attack types were launched against victim UNlX and Windows NT hosts in tree weeks of training data and two week of test data. For each TCP/IP connection, 41 various quantitative and qualitative features were extracted. Attacks fall into four main categories: DOS: denial of service R2L: unauthorized access from a remote machine U2R: unauthorized access to local super user (root) Probing: surveillance and other probing This dataset was acquired from the 1998 DARPA intrusion detection evaluation program consists of 11000 connections and 22 types of intrusions in the test set. The test set consists of 9 subtests .In heuristic clustering the relations between categorical and continuous features are handled naturally, without any forced conversions (kmeans) between these two types of features. The decision tree on each cluster refines the decisions by learning the subgroups within the cluster.
Actual \ Normal
Predicted
Abnormal
Class Normal
True positive (TP)
False negative (FN)
Abnormal
False positive (FP)
True negative (TN)
The detection rate is computed using the equations
DetectionRate =
TN * 100 ( FN + TN ) FP * 100 (TP + FP)
(12)
( FP + FN ) * 100 (TP + FP + FN + TN )
(13)
(11) FalsePositiveRate =
ErrorRate =
The experimental results obtained from the experiments conducted on 10% of the KDD dataset using Cascading without nearest consensus rule is given in Table II. The results obtained from heuristic clustering based average intrusion detection accuracy is 85.05 percent at a false-positive-rate of 10.08 percent on 10% subset of the training dataset KDDcup1999. The results obtained from C4.5 based average intrusion detection accuracy is 83.21 percent at a false-positive-rate of 9.28 percent. The experimental results obtained from three experiments conducted on 10% of the KDD dataset using Cascading with Nearest Consensus (NC) rule is given in Table III.
The following measures are used for evaluating the performance
Velammal College of Engineering and Technology, Madurai
Page 382
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE II
Iteration
RESULTS FOR CASCADE ANOMALY DETECTION WITHOUT NC RULE
Iteration
Training
Detection
False
Error
instances
Rate
Positive
Rate
Training
Detection
False
Error
Instances
Rate
Positive
Rate
Rate
Rate
0
415
90.9%
2.8%
3.4%
1
820
85.7%
2.5%
2.9%
0
415
78.1%
5.0%
7.2%
2
1200
80.0%
4.4%
5.4%
1
820
80.0%
4.4%
7.0%
3
3000
80.0%
6.4%
7.5%
2
1200
79.4%
4.2%
7.3%
4
5200
83.3%
2.4%
3.1%
3
3000
77.7%
6.8%
10.1%
5
6130
85.7%
1.9%
2.7%
4
5200
82.0%
6.3%
9.3%
6
7117
83.3%
3.1%
3.9%
5
6130
80.5%
5.9%
8.7%
7
8045
77.7%
6.1%
7.3%
6
7117
80.4%
6.4%
9.4%
8
9300
87.5%
2.8%
3.6%
7
8045
79.5%
6.7%
10.5%
9
11245
85.7%
5.3%
5.4%
8
9300
82.6%
5.5%
7.6%
9
11245
80.4%
6.4%
10.7%
95.00%
90.00%
From Table II, it is observed that the average detection rate is 83.98%, the average false positive rate is 3.77% and the average error rate is 4.52% for intrusion detection by cascading without Nearest Consensus rule using KDD dataset. From Table III, it is observed that the average detection rate is 89.98%, the average false positive rate is 3.77% and the average error rate is 4.52% for intrusion detection by cascading without Nearest Consensus rule using KDD dataset. 8.00%
85.00% Detection rate
7.00%
6.00%
80.00%
5.00% FPR 75.00% 4.00%
3.00% 70.00% 0
1
2
3
4
5
6
7
8
9 2.00%
iteration without NC rule
with NC rule 1.00%
Fig. 3 Performance of the heuristic + C4.5 method without Nearest Consensus ( NC-Rule) TABLE III RESULTS FOR CASCADE ANOMALY DETECTION WITH NC RULE
Velammal College of Engineering and Technology, Madurai
0.00% 0
1
2
3
4
5
6
7
8
9
Iteration without NC rule
with NC rule
Page 383
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 4 Performance of the heuristic + C4.5 method with Nearest Consensus (NC-Rule)
From Fig.3, Fig.4 it is observed that the cascading with NC has better performance in terms of Detection rate and false positive rate. CONCLUSIONS AND FUTURE WORKS VI. This paper Network intrusion detection with heuristic clustering and C4.5 incorporates the idea of applying data mining techniques to intrusion detection system to maximize the effectiveness in identifying attacks, thereby helping the users to construct more secure information systems. The main advantage of our algorithm is that the relations between categorical and continuous features in KDD are handled naturally, without any forced conversions (k-means) between these two types of features. C4.5 decision tree built on each cluster learns the subgroups within the cluster and partitions the decision space into finer classification regions; thereby improving the overall classification performance. For future work, we need to verify performance of the clustering algorithm over real data and make a new benchmark dataset for intrusion detection, because KDD Cup 1999 dataset was generated in the virtual network and the attacks included in it are greatly old-fashioned. Future works carried out for evaluating the performance of anomaly detection methods with new clustering, labeling, cascading various data mining methods with dataset drawn from various application domains.
security “, IEEE Trans.Syst.,Man and Cybern., vol.35, no.2,pp.302,Apr.2005. [6] M. Thottan and C. Ji, “Anomaly Detection in IP Networks,” IEEE Trans. Signal Processing, vol. 51, no. 8, pp. 2191-2204, May.2003. [7] C. Kruegel and G. Vigna, “Anomaly Detection of Web-Based Attacks,” Proc. ACM Conf. Computer and Comm. Security, Oct. 2003. [8] Z. Zhang, J. Li, C.N. Manikopoulos, J. Jorgenson, and J. Ucles,“HIDE: A Hierarchical Network Intrusion Detection System Using Statistical Preprocessing and Neural Network Classification,” Proc. 2001 IEEE Workshop Information Assurance, pp. 85-90, June 2001. [9] S.T. Sarasamma, Q.A. Zhu, and J. Huff, “Hierarchical Kohonen Net for Anomaly Detection in Network Security,” IEEE Trans.Systems, Man, and Cybernetics-Part B, vol. 35, no. 2, pp.450,Apr. 2005. [10] J. Gomez and D.D. Gup ta, “Evolving Fuzzy Classifiers for Intrusion Detection,” Proc. 2002 IEEE Workshop Information Assurance, June 2001. [11] A. Ray, “Symbolic Dynamic Analysis of Complex Systems for Anomaly Detection,” Signal Processing, vol. 84, no. 7, pp. 1115-1130, 2004. [12] N. Ye, S.M. Emran, Q. Chen, and S. Vilbert, “Multivariate Statistical Analysis of Audit Trails for Host-Based Intrusion Detection,” IEEE Trans. Computers, vol. 51, no. 7, pp. 810-820, 2002. [13] H.S. Javitz and A. Valdes, “The SRI IDES Statistical Anomaly Detector,” Proc. IEEE Symp. Security and Privacy, Vol.87,no.7,pp. 316326, May 1991. [14] J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas, “On Combining Classifiers,” IEEE Trans. Pattern Analysis and Machine Intelligence,vol. 20, no. 3, pp. 226-239, Mar. 1998. [15] L.I. Kuncheva, “Switching between Selection and Fusion in Combining Classifiers: An Experiment,” IEEE Trans. Systems, Man, and Cybernetics, vol. 32, no. 2, pp. 146-156, Apr. 2002. [16] R.P. Lippman, D.J. Fried, I. Graf, J. Haines, K. Kendall, D.McClung, D. Weber, S. Webster, D. Wyschogrod, R.K. Cunningham, and M.A. Zissman, “Evaluating Intrusion Detection Systems:The 1998 DARPA Off-Line Intrusion Detection Evaluation,” Proc.DARPA Information Survivability Conf. and Exposition (DISCEX ’00), pp. 12-26, Jan. 2000. [17] The third international knowledge discovery and data mining tools competition dataset KDD99-Cup, http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html, 1999.
ACKNOWLEDGEMENT This research is supported by Bose of Anna University for his support in completion of this work. Special thanks to Sam D Raj Kumar for their efforts in evaluating the system. Finally I am also very thankful to the reviewers for their detailed reviews and constructive comments, which have helped to significantly improve the quality of this paper. REFERENCES [1] S.R.Gaddam, V.V. Phoha, and K.S. Balagani, “KMeans+ID3: A Novel Method for Supervised Anomaly Detection by Cascading K-Means Clustering and ID3 Decision Tree Learning Methods ,” IEEE Trans.Knowledge and Data Eng, vol.19, no.3,pp.345, Mar.2007. [2] Zhi-Xin Yu; Jing-Ran Chen; Tian-Qing Zhu, “A novel adaptive intrusion detection system based on data mining”, Proceedings of Fourth International Conference on Machine Learning and Cybernetics Volume 4, 1821 Aug. 2005. [3] Jungsuk SONG, Kenji OHIRA, Hiroki TAKAKURA, Nonmembers, Yasuo OKABE,and Yongjin KWON, “ A Clustering Method for Improving Performance of Anomaly-Based Intrusion Detection System ,” IEICE Trans. Inf. & Syst., vol 91–d, no.5,pp.350, May 2008. [4] A. Lazarevic, A. Ozgur, L. Ertoz, J. Srivastava, and V. Kumar, “A Comparative Study of Anomaly Detection Schemes in Network Intrusion Detection” , Proceedings of SIAM International Confonference on Data Mining, May 2003. [5] Suseela T.Sarasamma , Quiming A.Zhu , and Julie Huff , “ Hierarchial KMeans+ID3: A Novel Method for Supervised Anomaly Detection by Cascading K-Meaohenen net for anomaly detection in network
Velammal College of Engineering and Technology, Madurai
Page 384
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Membrane Computing - an Overview R.Raja Rajeswari #1, Devi Thirupathi *2 #
Assistant Professor of Computer Science, M.V.M Govt Arts College(W), Dindigul, Tamil Nadu, India. 1
[email protected] *
Head Incharge, Department of Computer Applications, Bharathiar University, Coimbatore 641 046, Tamil Nadu, India. 2
[email protected]
Abstract— Membrane Computing is an emerging natural computing paradigm that emulates membrane operations on an imaginary membrane platform to solve computational problems. P Systems form the base for Membrane computing and there are three types of P Systems, Cell like P Systems, Tissue like P Systems and Neural like P Systems. Transition P Systems or P Systems, in general belong to the first category, Cell like P Systems which simulate a living cell for computing. This research paper gives an introduction to P Systems based on a preliminary survey of existing literature. Keywords— Membrane Computing, P Systems, Membrane Structure, Evolution Rules, Generative Device.
I. NATURAL COMPUTING Nature has always fascinated mankind through all its processes. Inspired by Nature, many inventions have happened: from aeroplanes that simulated birds to computers which simulate brain. Now after the buzz of Information and Communication technologies, this is the age of Nano-technology which closely resembles Nature’s technology. Nano-technology is engineering of nano particles, i.e. particles of size 1 to 100 nm (1 nm = 10 -9 m). Examples of Nano particles are atoms, Zinc oxide particles and DNA molecules. If nature is closely observed, it can be inferred that any natural process is a nano process at basic level. (e.g.) Raindrops are formed using nano particles called condensation nuclei. And a pumpkin under SEM (Scanning Electron Microscopy) will fascinate anyone with marvellous nano structures which cannot be artificially manufactured. Thus Nature has been the source of inspiration behind any man made technology.
Computer scientists were also fascinated by Nature which led to the budding of a new branch in computer science called Natural Computing or Biologically Inspired Computing. Natural computing dates back to the evolution of Neural networks and Genetic algorithms which emulate brain and genetic operators in computer science respectively, long back. Recent trends in Natural Computing saw the use of Swarm Intelligence in Soft computing and the budding of a new branch called Molecular computing. Molecular computing is a branch of Computer Science which is instigated by the biomolecular interactions in living organisms. There are two forms of molecular computing: (i) DNA computing and (ii) Membrane computing. DNA Computing DNA Computing [6] is computing with DNA strands to solve computational problems. Its history dates back to 1994, when the Computer Scientist Adleman solved the famous NP - complete problem, Hamiltonian Path Problem using DNA strands in a test tube. The advantages of DNA Computing are its massive parallelism and eco-friendly nature, which reduces the environmental hazards posed by conventional computers to nature and hence comes close to Green Computing. Its main draw backs are huge requirements of DNA and errors due to DNA operations, when it comes to practical implementation. But theoretically DNA Computing has developed well and has given solutions to NP Complete class problems of Computer Science. D.
E.
Membrane Computing The Second form of Natural computing, Membrane Computing does the computing in an imaginary membrane
Velammal College of Engineering and Technology, Madurai
Page 385
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
platform. It simulates the molecular interactions inside the cell and also the membrane operations. An initial survey on existing literature [1] infers the following results. Gheorge Paun [7] who initiated this field in 1998 defined Membrane computing as “Membrane computing is a branch of Natural computing which abstracts computing models from the structure and the functioning of living cells, as well as from the organization of cells in tissues or other higher order structures”. The next section discusses the P Systems which form the basis for Membrane Computing as defined in [8], [9] and [10]. II. P SYSTEMS P Systems form the basic models of Membrane Computing and there are three main types of P systems: 1 Cell like P systems, 2 Tissue like P systems and 3 Neural like P systems. The first type is developed with cell and its basic ingredients as the base. Tissue like P systems simulate several one membrane cells evolving in a common environment. Neural like P systems are similar to Tissue like P systems but they have also a state which controls the evolution. And cell like P systems are the ones which are addressed simply as P systems or Transition P systems in general. There are many variants of this P system like P systems with Symport / Antiport rules. But this research paper attempts only to analyze P Systems or Transition P Systems briefly. The three fundamental features of cells which are used in this computing model are: Membrane Structure, Multisets of chemical compounds and Evolution Rules which are analogous to Functions, Variables and Statements in a Programming Language. A. Membrane Structure The fundamental ingredient of a P system is the membrane structure as depicted following as a Venn diagram in Fig 1. 2 1
4
5
region 3
elementary membrane environment
Velammal College of Engineering and Technology, Madurai
Fig.1 Example
A membrane structure is a hierarchically arranged set of membranes, contained in a distinguished external membrane. Several membranes can be placed inside the skin membrane. A membrane without any other membrane inside it is said to be elementary. Each membrane determines a compartment, also called region, the space delimited from above by it and from below by the membranes placed directly inside, if any exists. Usually the membranes are identified by labels from a given set of labels. The symbolic representation of the above said membrane structures is 1 [ 2[ ] 2 3 [ ] 3 4 [5[]5]4]1. The number of membranes in a membrane structure is called the degree of the membrane structure. B. Evolution Rules In the basic P system, each region contains a multiset of symbol objects which correspond to the chemicals swimming in a solution in a cell compartment. The objects evolve by means of evolution rules [4] which are also localized, associated with the region of the membrane structure. And there are three main types of rules: 1 Multiset – rewriting rules (evolution rules) 2 Communication rules and 3 Rules for handling membranes. And Transition P systems use only the first type rules. They are of the form u Æ ν where u and ν are multisets of objects. Also we can add target indications like here, in, out to the rules indicating here is the same region, in goes immediately into a directly lower membrane and out indicates that the object has to exit the membrane, thus becoming an element of the same region surrounding it. (eg) aab Æ (a, here) (b, out) (c, here) (c, in) After using this rule in a given region of a membrane structure, two copies of a and one b are consumed and one copy of a, one of b and two of c are produced; one a and one c remain in the same region, new copy of b exits the membrane and one c enters into one of child membranes. The above said rules with atleast two objects in the L.H.S. are called cooperative rules. Rules of the form ca Æ cν, where c is a catalyst are called catalytic rules. Rules of the form a Æ v where a is an object and v is a multiset are called non cooperative. Rules of the form u Æ νδ, where δ denotes the action of membrane, dissolves the corresponding
Page 386
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
12 2 1 w1 = aac w2 = a R1 = { r1 : c---> (c, in2), r2 : c .---> (b, in2)} P1 = { r1 > r2 } R2 = {a --> (a, out), ac --> δ) P2 = Ǿ
membrane. Also in a P system rules are used in the maximally parallel manner, non deterministically choosing the rules and the objects. A sequence of transitions constitutes a computation. A computation is successful if it halts, it reaches a configuration where no rule can be applied to the existing objects, and the output region i0 still exists in the halting configuration. If the output region specified is an internal region, then it is called an internal output. And the objects present in the output region in the halting configuration is the result of the computation. If io = 0, then count the objects which leave the system during the computation and this is called as external output. A possible extension of the definition is to consider a terminal set of objects, T and to count only the copies of objects from T, present in the output region. Thus basically a P system computes a set of numbers. And a system is said to be propagating if there is no rule which diminishes the number of objects in the system. III. DEFINING TRANSITION P SYSTEMS A transition P system of degree n, n>1 is a construct π = (V, µ, w1, .... wn, (R1, P1) .....(Rn, Pn), i0), where : i) V is an alphabet; its elements are called objects µ is a membrane structure of degree n, with ii) membranes and the regions labeled in a one to one manner with elements in a given set. wi, 1 ≤ i ≤ n are strings from V* associated iii) with the regions 1, 2 .... n of µ; Ri, 1 ≤ i ≤ n, are finite sets of evolution rules iv) over V associated with the regions 1, 2… n of µ; Pi is a partial order relation over Ri 1 ≤ i ≤ n specifying a priority relation among rules of R i. An evolution rule is a pair (u,v) which denotes u Æ v where u is in V and v = v’ or v = ν’δ where v’ is a string over (V x { here, out} ) U (V x { inj /1 ≤ j ≤ n } ) and δ is a special symbol not in V. The length of u is called the radius of the rule u Æ ν v) i0, is a number betwen 1 and n which specifies the output membrane of π Example : A P system of degree 2 can be defined as follows: π = (V, µ, w1, w2, (R1, P1) , (R2, P2) , 1), V = { a, b, c, d} µ= [[ ] ]
Velammal College of Engineering and Technology, Madurai
Generally a P system can be used for solving two types of tasks: as a generative device – start from an initial configuration and collect all sets of natural numbers describing the multiplicities of objects from an output membrane or ejected from the membrane system at the end of a successful computation; and as a decidability device: introduce a problem in an initial configuration, in a specific encoding and look for the answer after a specified number of steps. The next section introduces an example for this [3]. IV. P SYSTEM AS A GENERATIVE DEVICE Consider the P system of degree 4. Π=(V, M, w1, w2, w3, w4, (R1, P1), (R2, P2), (R3, P3), (R4, P4), 4) V = { a, b, b, c, f}, M=[ [ [ ] [ ] ] ] 1 23 34 4 2 1 w1 = λ, R1 = Ǿ, P 1 = Ǿ w2 = λ R2 = {b’ ---> b, b ---> b (c, in4)}, P2 ={ r1> r2), r1 : ff --> af, r2 : f --> a δ }, w3 = af, R3 = { a--> ab’, a --> b’ δ, f --> ff}, P3 = Ǿ w 4 = Ǿ , R 4 = Ǿ , P4 = Ǿ
af a ---> ab' a ---> b' δ8 f ---> ff
b' --> b b --> b c(c, in 4 ) ff --> af > f --> a8δ
4 3
2 1
Page 387
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.2 Initial configuration
Initially w1 = Ǿ, w2 = Ǿ as shown in Fig.2. Hence start with µ3. Here the variables objects are a, f and the rules that can be applied are ……. 1 aÆ ab’ ……. 2 a Æ b’δ f Æ ff ……. 3 Using rules 1 and 3 parallely after n steps n>= 0, ab’n and 2n occurrences of f can be obtained. Now apply rules 2 & 3. That results in b’n+1 and 2n+1 occurences of f. And Membrane3 is dissolved and the objects pass onto Membrane2. At this moment rules of Membrane3 becomes inactive and rules of Membrane 2 are active. Applying rules b’Æb, ffÆaf parallely yields bn+1 and the number of f occurrences gets divided by two. In the next step, from bn+1, n+1 occurrences of c are introduced in membrane 4 using the rule b-Æb(c,in 4). At the same time, the number of f occurrences is again divided by two. This can be continued and at each step further n+1 occurrences of c are introduced in the output membrane. This can be done n+1 steps: n times when the rule ffÆaf and once when the rule fÆaδ is used. In this moment, Membrane2 is dissolved and no further step is possible. The obtained configuration is [1a2n+1bn+1[4 c(n+1)2 ]4 ]1 Consequently N(π)={m2/ m ≥ 1} Thus the above P System collects in its output membrane different values of m2, m>=1. The words “there’s plenty of room at the bottom” of Richard Feynman for Nano science, suit Membrane Computing also.
ACKNOWLEDGEMENT I record my sincere thanks to Prof. Kamala Krithivasan, Department of Computer Science and Engineering, IITM, Chennai for kindling my interest in Membrane Computing. Also I express my regards to Dr.Mrs.A.Pethalakshmi, Head, Department of Computer Science,M.V.M Govt Arts College(W), Dindigul who constantly encouraged all my research activities.
REFERENCES [10] P Systems webpage. [Online]: http://ppage.psystems.eu [11] Bernardini.F, “Membrane Systems for Molecular Computing and Biological Modelling”, Ph.D thesis, Department of Computer Science, University of Sheffield, Sheffield, 2005. [12] Krishna.S.N, “Languages of P Systems”, Ph.D thesis, Department of Mathematics, IIT Madras, Chennai, 2001. [13] Madhu.M, “Studies of P Systems as a model of cellular computing”, Ph.D thesis, Department of Computer Science and Engineering, IIT Madras, Chennai, India, 2003. [14] Oswald.M, “P Automata”, Ph.D thesis, Faculty of Computer Science TU Vienna, Vienna, Austria, 2003 . [15] Paun.Gh et al, DNA Computing New Computing Paradigms, Texts in Theoretical Computer Science, 1998. [16] Paun.Gh, “Computing with membranes”, Journal of Computer System Sciences, 2000. [17] Paun.Gh et.al.,“A guide to membrane Computing”, Theoretical Computer Science, 2002. [18] Paun.Gh, “Introduction to Membrane Computing”, Proc. Membrane Computing Workshop, 2004. [19] Paun.Gh, “Membrane Computing. An Introduction”, Springer, Berlin, 2002.
V. CONCLUSIONS Through this research paper, a preliminary review of membrane computing has been done. And membrane computing can be applied in computer science to simulate distributed computing environment and in solving NPComplete Class of problems. Also there are many variants [5] of this basic P system and some variants are used in modeling operating system. With the advent of Cloud Computing and its huge parallel processing power, P systems can facilitate development of a parallel processing language. Also in BioNano technology, Membrane Computing is used to model biological processes [2]. Still there are many avenues where membrane computing can be applied.
Velammal College of Engineering and Technology, Madurai
Page 388
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Entrustment Based Authentication Protocol For Mobile Systems R.Rajalakshmi1 R.S.Ponmagal2 1. Senior Lecturer, CSE, MAM College of Engineering 2. Lecturer, CSE, Anna University Tiruchirapalli
ABSTRACT - The aim of the paper is to describe the development of an efficient entrustment-based authentication protocol for mobile communication systems .A backward hash Chain Method is Proposed to use to ensure the authentication messages in offline authentication process in Visited Location Register (VLR), which includes the reduction in Computational Cost. The Entrustment-Based Protocol is claimed to provide non-repudiation in on-line and off-line authentication. The efficient Protocol preserves the security properties of the original scheme and reduces the computational cost in Mobile Station (MS).
authentication processes. This weakness is that any legal VLR can forge authentication messages without the help of the mobile user. The Entrustment protocol achieves non-repudiation in both on-line and off-line authentication processes, and thus avoids the weakness described above.
1. INTRODUCTION Mobile systems provide mobile users with global roaming services. To support that, numerous authentication approaches employ the public-key system to develop their protocols. A private authentication protocol to prevent the home location register(HLR) from eavesdropping on communications between the roaming station (MS) and the visited location register (VLR).Due to hardware limitations, MS cannot support heavy encryption and decryption, and therefore wastes a lot of time in exponential computations. Lee and Yeh [5] proposed a delegation based authentication protocol to solve the problems of data security, user privacy, computational loads and communicational efficiency in PCSs. Their protocol also adopted the public-key system to achieve the security requirements. To increase the communicational efficiency, and save authentication time, their protocol employs off-line authentication processes; such that VLR does not need to contact HLR frequently, and can rapidly re-authenticate MS. Therefore, compared with, the protocol of Lee and Yeh not only has a lower computational load for MS than related approaches such as GSM and MGSM [5][6][7], but also provides greater security. Though the protocol of Lee and Yeh exhibits non – repudiation in on-line authentication process, it still has a weakness in off-line
2. METHODOLOGY OF ENTRUST MENT BASED AUTHENTICATION PROTOCOL The enhanced protocol employs a backward hash chain to provide authentication and non- repudiation in off-line authentication processes. Setup: HLR and MS have their private/public key pairs (x, v) and (σ,K), respectively. The key pair (σ,K) is also stored in the MS’s SIM card. Besides, MS generates random number n1, pre-computes a hash chain h(1)(n1), h(2)(n1), . . . , h(n+1)(n1) and stores them in its database, where h(1)(n1) = h(n1) and h(i+1)(n1) =h(h(i)(n1)) for i = 1, 2, . . ., n. 2.2 On-line authentication Step 1. MS→VLR: K MS sends K to VLR. Step 2. VLR→MS: n2, IDV VLR generates a random number n2 and sends n2 and IDV to MS. Step 3. MS→VLR: r, s, K,N1, IDH, IDV MS generates a random number t, picks N1 =h(n+1)(n1) stored in its database, signs N1, n2 and IDV , and sends r, s,K,N1, IDH and IDV to VLR, where r = gt (mod p) and s = σ × h(N1_n2_IDV ) + t × r (mod q). Step 4.VLR→HLR: [N1_n2_K]KHV IDH, IDV If VLR successfully verifies the received messages by checking gs = vKK)h(N1_n2_IDV )rr (mod p), then sends [N1_n2_K]KHV , IDH and IDV to HLR. Otherwise, VLR rejects MS’s request. Step 5. HLR→VLR: [[N1, n3, IDV ]σ_n2_l_C1]KHV , IDH, IDV HLR decrypts [N1_n2_K]KHV and obtains K. If he successfully searches the corresponding σ in its database according to K, then he computes C1 = h(N1_n2_n3_σ) and l = N1, where n3 is a random number, and sends [[N1, n3, IDV ]σ_n2_l_C1]KHV , IDH and IDV to VLR. Step 6. VLR→MS: [N1, n3, IDV ]σ, IDVVLR decrypts [[N1, n3, IDV ]σ_n2_l_C1]KHV and obtains [N1, n3, IDV ]σ, n2, l and C1. Then he checks n2 and l, sets the
Velammal College of Engineering and Technology, Madurai
Page 389
Key Words: - Mobile station, security, computational cost, non- repudiation.
VLR,
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
current session key SK = C1 used by VLR and MS, and forwards [N1, n3, IDV ]σ and IDV to MS. Finally, MS decrypts [N1, n3, IDV ]σ, checks N1 and computes the current session key SK = C1. In the on-line authentication process, VLR connects with HLR when MS accesses the network through VLR and demands authentication. HLR also leaves secure authentication tokens with MS and VLR in this process. The off-line authentication process means that VLR need not contact HLR frequently, and can rapidly authenticate MS by verifying the secure token when MS accesses the network through VLR again. A newly token used for next authentication is simultaneously generated by using the old token. 2.3 Off-line authentication MS→VLR: [h(n−i+1)(n1)]Ci MS picks h(n−i+1)(n1) stored in his database and sends [h(n−i+1)(n1)]Ci to VLR for i = 1, 2,…., n, Where a predefined constant n is the limited times of off-line authentications, On receiving the authentication message from MS, VLR checks whether h(h(n−i+1)(n1)) and l are equal, updates l = h(n−i+1)(n1) and computes the session key Ci+1 = h(l,Ci). He also updates the count i = i + 1 and checks i ≤ n. For example, let n = 5. After the on-line authentication, MS stores h(4)(n1), h(3)(n1), . . . , h(1)(n1) in his/her database; VLR obtains the authentication token l = h(5)(n1); and C1 is the session key shared between MS and VLR. Subsequently, in the first off-line authentication, on receiving the authentication message [h(4)(n1)]C1 from MS, VLR checks h(h(4)(n1)) =?l and updates l = h(4)(n1) for the second offline authentication. Finally, MS and VLR have the common session key C2 = h(l,C1). Accordingly, MS and VLR are able to use the same method to execute the subsequent offline
illustrates the concept of delegation. In a business corporation, the manager uses his private key to sign a document and his staff can verify the document based on his public key. If the manager cannot sign a document because the is away on business, he can delegate his signature authority to his trustworthy assistant to sign the document without giving the assistant his private key. His staff verifies that the document is still based on his public key. This authorized signature technique is called proxy signature. This new type of digital signature gave us the inspiration for our model. The assistant is authorized to sign the document when the manager is absent, but the staff can still use the manager’s public key to verify the document. This implies that, even if the staff can distinguish the signature of the assistant from that of the manager, the staff cannot know the real identity of the assistant. The manager cannot deny the signature if a dispute arises. Of course, the manager should have the ability to identify a dishonest assistant. The previous example can be applied to our model. HLR gives MS the power to sign and VLR can verify the signature based on the public key of HLR. VLR can only verify the legality of MS but it does not know the real identity of MS. The model can provide user privacy and non_ repudiation features, and key management is easier than in the pure public-key system model since only the public key of HLR should be managed. Fig. 2 illustrates the concept of our model.
authentications.
Fig 1. Communication systems 3. CONCEPT OF DELEGATION In order to introduce our scheme, it is necessary to briefly describe the concept of delegation. Our method is inspired by the Proxy signature, which is the delegation of the power to sign messages. The following example
Velammal College of Engineering and Technology, Madurai
3.1. User Identity Privacy GSM, MGSM and public-key based protocols provide weak user identity privacy since MS must deliver his real identity to the network for authentication. In our protocol, the real identity of MS is never transmitted over the entire network for authentication purposes. Because we use pseudonym generated by HLR in the registration phase to represent the identity of MS in the network, no one except HLR can obtain any information about the identity of MS. Even VLR can only verify the legality of MS based on the public key of HLR, nothing about the identity of MS but only implies that MS is authorized by HLR. Hence, our scheme provides stronger user identity privacy than GSM, MGSM and public-key based protocols. 3.2. Non - repudiation GSM and MGSM are based on the secret-key system, so they cannot provide the feature of non - repudiation. No doubt, public-key based systems can greatly benefit by
Page 390
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the non - repudiation feature of the public-key cryptosystem. In our scheme, each MS gets a different pair key from HLR in the registration phase. The key implies the authorization from HLR. This authorization makes VLR transfer his trust in HLR to the requested legal pseudonym MS. Because only HLR has the ability to authorize MS to sign on his behalf, HLR cannot deny this in the event a disputation occurs. Of course, HLR has the ability to identify the misused MS. Thus, our scheme can also provide the feature of non - repudiation. 3.3. Mutual Authentication between MS and VLR GSM and MGSM only provide the mechanism for VLR to authenticate MS, and the public-key based protocols can provide mutual authentication services. In our scheme, it is easy for VLR to authenticate MS by verifying the proxy signature made by MS using the proxy key authorized by HLR. If MS is authenticated by VLR. On the other hand, MS can authenticate VLR by decrypting the message received in Step 6) of the on-line authentication phase to get and checking whether is the same as what he sent to VLR in Step 3) of on-line authentication phase. Because HLR is trustworthy, only the legal VLR can decrypt the message received from HLR to get the correct. There is no way for an attacker to pretend to be a legal MS or VLR. Besides, without knowing the secret keys and, impersonating HLR is impossible. Thus, our protocol can provide mutual authentication service between MS and VLR. Furthermore, MS gets a proxy pair key from HLR over a secure channel in the registration phase. The relation between the pair key and the corresponding real identity of the MS are protected in a secure database located in the HLR. In our scheme, MS uses the proxy key to sign and VLR is responsible to verify the MS according to (3). If a MS-A acquires MS-B’s pseudonym and impersonates B to ask authentication request, (3) cannot hold since MS A has no idea about B’s proxy key. Therefore, this impersonating attack cannot succeed. In the off-line authentication process, MS generates the message and sends it to VLR. VLR decrypts the message to get. It is very difficult to compute according to, since is a one way hash function which is relatively easy to compute, but significantly harder to reverse [14]. If any attacker tries to replay this message to pass the authentication process, he cannot succeed since VLR will find out that the value does not equal to which is updated to. If any attacker tries to forge this message to pass the authentication process, he cannot succeed since the message is encrypted by a session key which is different from time to time and only known to MS and VLR. For security considerations, it is not reasonable to do off-line authentication all the time. Hence, a predefined constant should be set to a reasonable constraint on the times to do off-line authentication. 3.4. Key Management
Velammal College of Engineering and Technology, Madurai
In public-key-system-based protocols, the verification of MS is based on the public key of MS. However, it is not easy to do so in practice. Because there are many mobile users in the system, the complexity of the public-key infrastructure (PKI) will be introduced into such protocols. In our protocol, HLR authorizes MS to sign the message and VLR is merely needed to verify MS based on the public key of HLR. The number of HLR is much less than that of MS, so the complexity of PKI is dramatically reduced. Besides, theoretically, the key of HLR must be more strictly defined and protected than that of MS, and it should be a long-term key that can be used without being frequently updated. Key management will become easier than in the protocols based on public-key systems. The key management of GSM and MGSM is also easy since secret key is a long-term key kept permanently in the SIM card. If HLR has to change its public/private key pair for some security reason, he should generate a new proxy key pair for each user and send the key to the corresponding user securely. This situation is the same as when CA changes its public/private key pair [16]; the old certificates, which CA signed before, should be collected back and destroyed, and the new certificates must be generated for the users. There seems to be no good solution to avoid this ugly situation. However, compared to the original public-key-based protocols, the key management of our protocol is much easier since the number of HLR’s public key is much less than that of MS’s in public-key-system-based protocols and the public key of HLR is not necessarily updated frequently.
4. ANALYSIS BETWEEN EXISTING SYSTEM AND PROPOSED SYSTEM Previous Scheme: Token i and tokeni+1 are independent.HN can forge token i. Have not non-repudiation Charge Problem:Mobile users deny has used services and refuse to pay. Overcharge mobile users for services that he did not request. Proposed Scheme: All token i are chained by backward hash-chain and are decided by MS. HN cannot forge token I;Have non-repudiation ;PreCompute and reduce the computational cost in MS. 5. CONCLUSION The weakness of the delegation based authentication protocol which cannot exhibit non-repudiation in offline authentication processes. The proposed Entrustment authentication protocol provides Online and Offline authentication In the Entrustment based authentication protocol; MS uses a backward hash chain to guarantee
Page 391
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
that no-one can forge the authentication messages to trick HLR and provides security.
REFERENCES [1] H.-Y. Lin, “Security and authentication in PCS," Comput. Elect. Eng.,vol. 25, no. 4, pp. 225-248, 1999. [2] A. Merotra and L. Golding, “Mobility and security management in the GSM system and some proposed future improvements,” Proc. IEEE, vol. 86, pp. 14801496, July 1998. [3] M. Long, C.-H. Wu, and J. D. Irwin, “Localised authentication for internet work roaming across wireless LANs,” IEE Proc. Commun., vol. 151,no. 5, pp. 496500, Oct. 2004. [4] T.-F. Lee, C.-C. Chang, and T. Hwang, “Private authentication techniques for the global mobility network,” Wireless Personal Commun. vol. 35, no. 4, pp. 329-336, Dec. 2005 [5] W.-B. Lee and C.-K. Yeh, “A new delegation-based authentication protocol for use in portable communication systems,” IEEE Trans. Wireless Commun., vol. 4, no. 1, pp. 57-64, Jan. 2005. [6] M. Rahnema, “Overview of the GSM system and protocol architecture," IEEE Commun. Mag., pp. 92100, Apr. 1993.
Velammal College of Engineering and Technology, Madurai
Page 392
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Modeling A Frequency Selective Wall For Indoor Wireless Environment Mrs. K.Suganya1 , Dr.N.Suresh Kumar2 , P.Senthil Kumar3, 1 [email protected] 1. Assistant Professor III, Velammal College of Engg and Tech, Madurai 2.Principal, Velammal College of Engg and Tech, Madurai 3. Research Scalor , Bharath University Chennai.
ABSTRACT: People today are demanding a communications infrastructure that is both secure and mobile. Meeting both these requirements cause considerable challenge. Currently, mobile communication networks and systems are designed on the basis of detailed analysis of RF coverage and capacity requirements. Security and privacy issues can be addressed through good design, but ‘eavesdropping’ remains a real vulnerability An emerging technology named FSS (frequency selective surface) is increasingly being proposed as an answer to the deployment of secure wireless systems for indoor wireless environments, taking advantage of innovative techniques in building design and the use of attenuating materials. FSS’s could be deployed to allow certain frequencies to propagate into a room, while reflecting other frequencies. In this paper we analyzed a Frequency Selective Surface (FSS) attached onto existing common construction material at certain distance that transforms it into frequency selective filtering wall which isolates indoor wireless system from external interference.
INTRODUCTION The interference between the co-existing indoor wireless systems is becoming an important issue. Unwanted interference not only degrades system performance but also compromise security. Interference mitigation can be achieved using advance signal processing or antenna design but in indoor environments, a simpler and more effective approach may be to physically modify the propagation environment. More specifically, a building wall might be transformed into frequency selective (FS) filter which filters
cost FSwall can be created by covering a building wall with a custom designed frequency selective surface (FSS). The custom designed FSS
discussed here is constructed using conducting aluminum foil square loops, with loop dimensions shown fig.1. A bandstop response can be obtained with attenuation of about 50dB greater than that of an uncovered wall and the resonant frequency. This could permit a substantial interference reduction and thus much better signal to interference ratio (SIR), and accordingly a better system performance. However [1] also indicates that a response of a FS structure is highly dependent on substrate (i.e., the wall board) and the illumination angle.
I.
out undesired interference (e.g., 2.4GHz WLAN signals), but still allows other desired signals (e.g., cellular telephone, TV & radio broadcasting) to pass through as usual. This is the ultimate research reported in this paper [2]. In [1] it has been demonstrated that a simple and low
Velammal College of Engineering and Technology, Madurai
BFSS BWALL CFSS DFSS DWALL
AAIR
BAIR
CAIR
DAIR
AFSS AWALL CWALL
Fig.1. the matrix cascading technique, as applied to the FSS problem in this paper. Each layer is characterized by an ABCD matrix. The angle sensitivity is, to some extent a function of the FSS pattern and is under the control of the FSS design engineer. However the properties of building walls are not. For practical reason, it would be ideal if the wall and FSS cover could be modeled independently, and cascaded to
Page 393
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
obtain the combined performance. Modeling a FSS with a simple equivalent circuit has been investigated in [4].
II. FACTORS INFLUENCING THE FSS PERFORMANCE AND DESIGN The performance and behavior of the FSS filters depends on the following factors. The conductivity of the FSS conductor. The geometry of the FSS element (shape, with of conductive strip lines, proximity of conductive strip lines, thickness of conductor). The permittivity of the FSS substrate. The period of the FSS array. The number of FSS arrays when these are employed in a cascade. The electrical distance between the FSS arrays in cascade configurations. The choice of element types in hybrid FSS configuration. The finite number of periods and the metallic frames surrounding the FSS window. III. EQUIVALENT CIRCUIT METHOD Equivalent circuit modeling is the preferred modeling technique used in this research. Although it can only be used for linear polarization and is limited to certain element shapes, it offers the designer a quick and simple way to evaluate possible FSS solutions without intensive computing. EC modeling is ideal for use as an initial tool in FSS design. Each layer of square loop FSS can be represented by an equivalent circuit with an inductive component in series with a capacitive component’s. The FSS is modeled by separating the square loop elements into Vertical and horizontal strips (gratings). For TE wave incidence, the vertical strips Contribute to the inductive impedance in the equivalent circuit, where the value L is calculated according to equation 1. XL/Z0=ωL=d/p*F (p, 2s,λ,θ)………. ...(1) Similarly, the horizontal gratings correspond to the capacitive component and the C value is calculated according to equation 2. BC/ZO=ωC=d/p*4*F (p, g,λ,θ)*εeff……..( 2) p
s
d
g
c c
L
Equivalent circuit model for the square loop FSS. The field F is given by, F(p,w,λ,θ)=p/λ*cosθ[ln(cosec(πw/2p))+ G (p, w,λ,θ)] G (p, w,λ,θ)] = {0.5*(1-β²)²[(1-(β²)/4) (A+ + A_)+4β² A+ A_]} {[1(β²/4)] +β²[(1+ (β²/2) (β^4/8)] (A++A_) + 2β^6 A+ A_} A± ={1/[1±(2psin θ/λ)-(p cos θ/λ) ²]ˆ½}-1 β=sin (ωπ/2p) It was suggested that with sufficient spacing (comparable to a wavelength), the mutual influence between a wall and its FSS cover can be neglected. This paper considers the minimum spacing required between the FSS and the wall in order that they may be modeled as acting essentially independently.
IV. SIMULATION AND CASCADING MATRICES The measurements presented in this paper were simulated at S band frequency. The frequency responses of FS structure were obtained at various illumination angles. The transmission (S21) and reflection (S11) coefficient of both the FSS cover and existing wall are obtained. Simulations were preceded with different air spacing between the FSS and wall surface (0, 10 and 20mm). The FSS and wall should act sufficiently independent for simple wave propagation. V. INTERACTION BETWEEN FSS AND WALL. The simulations of FS wall performance at 0° signal incidence with air spacing is shown in fig.3. When the air spacing is 0mm fr for FS wall is at 2GHz from the designed fr at 2.4GHz which suggest that the overall response was significantly influenced by the wall structure conversely when the air is increased to 10mm or 20mm, fr did not shift from 10GHz. Therefore, 10mm air spacing provides sufficient independence to ensure insignificant interaction between FSS and the wall. VI. RESULTS The simulation carried with 10mm air spacing at 0º incidence is obtained at 2.15 GHz with 50dB attenuation.
Fig .2. Square loop FSS and the appropriate equivalent circuit.
Velammal College of Engineering and Technology, Madurai
Page 394
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.6. FSS with 0degree incidence and 10mm air spacing Fig.3. FSS alone in air At other incident angles discrepancy in fr value occurs. This suggests that at 0mm intervening air spacing, the FSS response is greatly influenced by the wall. The FSS and the wall interact with each other in a complicated manner. Such interaction detunes the performance of FSS and prevents the FS wall from being modeled by cascading the effects of the FSS and the wall individually. By increasing the air spacing to 10mm the response is obtained for the predicted frequency. At 0º incidence the
resonance occurs at 2.4GHz. In other words, the FSS characteristics are preserved, with the FSS remaining unaffected by the difficult-to-model dielectric properties of the wall. This indicates that the interaction between the FSS and the wall is almost insignificant with 10mm air spacing. More specifically, the air spacing allows the FSS and the wall to be treated independently, and to be represented by the FSS and the wall matrices. Accordingly, the proposed matrix cascading technique can be adopted for modeling such a FS wall. Increasing the air spacing to 20mm, yields no additional performance benefit over a 10mm spacing. The FSS matrix is multiplied directly with the wall matrix. The focus has been placed on determining the minimum air spacing required, so that the FSS and the wall can act independently without influencing the ‘fundamental’ response of each other. Provided that the wall is not highly reflective, the impedance transformation caused by the 10mm air section is regarded as insignificant in the investigation considered here.
VII.
Fig.4 FSS with 0degree incidence and 0mm air spacing
Fig.5. FSS with 0degree incidence and 5mm air spacing
Velammal College of Engineering and Technology, Madurai
APPLICATION OF FSS. Selective shielding of the electromagnetic interference from high power microwave heating machines adjacent to wireless communication base-stations. Selective shielding of frequencies of communication in sensitive areas (military installations, airport, police etc.). Protection from harmful electromagnetic radiation especially in the 2-3GHz band arising externally (wireless communication base stations) or internally (microwave ovens) in the domestic environment, schools, hospitals etc. Control of radiation at unlicensed frequency bands (e.g. Bluetooth applications,2.45GHz). Pico cellular wireless communications in office environments such as the Personal Handy-phone System in offices whereas to improve efficiency each room needs to prevent leakage of radio waves into another room. This implies that windows, floor and ceiling need to be shielded. Isolation of unwanted
Page 395
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
radiation. FSS windows can be incorporate in trains to prevent mobile phone frequencies. Note: that in the above applications one wishes to prevent certain frequency bands of Electromagnetic radiation to be transmitted whereas others are required to pass (frequencies related to emergency services for example). Hence the use of a broadband shielding material is not an option.
Fig8. 5mm 45deg
VIII. CONCLUSION AND FUTURE SCOPE It is concluded that 10mm air spacing is sufficient to provide independence between the FSS cover and the wall. This allows engineers to focus on designing a FSS with desired response without knowledge of the specific properties of each building/office wall. The wall material may affect the absolute attenuation but not the frequency selectivity. However, the spacing required could be reduced down to almost 0mm with an appropriate FSS design and careful choice of dielectric substrate to provide the spacing rather than air. The future works are concerned with simulating the FSS surface with various structures.
REFERENCES
Fig7. 5mm 15deg
Fig7. 5mm 30deg
Velammal College of Engineering and Technology, Madurai
[1] G.H.H.Sung, K.W. Sowerby and A.G.Williamson, “The impact of frequency selective surfaces applied to standard wall construction materials.” Proc.IEEE 2004 Antennas Propag. [2] A.Newbold, “Designing buildings for the wireless age.” Compat. Contn. Eng. J., vol.15. [3] G.H.H.Sung, K.W. Sowerby, “Modeling a lowcost frequency selective wall for wireless friendly indoor environments”.IEEE Antennas and wireless propagation letters.vol.5.2006. [4] G.H.H.Sung, K.W. Sowerby and A.G.Williamson, “Equivalent circuit modeling of a frequency selective plasterboard wall.” In Proc.IEEE 2005 Antennas Propag.Symp., vol.4A. 2005. [5] D.M.Pozar, Microwave Engineering. AdisonWesley.1990. [6] “An investigation into the feasibility of designing frequency selective windows employing periodic structure” C.Mias, C.Oswald and C.Tsakonas- The Nottingham Trent University.
Page 396
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Multi-Agent Based Personalized e-Learning Environment 13 2
T. Vengattaraman1, A. Ramalingam2, P. Dhavachelvan3, R.Baskaran4, Department of Computer Science & Engineering, Anna University, Chennai, India.
Department of Master in Computer Applications, Sri Manakula Vinayagar Engineering College, Puducherry, India. 4
Department of Computer Science, Pondicherry University, Puducherry, India.
Abstract— Personalizing e-learning services becomes more
important and it is needed for learners when e-learning takes place in an open and dynamic environment. It is of great importance to provide personalized e-learning services which can automatically adapt to the interests and levels of learners. In this paper, we propose the personalization of e-learning services with multiple feedback measures and adapting at individual level also at group level using multi agent architecture. Multiple feedback measures are collected from the user and stored in user profile. The feedbacks are adapted using an adaptation algorithm and thereby providing personalized e-learning content to the individual user and group of users. VI. INTRODUCTION E-learning is being viewed as an important activity in the field of distance and continuing education. It offers obvious advantages for e-learners making access to educational resource at any time or at any place. With the development of e-learning technologies, learners can be provided more effective learning environment to optimize their learning. Web personalization is an advanced stage in the e-learning systems evolution and it alleviates the burden of information overhead by tailoring the information presented based on an individual user’s needs. The learners have different learning styles, objectives and preferences, fact that leads to different efficiency and effectiveness from individual to individual [1]. Hence all the learners can't be treated in a uniform way. So there is a need for personalized system which can automatically adapt to the interests and levels of learners. A promising approach towards the personalization of e-learning services is user profiling. User profile includes interests, levels and learning patterns can be assessed during the learning process. Based upon the profile, personalized learning resource could be generated to match the individual preferences and levels. User profiling finds its importance in recommendation systems. Typically a recommender system compares the user profile to some reference characteristics to find similarity and differences and thereby provide suggestions to the users. For personalized search, user profiles can be
Velammal College of Engineering and Technology, Madurai
used for re-ranking the results returned to the user. Comparing to recommender systems and personalized search, user profile is important and it is needed in elearning as it is a continuous process. Feedback is the information the user receives from the system as the result of his or her action. It occurs in elearning not only in the assessment process, but can be provided to a student during navigation through learning materials, communication and collaboration with other students. Even the alerts and remainders that often appear in the e-learning system can be considered as feedback [2]. Mostly e-learning systems are personalized using either implicit or explicit feedbacks. In Implicit profiling, the user's behavior is tracked automatically by the system whereas in explicit profiling each user is asked to fill in a form when visiting the website. In our approach we are not following either of these methods. Instead we combine multiple feedback measures to get accurate and more complete user profiles for personalization. The learners with common interests and levels can be grouped, and multiple feedbacks of one person can serve as the guideline for information delivery to the other members within the same group. By adapting multiple feedbacks, personalized effective e-learning content can be delivered at the individual level and at the group level. Individual adaptation means that feedback is adapted to each student based on their characteristics. Group adaptation means adapting the feedback to common characteristics of a group of users. RELATED WORK AND LITERATURE SURVEY VII. Many approaches uses semantic Web or Web service technologies to provide dynamic as well as personalized support for learning objectives. Few approaches [3][4] are concerned with bridging learning contexts and resources by introducing semantic learning context descriptions. This allows the adaptation to different contexts based on reasoning over the provided context ontology, but does not provide solutions for building complex adaptive learning applications by reusing distributed learning functionalities. Some proposals [5] follows the idea of using a dedicated personalization web service which makes use of semantic learning object descriptions to identify and provide
Page 397
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
appropriate learning content. Integration of several distributed learning services or service allocation at runtime is not within the scope of this approach. The related research on a Personal Reader Framework (PRF) introduced in [6] and [7] allows mediation between different services based on a”connector service”. The work described in [8] utilize semantic web as well as web service technologies to enable adaptation to different learning contexts by introducing a matching mechanism to map between a context and available learning data. However, neither it considers approaches for automatic service discovery nor it is based on common standards. Hence, the reuse and automatic allocation of a variety of services or the mediation between different metadata standards is not supported [9]. Whereas the majority of the described approaches enable context-adaptation based on runtime allocation of learning data, all of them do not enable the automatic allocation of learning functionalities neither it does enable the integration of new functionalities based on open standards. Nevertheless, all approaches do not envisage mappings between different learning metadata standards to enable interoperability not only between learning contexts but also between platforms and metadata standards. VIII.
APPROACHES TO PERSONALIZATION OF ELEARNING SERVICES
Many approaches to personalization of e-learning services consider re-authoring of existing learning material. Mechanisms that facilitate the reuse of learning content increase the life span of the e-learning content but they are too expensive. To reduce the implementation overhead of the designer it is imperative to facilitate the maximum reuse of learning resources. The fulfilment of the individual needs of each user – learning personalization, educational content re-usability on large scale – content reuse, assurance of the communication between e-learning systems as well as with other human resources management systems – interoperability are the main objectives of the researches in this domain. The following subsections explain about the research works carried out by different authors in personalization of e-learning systems. A. Personalized E-learning system Blochl et al [10], proposed an adaptive learning system which can incorporate psychological aspects of learning process into the user profile to deliver individualized learning resource. The user profile is placed in multidimensional space with three stages of the semantic decisions: cognitive style, skills and user type. However,
Velammal College of Engineering and Technology, Madurai
both the means to acquire user's feedback and the algorithms to update user profile have not been addressed in the presentation. SPERO is a [12] based on the IEEE Learning Technology Systems Architecture (LTSA). It could provide different contents for the foreign language learners according their interests and levels. The problem of SPERO system is that it is largely using questionnaires and e-surveys to build user profiles, which costs the users too much extra work. B. Recommendation Systems User profiling is the key process of recommendation systems, which collect user feedback for items in a given domain and assess user profiles in terms of similarities and differences to determine what to recommend. Depending on underline technique, recommendation systems can be divided into collaborative filtering-based content-based [11] and hybrid approaches. Classified by means to acquire feedback, they can be categorized as explicit rating, implicit rating and no rating needed [11] systems. In fact, user's feedbacks are so important that only very few content-based recommendation systems require neither explicit rating nor implicit rating. For example, Surflen is a recommendation system using data mining techniques to assess the association rules on web pages through user’s browsing history without the feedbacks. However, it's hard to find user's exact interests just based on the browsing history, since it always happens that users open a page they don't like or just by mistake. This problem becomes even more severe in the situation that the system is sparsely used. Many methods of user profiling are heavily depending on the user feedbacks to construct user profiles. The feedback can be assessed explicitly by rating, or implicitly by the user behaviors such as print and save. The major drawback of existing e-learning platform is personalization with implicit feedbacks and explicit feedbacks. The proposed approach is trying to improve the personalized content delivery to the e-learners. C. Feedback Adaptation Individual adaptation means that feedback is adapted to each student and his/her individual (combination of) characteristics [12]. For example, the individual characteristics could include the user’s knowledge of the subject being studied (knowledge of the main concepts, formulas, etc) and the number of mistakes the user makes during the testing. The time and the way of feedback presentation could be personalized to these individual characteristics. For example, if the user has started to make some mistakes more often the system can present the feedback more often and include more detailed explanations in the feedback (compared to the feedback given to a user who only occasionally makes a mistake). The information that is presented in the feedback can be
Page 398
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
also personalized by relating to concepts that are already mastered by the user. Group adaptation supposes that the system adapts the feedback to common characteristics of a group of users. For example, they can be grouped according to their learning style. Immediate feedback could be presented in a brief form for active users, while detailed elaborated feedback could be presented to reflective learners. Another example of the group adaptation of feedback is personalization of the feedback to students who has passed the same courses before. The feedback could include references to the previous course (if the student has passed it) or the detailed explanation (in the case the material might be unknown to the user as s/he has not taken the course). IX. PROPOSED APPROACH In our approach, we propose the personalization of elearning services with multiple feedbacks and adapting feedbacks at individual level and group level using multiple agents. Knowledge about a user inferred from user interactions with the e-learning systems is stored in user profile and they are used to adapt offered e-learning resources and guide a learner through them. The learning material is accessible as other information sources on the web. Hypertext interlinks related pieces of information and allows the user to browse through the information space. The links are provided either explicitly, encoded by authors of the pages, or they generated automatically based on the test. Generating links automatically based on user profiles is an attractive option. However, it creates challenges as well. The data which are collected about the user can be classified into two types: static data and dynamic data. The static data are those which are not altered during the student-system interaction. The dynamic data are that changes according to the student learning progress and with the system interaction. The static data comprises five different parts. They are personal, personality, cognitive, pedagogical and preference data. Each one is an aggregation of student characteristics, which are not usually changed during an elearning session. Personal data comprises the biographical information about the student, and can be easily obtained from the course enrolment form. These information are: Student name • Student’s professional activities • • List of degree and qualifications The personality data models the student characteristics that represent the type of person the student is. These
Velammal College of Engineering and Technology, Madurai
characteristics can be inferred from personality tests. The attributes of personality data are: Personality type • • Concentration skills –based on the average time spent in the learning contents. Collaborative work skills based on the • participation in group works Relational skills based on the interactions with • students and teacher. The cognitive data models the student characteristics that represent the type of cognition the student possess. These characteristics can e inferred from cognitive tests. Based on the cognitive data, the contents can be tailored to meet the student needs. The cognitive data are Cognitive style • • Level of experience the student possesses in using the e-learning system • Student experience in using computers The pedagogical data defines the student characteristics that deal directly with the learning activity. This data intends to model the student’s behavior in learning situations, comprising two personal properties Learning style • • Learning approach The preference data stores a set of student preferences regarding the system customization. Most of the preferences are gathered from the student, but some of them are defined by the system administrator. The attributes are: Preferred presentation format • Preferred language for content display • • Web design personalization • Video speed Sound volume • The dynamic data comprises two sets of data. They are performance data and the student knowledge data. The performance data comprises level of motivation, confidence for e-learning, ability to formalize and comprehend course concepts, global performance level by the student in the course, level of effort spend by the student in the course, and portfolio that stores all the results obtained by the student in the current course. Data is constantly being gathered in order to keep an updated data and this data can be collected from the student-system interaction. The student knowledge data comprises of the concepts referred in the course, student’s progress regarding the knowledge concepts describes the knowledge that the student must possess for the current course and must possess until the end of the course. This set of data also gathers information about the student’s progress during the course sessions. This data is also gathered from the studentsystem interaction. User profile which includes user preferences contains three versions of profile data. They are:
Page 399
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
• Initial login profile • Learner’s skill profile Learner’s type profile • Initial profile formed on first time login and based on the query answered by the learner. It can be dynamically changed over a period of time, because of every action of learner is monitored and updated in learner profile. The feedbacks which are stored in user profile can be adapted to provide personalized e-learning service to the users. For adapting feedback, a feedback adaptation algorithm is proposed. X. IMPLEMENTATION The development of e-learning platform can be done with java. Basically an e-learning platform consists of login validation service module, registration service module, etest service module. Apart from this, in our project we are adapting multiple feedbacks for providing personalized elearning content to the individual user and group of users. For this we are going to propose an algorithm for adaptation of feedbacks. Our proposed system consists of profiler service, learner service, monitoring service, feedback service, content service, adaptation service. The fig. 1 depicts the system architecture using multiple agents. C. Learner service The learner service consists of login service, registration service, pre assessment test service. 1) The login service: The login service module consists of on-line registration, validation. Before taking online course, the student has to register and do the validation. The validation part consists of checking the values with the database and validating it. After validation, the student can take online course. 2) The registration service: The registration service module consists of registering the form if the user is a new user. Usually the user has to get registered with the website before taking the online course. 3) The pre assessment service: The pre assessment test module provides the on line test to the e-learners in order to check for the level of the learners in the subjects. The test marks will be calculated and his efficiency will be rated based on the marks. D. The profile service The profile service consists of two tasks. They are assessment of expertise level, providing guideline for delivery. The levels of learners are determined by the average preferences. The webpage user has read on any topic which could have different levels in terms of beginning, intermediate and advanced. The information
Velammal College of Engineering and Technology, Madurai
delivery to individual user and group of users is done by adaptation algorithm. E. Feedback extraction The feedbacks are extracted to make a final assessment of user preference. The feedbacks which are employed in system including reading time, scroll, print/save and relational index. Reading time: return 1 if user read pg longer than • Φt, where Φt is a predefined threshold; 0 otherwise. Scroll: return 1 if the number of user scrolls • (either mouse or keyboard pageup/pagedown) on pg is greater than pies, where Φs is a predefined threshold; 0 otherwise. Print/save: return 1 if user prints/saves pg; 0 • otherwise.
System Interface (GUI)
Profiler service (Agent)
Learner Service (Agent)
Content Service (Agent)
Monitoring service (Agent) Profile Database
Feedback service (Agent)
Adaptation service (Agent)
• elational index: return 1 if keywords of pg appear in user's chatting history ch more than Φt times, where Φt is a predefined threshold; 0 otherwise. Fig. 1 System Architecture
F. Feedback service Feedback service consists of collecting multiple feedbacks such reading time, no of scroll, no of print/save and relational index on chatting history and storing it in user profile.
Page 400
R
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
G. Adaptation service The feedbacks which are maintained in user profile can be adapted and provided to the individual user and group of users. The adaptation service consists of adaptation algorithm which adapts the contents and provides the personalized content to the e-learners. H. Adaptation algorithm The web pages are organized by the topics which are structured in e-content ontology. Each topic is attached with several keywords. XI. EXPERIMENTAL RESULTS A prototype system has been implemented and multiple feedback measures are recorded for feedback extractor. From the feedback we obtained the learning performance curve over various iterations and the fig.2 shows the improvements. For system training purpose, we ask a group of students to do the following experiments: Step 1: Select a topic such as “Data Structures” • and “Object Oriented Programming”, let the students indicate their levels on it in terms of beginning, intermediate, or advanced. • Step 2: Evaluate the performance. • Step 3: Prepare proficiency profile for each student and store it in database. • Step 4: Customize the learning content based on the proficiency profile. Fig. 2
measures are collected from the user and stored in user profile. The feedbacks are adapted using an adaptation algorithm and thereby providing personalized e-learning content to the individual user and group of users. The experimental results show that there is a great improvement in the learning curve. REFERENCES [20] Zhang Yin, Zhunag Yueting, WU Jiangqin, “Personalized Multimedia Retrieval in CADAL Digital Library”, Advances in Multimedia Information Processing - PCM 2008, Volume 5353/2008, pp703-712, 2008. [21] Okkyung Choi, SangYong Han, “Personalization of Rule-based Web Services”, Sensors 2008, 8(4), pp-2424-2435, 2008. [22] Knight, C., Gašević, D., & Richards, G., “An Ontology-Based Framework for Bridging Learning Design and Learning Content”, Journal of Educational Technology & Society, 9 (1), pp-23-37, 2006. [23] M. Baldoni, C. Baroglio, V. Patti, and L. Torasso, “Using a rational agent in an adaptive web-based tutoring system”, In Proc. of the Workshop on Adaptive Systems for Web- Based Education, 2nd Int. Conf. on Adaptive Hypermedia and Adaptive Web-based Systems, pp- 43-55, Malaga, Spain, 2002. [24] M. Baldoni, C. Baroglio, I. Brunkhorst, N. Henze, E. Marengo and V. Patti, “A Personalization Service for Curriculum Planning”, ABIS 2006 - 14th Workshop on Adaptivity and User Modeling in Interactive Systems, Hildesheim, pp-17-20, October 9-11 2006. [25] Henze, N., “Personalized e-Learning in the Semantic Web. Extended version of 4”, International Journal of Emerging Technologies in Learning (iJET), Vol. 1, No. 1, pp-82-97, 2006. [26] Nicola Henze, Peter Dolog, and Wolfgang Nejdl, “Reasoning and Ontologies for Personalized E-Learning”, Educational Technology & Society, Vol. 7, Issue 4. pp-82-97, 2004. [27] Schmidt, A., Winterhalter, C., “User Context Aware Delivery of ELearning Material: Approach and Architecture”, Journal of Universal Computer Science (JUCS), vol.10, no.1, pp-28-36, January 2004. [28] Simon, B., Dolog., P., Miklós, Z., Olmedilla, D. and Sintek, M., “Conceptualising Smart Spaces for Learning. Journal of Interactive Media in Education”, Special Issue on the Educational Semantic Web, ISSN:1365-893X [http://wwwjime. open.ac.uk/2004/9], 2004. [29] Martin Blöchl, Hildegard Rumetshofer, Wolfram, “Individualized E-Learning Systems Enabled by a Semantically Determined Adaptation of Learning Fragments”, Proceedings of the 14th International Workshop on Database and Expert Systems Applications, pp-640, 2003. [30] Xiaobin Fu, Jay Budzik, Kristian J. Hammond, “Mining navigation history for recommendation”, Proceedings of the 5th international conference on Intelligent user interfaces, United States, pp-106 - 112, 2000. [31] Xin Li and S. K. Chang, “A Personalized E-Learning System Based on User Profile Constructed Using Information Fusion”, The eleventh International Conference on Distributed Multimedia Systems (DMS'05), Banff, Canada, pp. 109-114, Sep.2005.
Learning Improvement Graph
The experimental results show improved learning curve. We used scoring in the range of 1 to 5 for each student; the overall performance of the students improves in the subsequent iterations. CONCLUSION XII. This paper proposes the personalization of e-learning services with multiple feedback measures and adapting at individual level also at group level. Multiple feedback
Velammal College of Engineering and Technology, Madurai
Page 401
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Identification in the E-Health Information Systems Ales Zivkovic University of Maribor FERI Maribor [email protected] Abstract Fast development of health informatics promise to offer new services supported with information technology. Different stakeholders in the public health care systems are already using partially integrated solutions in order to exchange electronic health records. Patient identification is one of the key elements in the future electronic health systems. While technology provides several solutions for patient identification, processing of personal data related to health is strictly regulated. The paper addresses some of the most common issues in regard to privacy, discuss technical opportunitiesin the identification problem space and presents current identification solution in Slovenia.
1. Introduction The use of most information systems depends on user identification. The purpose of identification is representation of the user or another system to the information system being used. Based on the identification, the information system will allow or deny access to the system or one of its parts. With the integration of information systems, high availability of e-services and ubiquitous solutions, the importance of identification mechanisms and efficient authorization schemes is also becoming more and more important. The identification mechanisms should guarantee first level security to the information systems as well as the user, while the authorization prevents unauthorized access to the information, data and services. The costs of public health care schemes are substantially increasing and governments are calling for new strategies [1]. One of the solutions is open integrated electronic health system that is able to exchange the information between all stakeholders in the public health care system. The key stakeholders are:hospitals, medical centers, health centers, insurance companies, government, patients, employers, pharmacies, drugstores, pharmaceutical companies, health care professionals (i.e. doctors, nurses, pharmacists) and others. The central part of the electronic health system is electronic health record (EHR), the electronic document that contains the comprehensive data of the past and present physical and mental state of health of an individual. The EHR is highly sensitive in regard to personal data protection as well as confidentiality, safety and accuracy. The identification plays an important part in fulfilling these requirements. The three basic principles of identification are [2]:
Velammal College of Engineering and Technology, Madurai
Something you have - the user poses something • that is then used for identification since other people do not have this. For example EHR smart card, hardware key generator andpublic key certificate. Something you know - the identification element • is secret information that is known only to the user of the system. For example password, paraphrase and Personal Identification Number (PIN). Things you are - the principle is based on the • physical characteristics or behavior of the user that can be used for identification. For example voice recognition, fingerprint, biometric signature, palm and hand print recognition, iris and retina recognition. The most common principle in use today is "something you know" since passwords and PIN numbers are widely used for computer and system login while ATM machines and mobile phones require PIN identification. The principle is simple to implement, easy to use and inexpensive. The only disadvantage is password vulnerability. Therefore it is better to implement and use other identification mechanisms. The paper is organized as follows. The second section gives a brief overview of the standards and regulations in health informatics and identification. In section three, technical solutions for user identification are described and compared. Section four gives an overview of the recent identification solution in Slovenian public health system. Section five concludes the paper.
2. Standards and Regulations There are many standards, laws and directives related to protection of personal data, security, identification, health information systems and biometrics. The aim of this section is to point out some of the related standards and regulations and provide the overview of the legal framework in Slovenia. The ISO/DIS 21091:2008 Health informatics - Directory services for security, communications and identification of professionals and patients [3] standard is aimed at providing key requirements for secure communication of the health care professionals in conducting clinical and administrative functions. The standard extends the public key infrastructure with additional requirements for its efficient use within the health care transactions. The health care directory includes the information about valid certificates, certificate revocation lists (CRL), identification
Page 402
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of individual roles within the health care system, trust chains and secure communication over the public networks. The directory services are using the X.500 framework. The specification also defines the directory security managementframework where the following standards should be considered for protection of the directory infrastructures: ISO 22600-2, ISO/IEC TR 13335-1 and Control Objectives for Information and Related Technologies (COBIT) [4]. According to the Technical specification CEN/TR 15872 Health informatics - Guidelines on patient identification and cross-referencing of identities [5], the accuracy of identification and information is essential requirement within the healthcare information system. The individual needs to be uniquely identified in order to ensure appropriate and relevant care. Currently the patient may have several identifiers (i.e. domestic health care identifier, international insurance identifier while traveling, temporary identifiers for various medical services outside the public health care system) corresponding to different geographical locations, different health care organizations and services. Consequently, different identifiers increase the risk of identification error and potentially compromise the patient's safety. The quality of identification ensures that health care professionals have access to patient information, facilitating closer coordination and continuity of care and improving medical service in terms of preventive as well as follow-up activities. In the technical report, the solution for providing continuity of medical care and patient data exchange that is based on the reliable consolidation of different identities in the patient identifier domain. It also provides solutions for cross-referencing of identities when the medical professionals needs to access all patients data that resides in different information systems managed by different health organizations. The main legal instrument of EU data protection law Directive 95/46/EC [6], defines health information as sensitive data. Sensitive data cannot be processed unless the subject of the data (the patient in case of the EHR) gives explicit consent, or another exemption applies. Article 8 of the Directive expressly provides that Member State laws may prohibit the processing of certain sensitive data, irrespective of the individual's consent. Without obtaining consent, organizations may only process the following data: Data necessary for exercising the organization's • obligations or rights related to employment. Data necessary to protect the vital interests of the • individual when the individual is physically or legally incapable of giving consent. Data necessary for the establishment, exercise or • defense of legal claims. Data necessary for the purposes of preventive • medicine, medical diagnosis, the provision of care or treatment or the management of health-care services, and where those data are processed by a health professional
Velammal College of Engineering and Technology, Madurai
subject to professional secrecy or by another person subject to an equivalent obligation of secrecy. Data that were clearly made public by the • individual itself. In addition, the processing of sensitive data generally requires prior approval from national data protection authorities. In Slovenia, the law requires the approval of the Information Commissioner. In Italy a detailed security policy document is required, and specific technical requirements must be met. In Spain, the processing of health-related data triggers a requirement for more rigorous security measures under Royal Decree 994/1999. After the Member States have implemented the exceptions differently and inconsistently, theEuropean Commission Report on the implementation of the Data Protection Directive (95/46/EC [6]) recognized the problem and the Data Protection Working Party, an independent European advisory body on data protection and privacy, issued the Working Document on the processing of personal data relating to health in electronic health records (EHR).The document provides guidelines on the interpretation of the applicable data protection legal framework for EHR systems, presents the general principles and recommendations on eleven topics namely respecting self determination, identification and authentication of patients and health care professionals, authorization for accessing EHR in order to read and write in EHR, use of EHR for other purposes, organizational structure of an EHR system, categories of data stored in EHR and modes of their presentation, international transfer of medical records, data security, transparency, liability issues and control mechanisms for processing data in EHR. In 2004 Slovenia adopted the Personal Data Protection Act [7] that determines the rights, responsibilities, principles and measures to prevent unconstitutional, unlawful and unjustified encroachments on the privacy and dignity of an individualin the processing of personal data. The enforcement of the law is under the supervision of the Information Commissioner of Republic of Slovenia. In 2008 the commissioner issued two important guidelines [8]: Guidelines for safeguarding the personal data in 1. the Hospital Information Systems (HIS) and Guidelines regarding the use of biometric 2. technologies. In the first guideline the commissioner points out that the logging of user interaction could be at different levels: (1) change log, (2) data access log and (3) full audit trail. While the law does not require the audit trail, technically it is possible to implement it within the HIS. The second level is mandatory for sensitive data and must be implemented regardless of the circumstances. The use of group login accounts is not advisable and should be used only in special cases (i.e. emergency team). However, for these exceptions additional measures should take place
Page 403
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
making clear who were the members of the emergency team in the particular time frame. The second guideline first explains the difference between identification and authentication. When implemented with biometric techniques, both approaches are treated equally and may be used only when it is impossible to fulfill the same aim with other means. The use must be regulated. In the public sector the biometric solutions may be used for entering working premises or work registration. In private sector, the use is allowed only when higher security measures are required for protecting people or trade secrets. It can be used only for company employees with prior written notice. Before introducing the biometric, the company must acquire approval from the government supervisory office. According to the Slovenian and EU laws the biometric data are always treated as sensitive personal data.
3. Technical Solutions Something you know (usernames and passwords, PIN) is the most widely used identification principle for accessing IT resources today. The administrator or the system itself, using the algorithm, sets the username and password for the new user. The user can later partially (only password) or fully (username and password) change it. This identification principle has several drawbacks: Reliability of identification - when the system • receives the correct username and password it only knows that the identification data matches the data in the database. It can not guarantee that the person that entered the data to the system is the right person. The data could be stolen or the user gives it to someone else in order to perform the task for him. Insufficient control - even if the company • introduced the password policy, there are many problems when used in practice (i.e. the password length, complexity, the password change frequency, storing the passwords). The compromised password detection - usually it • is quite difficult to detect that the password was compromised. The attacker could use the stolen data for days, months or even years before anybody founds out. The compensation control would be to change the password regularly to prevent long time exposure. The system could also try to recognize unusual logins (after the premises were closed, weekends). 3.1 Biometric identification The biometric identification could be divided into two subgroups: • Physiological biometry - includes the methods that use physiological properties of the person like fingerprints, iris and retina recognition and palm geometry. The variability of the properties in this group is very low. Behavioral biometry - uses person's behavioral • patterns. In this group are voice, typing and handwriting. The biggest problem of the methods in this group is high
Velammal College of Engineering and Technology, Madurai
variability of the property being measured for identification. The recognition accuracy is also lower in comparison to the physiological biometry. The variability could be due to the changes a person is going through (i.e. illness, stress). The advantage is better user acceptance. The biometric identification has several advantages: the biometric properties are unique, • • the property can not be transferred to another person, it can not be forgotten, lost or stolen, • • it is difficult to copy or forge, • it can be used when person is unconscious and it is hard for someone to hide or change the • property. Therefore, the use of biometric identification could be more appropriate in healthcare systems for some users (for example children, senior users and people with disabilities) and special use cases (for example emergency). 4. Identification in the Slovenian Health System In the year 2000 the Health Insurance Institute of Slovenia (ZZZS) [9] introduced the Slovene Health Insurance Card (KZZ) [9] as the official document applied in the implementation of the rights deriving from the compulsory and voluntary health insurance in Slovenia. Slovenia was the first country to introduce an electronic card at a national scale within the EU. The common EU member countries objective is to introduce an electronic document applicable within a country and across its borders. The Slovenian card is issued, free of charge, to every person upon the first regulation of the compulsory health insurance status in Slovenia and is used for the identification of the policy owner, the patient, as well as the health service providers. The card is made of plastic, measures 8.5 x 5.5 centimeters and has the chip with the following data: thecard owner data, • • thedata about the mandatory insurance policy, thedata about the voluntary health insurance, • • the data about the personal doctor, • the data about the medical accessories, the data about the owner decision regarding the • organ donation in case of death, the data about the issued medicine and • • the payer's data. The access to the data is granted only together with the authorized professional card owned by the medical staff (doctors, nurses, physicians, chemist). In October 2008 the ZZZS issued the new KZZ card that will replace the old KZZ card.
Page 404
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Figure 1: The new Slovene Health Insurance Card (KZZ) with the reader on the right In addition to the data on the old KZZ card, the new KZZ card includes the digital certificates issued by the ZZZSCA and has additional space for two additional personal certificates (2x2048 bits). The default certificate issued by the ZZZS is used for accessing the health insurance on-line system, while the additional personal certificates could be used for accessing other resources on the web. Figure 1 shows the new card (left) and the card reader (right).
Figure 2: The software for the management of the new Slovene Health Insurance Card On Figure 2 the software for the managements of the KZZ card is presented. The screenshot shows the available space on the card, the certificates status,the card type and the data stored on the card. While the old card was used as the data carrier, the new card is used as the identification element the key for accessing the health data. This change enables on-line access to the patients' data using digital certificates. In the future the card will be integrated with the Electronic Health Record (EHR), e-prescription and authorizations within the health information systems (i.e. patients waiting lists) as well as integration with the identity card. 5. Conclusion The identification is becoming one of the most important elements of the future IT-based solutions. The paper discusses legal as well as technical issues of the identification in the health care domain. Slovenia was the first EU country that introduced the electronic health
Velammal College of Engineering and Technology, Madurai
insurance card on the national scale. The second-generation card was recently put into production enabling better security and introduction of new e-health services. While the biometric identification might be better solution for the healthcare domain, the strong national and EU regulations prevents IT solution providers to use the technology in the Health Information Systems (HIS). Consequently, the Health Insurance Institute of Slovenia decided to use smartcard solution together with the digital certificates. In the future the integration with the identity card is anticipated.
References 1. Working Document on the processing of personal data relating to health in electronic health records (EHR), WP131, 2007 Ramesh Subramanian, Computer Security,Privacy 2. and, Politics-Current Issues Challenges and Solutions, IRM Press,2008 ISO/DIS 21091:2008 Health informatics 3. Directory services for security, communications and identification of professionals and patients ISACA, Control Objectives for Information and 4. related Technology(COBIT), version 4.1, IT Governance Institute, 2007 CEN/TR 15872 Health informatics - Guidelines 5. on patient identification and cross-referencing of identities Directive 95/46/EC of the European Parliament 6. and of the Council, http://ec.europa.eu/justice_home/fsj/privacy/docs/95-46ce/dir1995-46_part1_en.pdf (accessed March 19, 2010) Personal Data Protection Act (Slovenia), 7. http://ec.europa.eu/justice_home/fsj/privacy/docs/impleme ntation/personal_data_protection_act_rs_2004.pdf (accessed March 19, 2010) Information Commissioner (Slovenia), 8. http://www.ip-rs.si/ (accessed March 19, 2010) Health Insurance Institute of Slovenia, 9. http://www.zzzs.si/indexeng.html
Page 405
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Privacy Preserving Distributed Data Mining Using Elliptic Curve Cryptography M.Rajalakshmi#1, T.Purusothaman*2 #
Department of CSE/IT, Senior Grade Lecturer, Coimbatore Institute of Technology, Coimbatore,Tamilnadu, India 1
[email protected]
*
Department of CSE/IT, Assistant Professor,Government College of Technology, Coimbatore, Tamilnadu, India 2
[email protected]
Abstract— Association rule mining is one of the most investigated fields of data mining. When it occurs on distributed data, privacy of participating parties becomes great concerns. Distributed association rule mining is an integral part of data mining aimed at extracting useful information hidden in distributed data sources. As local frequent itemsets are globalized from various data sources, sensitive information pertaining to the individual data source needs high protection. Different privacy preserving data mining approaches for distributed environment have been proposed but they all suffer from privacy, computation complexity and communication overhead. The proposed method finds global frequent itemsets in a distributed environment with minimal communication among sites and ensures higher degree of privacy with Elliptic Curve Cryptography(ECC). The experimental results shows that ECC is more secure than RSA cryptography and generates global frequent itemsets without affecting mining performance and confirms optimal communication among sites Keywords— distributed data mining, privacy, secure multiparty computation, frequent itemsets, sanitization, cryptography
1. Introduction Major technological developments and innovations in the field of information technology have made it easy for organizations to store a huge amount of data within its affordable limit. Data mining techniques come in handy to extract valuable information for strategic decision making from voluminous data which is either centralized or distributed [1], [11]. The term data mining refers to extracting or mining knowledge from a massive amount of data. Data mining functionalities like association rule mining, cluster analysis, classification, prediction etc. specify the different kinds of patterns mined. Association Rule Mining (ARM) finds interesting association or correlation among a large
Velammal College of Engineering and Technology, Madurai
set of data items. Finding association rules among huge amount of business transactions can help in making many business decisions such as catalog design, cross marketing etc. A best example of ARM is market basket analysis. This is the process of analyzing the customer buying habits from the association between the different items which is available in the shopping baskets. This analysis can help retailers to develop marketing strategies. ARM involves two stages i) Finding frequent itemsets ii) Generating strong association rules
1.1 Association Rule Mining: Basic concepts Let I = {i1,i2…im} be a set of m distinct items. Let D denote a database of transactions where each transaction T is a set of items such that T ⊆ I. Each transaction has a unique identifier, called TID. A set of item is referred to as an itemset. An itemset that contains k items is a k-itemset. Support of an itemset is defined as the ratio of the number of occurrences of the itemset in the data source to the total number of transactions in the data source. Support shows the frequency of occurrence of an itemset. The itemset X is said to have a support s if s% of transactions contain X. The support of an association rule XÆY is given by Support = (Number of transactions containing XUY)/(Total number of Transactions) where X is the antecedent and Y is the consequent. An itemset is said to be frequent when the number of occurrences of that particular itemset in the database is larger than a user-specified minimum support. Confidence shows the strength of the relation. The confidence of an association rule is given by, Confidence = (Number of transactions Containing XUY) (Total number of Transactions containing X)
Page 406
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1.2 Distributed Data Mining In the present situation, information is the key factor which drives and decides the success of any organization and it is essential to share information pertaining to an individual data source for mutual benefit. Thus, Distributed Data Mining (DDM) is considered as the right solution for many applications, as it reduces some practical problems like voluminous data transfers, massive storage unit requirement, security issues etc. Distributed Association Rule Mining (DARM) is a sub-area of DDM. DARM is used to find global frequent itemsets from different data sources distributed among several sites and interconnected using a communication network. In DARM, the local frequent itemsets for the given minimum support are generated at the individual sites by using data mining algorithms like Apriori, FP Growth tree, etc.[1], [11]. Then, global frequent itemsets are generated by combining local frequent itemsets of all the participating sites with the help of distributed data mining algorithm [5], [2]. The strong rules generated by distributed association rule mining algorithms satisfy both minimum global support and confidence threshold. While finding global frequent itemset, local frequent itemsets at individual sites need to be collected. Due to that the participating sites know the exact support count of itemsets of all other participating sites. However, in many situations the participating sites are not interested to disclose the support counts of some of their itemsets which are considered as sensitive information. Hence the privacy of sensitive information of the participating sites is to be preserved [8]. In such cases, classical data mining solutions cannot be used. Hence, Secure Mutiparty Computational (SMC) solutions can be applied to maintain privacy in distributed association rule mining [16]. The goal of SMC in distributed association rule mining is to find global frequent itemset without revealing the local support count of participating sites to each other. The subsequent sections of the paper are organized as follows. Firstly, related existing works are reviewed. Secondly, the proposed approach and its performance evaluation are discussed. Lastly, a suitable conclusion and future work for maintaining privacy is attempted.
2. Related Work While practicing data mining, the database community has identified several severe drawbacks. One of the drawbacks frequently mentioned in many research papers is about
Velammal College of Engineering and Technology, Madurai
maintaining the privacy of data residing in a data source [6], [20]. Privacy preserving data mining provides methods for finding patterns without revealing sensitive data. Numerous research works are underway to preserve privacy both in individual data source and multiple data sources. There are two broad approaches for privacy-preserving data mining namely data sanitization [21] and Secure Multiparty Computation (SMC) [10], [16]. A few representative works in SMC is discussed in this section. The concept of Secure Multiparty Computation (SMC) was introduced in [22]. In many applications the data is distributed between two or more sites, and for mutual benefit these sites cooperate to learn the global data mining results without revealing the data at their individual sites. The basic idea of SMC is that this computation is secure if at the end of the computation no party is unaware about the other participating sites except its input and the results. The secure computation protocols are presented in the form of combinational circuit [9]. The idea is that the function F to be computed is first represented as a combinational circuit and then the parties run a short protocol to securely compute every gate in the circuit. Every participant gets corresponding shares of the input wires and the output wires for every gate. Here the size of the protocol depends on the size of the circuit, which depends on the size of the input. This is inefficient for large inputs as in data mining. In paper [18] Vaidya et al. proposed a method for privacy preserving association rule mining in vertically partitioned data. Each site holds some of the attributes of each transaction. An efficient scalar product protocol was proposed to preserve the privacy among two parties. This protocol did not consider the collusion among the parties and was also limited to boolean association rule mining. In [14] Kantarcioglu et al. proposed a work to preserve privacy among semi-honest parties in a distributed environment. It works in two phases assuming no collusion among the parties. But in some real life situations collusion is inevitable. The first phase identifies the global candidate itemsets using commutative encryption (RSA algorithm) which is computationally intensive and the second phase determines the global frequent itemsets. This work only determines the global frequent itemsets but not their exact support counts. An algorithm using Clustering future (CF) tree and secure sum is proposed to preserve privacy of quantitative association rules over horizontally partitioned data [17] and fixing of proper threshold value for constructing the CF tree is not easy. The efficiency of the algorithm is
Page 407
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
unpredictable since it depends on the threshold value chosen to construct the CF tree. The present analysis proposes a novel, computationally simple and secure multiparty computation algorithm for semi-honest parties.
One basic condition for any cryptosystem is that the system is closed, i.e. any operation on an element of the system results in another element of the system. In order to satisfy this condition for Elliptic curves it is necessary to construct non-standard addition and multiplication operations.
3. Problem Definition
To encrypt P, a user picks an integer, k, at random and sends the pair of points (k*BP, P+k*PUBKEY)
Let S1, S2… Sn. be the set of participating sites where n>2. Let D1,D2…Dn be the data sources of sites S1, S2… Sn respectively which are geographically distributed and let I = {i1,i2…im} be a set of items. Each transaction T in Di such that T ⊆ I, where i=1 to n. Li be the local frequent itemset generated from a participating site Si and G be the global frequent itemset. To generate the global frequent itemset G, each site needs to send its respective support counts of its local frequent itemsets to the other participating sites. The intended goal of this proposed approach is to discover the global frequent itemsets without revealing the sensitive information of all the participating sites.
4. Proposed Approach A new approach is proposed in this paper to estimate the global frequent itemsets from multiple data sources while preserving the privacy of the participating sites. All sites in the mining process are assumed as semihonest. The semi-honest parties are honest but try to learn more from received information. Any site can initiate the mining process. To protect sensitive information of each participating site, elliptic curve cryptography and randomization are applied.
4.1 Overview of Elliptic Curve Cryptography (ECC) Many cryptographic algorithms and protocols have been proposed to encrypt and decrypt data. Among the algorithms, ECC is proved as more secured cryptographic technique. Also it is computationally efficient than RSA and Diffie-Hellman [23]. The use of elliptic curve groups over finite fields as a basis for a cryptosystem was first suggested by Koblitz [24]. The elliptic curve approach is a mathematically richer procedure than standard systems like RSA. The basic units for this cryptosystem are points (x,y) on an Elliptic curve, E(Fp), of the form Y2 =x3+ax+b, with x,y,a,b~Fp ={1,2,3 ....... p-2,p-1} Where Fp is a finite field of prime numbers.
Velammal College of Engineering and Technology, Madurai
To decrypt this message you multiply the first component by the secret key, s, and subtract from the second component, (P+ k*PUBKEY ) - s*(k*BP) = P+k*(s*(BP)) - s*(k*BP) =P We then reverse the embedding process to produce the message, m, from the point P. This system requires a high level of mathematical abstraction to implement.
4.2 Secure Mining of Support Counts Our proposed method finds the global frequent itemsets without disclosing the support counts of individual sites. Algorithm 1 given below computes global candidate itemsets in a secured manner using ECC technique. Algorithm 2 computes the globally frequent itemsets using random number addition.
Algorithm 1 Input: n local frequent itemsets. Each set, denoted by Ai (1≤ i ≤ n), belongs to each party. s is the Threshold support. Si denotes the local storage of each party. A1 denotes the commoner, i.e. i=1.LL is the count of locally frequent itemsets received so far by the commoner. Initially LL=0. CLL is the total count of locally frequent itemsets. Output: UiAi (1≤i≤n), that is the set union of locally frequent itemsets (Global candidate itemsets). 1. Compute the locally frequent itemsets at each site. Make an entry in the local storage, Si of each party. 2. Encryption traversal: Encrypt the locally frequent itemsets using the • public key of the commoner say pub1. • Send the encrypted message to the next party, which encrypts the locally frequent itemsets using its public key and forwards the result to the next site and the process continues until it reaches the commoner again. 3. Decryption traversal: • Decrypt the message using the private key of the commoner say priv1. • Send the decrypted message to the next party, which decrypts the message further using its private key and forwards the result to the next site and the process continues until it reaches the commoner again.
Page 408
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
4. Increment LL If any received subset is not present in the local storage Si Then add to Si 5. The process ends when LL equals CLL at the commoner. The above algorithm meets the requirement that no party can determine any itemset’s owner. Initially, each of the participating sites generates their locally frequent itemsets using any of the frequent itemset generation algorithms like Apriori, FP growth tree algorithm, etc. The generated locally frequent itemsets at each site, denoted as Ai, are added to the local storage, Si of respective sites. The purpose of the local storage is to avoid the infinite looping of any itemset before reaching the commoner. This is accomplished by forwarding any received itemset by verifying its presence in the local storage. Thus the local storage ensures that any itemset will pass through all the participating sites only once, that is at the worst case. The count of locally frequent itemsets generated at each of the participating sites is calculated to determine the end of the Algorithm 1. For this, the commoner chooses a random number to be added with its count of locally frequent itemsets. This value is sent to the next participating site, which adds its count of locally frequent itemsets to the received value and forwards the result to the next site. The process is repeated until all the participating sites have added their counts of locally frequent itemsets and when the value reaches back the commoner, it subtracts the random number to get the count of all locally frequent itemsets. It is denoted as CLL. The commoner maintains the count of received itemsets (LL). On receiving any forwarded itemset, the commoner increments LL and if it is not present in the local storage, an entry is made. If it is present already then the received itemset is discarded. Thus the commoner removes the redundancies. When the other sites receive any forwarded itemset, it is verified with the respective local storage and if an entry is present, then the itemset is directly forwarded to the commoner, else an entry is made and the itemset is forwarded to random destinations. Thus we avoid the repeated looping of any itemset. The end of Algorithm 1 is when the commoner’s count of received itemsets (LL) reaches CLL. The commoner’s local storage contains the union of the locally frequent itemsets of all the participating sites. The following algorithm privately computes the globally supported itemsets [5].
Algorithm 2 Input: Global candidate itemsets. Number of participating sites n >= 3. The site initiate the mining process can be the commoner.
The commoner calculates the global database size 1. as follows: • Add any random number R to the local database size. • Send the value to the next party, which adds its local database size to the received value and forwards the resulting value to the next site and the process continues until it reaches the commoner again. The commoner subtracts R from the received • value to get the global database size. • 2. For each item i in S1 do Add a random number to the local support count of I and send it to the next site. The next site adds its support count for i to the count and sends it to the next site and this process continues till the count reaches the commoner. End for. The global support percentage of an itemset, X is calculated as summation of local support over all the participating sites, 1≤i≤n divided by the summation of the database size (DB) over all the participating sites, 1≤i≤n. The commoner securely calculates the global support count of each itemset in the union of locally frequent itemsets. For each itemset in the union, the commoner the next site. The process is repeated until it reaches back the commoner. The commoner calculates the global support percentage for every itemset in the union. Those itemsets whose support is greater than or equal to the threshold support form the set of globally frequent itemsets.
5 Performance Evaluation Distributed environment with three participating sites was simulated to evaluate the performance of the proposed algorithm. At each site, a P-IV, 2.8 GHz, 2 GB RAM machine is run on Windows operating system. The proposed method is implemented using Java and MSAccess. Three different data sources of size 10K, 50K, and 100K are synthetically generated to study the performance of the proposed work. The local frequent itemsets of the participating sites which are the inputs for the algorithm is generated by using the FP tree algorithm for supports varying from 10% to 30%.
5.1 Timing analysis
Output: Global frequent itemsets.
Velammal College of Engineering and Technology, Madurai
Page 409
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Graph in Fig. 1 shows that the time taken to encrypt and decrypt the local frequent itemsets using existing RSA D|=100k,|I|=12,ATL=9; No. of sites = 3
60
140
50
120 100
40
Time (in sec.)
Tim e (in Sec.)
|D| = 10k, |I| = 10, ATL=7; No. of sites = 3
RSA 30
ECC
20 10
80
RSA
60
ECC
40 20
0
0
10
20
30
10
Cryptographic Technique and Our proposed ECC Technique. The result shows that the ECC cryptographic technique is more efficient than the existing cryptographic technique. Fig. 1.1 Time complexity for Encryption and Decryption of 10K data source
30
Fig 1.3 Time complexity for Encryption and Decryption of 100K data source
5.2 Accuracy analysis Accuracy is one of the main objectives of privacy preserving distributed data mining. Table 1 shows the number of global frequent itemsets generated by both methods that use traditional RSA method and proposed ECC method. Both are found to be identical.
|D|=50k,|I|=10,ATL=7; No. of sites = 3 140 120 Time (in Sec.)
20 Support (%)
Support (%)
100 80
RSA
60
ECC
40 20 0 10
20
30
Support (%)
Fig. 1.2 Time complexity for Encryption and Decryption of 50K data source TABLE 1
Accuracy comparison of traditional RSA and proposed ECC methodology
Data set 10K 50K 100K
Support
No. of Sites
10% 20% 10% 20%
3 3 3 3
10% 20%
3 3
Total No. of Global Frequent Itemsets RSA ECC 1023 1023 320 320 231 231 175 175
Velammal College of Engineering and Technology, Madurai
60 10
60 10
Accuracy 100% 100% 100% 100% 100% 100%
Page 410
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6 Conclusion and Future Work Several existing algorithms for preserving privacy in a distributed environment have been analyzed and an efficient algorithm for the same has been proposed and implemented. This proposed method uses a mathematically rich procedure of cryptography namely the elliptic curve cryptography. The proposed model believes that the participating sites are semi honest. We can extend the proposed model to work even for dishonest parties. The proposed model works for only homogeneous databases. We can improve the system to work even for heterogeneous databases where the attributes of the participating sites are different.
References [32] R.Agrawal, R.Srikant, “ Fast Algorithms for Mining Association Rules, “ in proceedings of the 20th VLDB Conference Santiago, Chile,1994, pp.487-499. [33] M.Ashrafi, D.Taniar, K.Smith, ” ODAM : An Optimized Distributed Association Rule Mining Algorithm,” IEEE Distributed Systems Online, 2004, 5th ed., Vol. 3. [34] M.Ashrafi, D.Taniar, K.Smith, “ Privacy-PreservingDistributed Association Rule Mining Algorithm.” in proceedings of International Journal of Intelligent Information Technologies, 1st ed. Vol.1, 2005, pp. 46-69. [35] M.Atallah, E.Bertino, A. Elmagarmid, M.Ibrahim, V.S.Verykios, “Disclosure limitation of sensitive rules,” in Proceedings of the IEEE Knowledge and Data Exchange Workshop (KDEX'99). IEEE Computer Society, 1999, pp. 45-52. [36] D. Cheung, V.Ng, A.Fu, Y.Fu, “Efficient Mining of Association Rules in Distributed Databases,” IEEE Transactions on Knowledge and Data Engineering. 1996, 8th ed., Vol. 6, pp. 911-922. [37] C.Clifton, D.Marks, “Security and Privacy Implications of Data Mining,” in Proceedings of the ACM SIGMOD Workshop on Data Mining and Knowledge Discovery, 1996, pp.15-19. [38] C. Clifton, “Secure Multiparty Computation Problems and Their Applications: A Review and Open Problems,” in Proceedings of the Workshop on New Security Paradigms, Cloudcroft, New Mexico, 2001. [39] C.Clifton, M. Kantarcioglu, J.Vaidya, “Defining privacy for data mining. Book Chapter Data Mining, Next generation challenges and future directions, 2004. [40] O.Goldreich, S, Micali, Wigderson, “How to play any mental game - a completeness theorem for protocols with honest majority,” in 19th ACM symposium on the theory of computing, 1987, 218-229. [41] O.Goldreich, “Secure Multiparty Computation (Working Draft)”, 1998. [42] J.Han, M. Kamber, ”Data Mining: Concepts and Technique”, Morgan Kaufmann Publishers, 2001. [43] J. Han, J. Pei , Y.Yin, R. Mao, “Mining Frequent Patterns without Candidate Generation : A Frequent Pattern Approach,” IEEE Transactions on Data Mining and Knowledge Discovery, 8th ed., Vol.1, 2004, pp.53-87. [44] Inan.A., Saygyn.Y., Savas.E., Hintoglu.A.A, & Levi.A, “ Privacy preserving clustering on horizontally portioned data,” in proceedings of
Velammal College of Engineering and Technology, Madurai
the 22nd International Conference on Data Engineering Workshops (ICDEW'06), 2006. [45] M. Kantarcioglu, M., C.Clifton, “Privacy-Preserving Distributed Mining of Association Rules on Horizontally Partitioned Data,” IEEE Transactions On Knowledge And Data Engineering, 2004, 16th ed. Vol.9. [46] G.Lee, C.Chang, A.L.P.Chen, “ Hiding sensitive patterns in association rules mining,” in Proceedings of the 28th Annual International Computer Software and Applications Conference (COMPSAC'04), 2004. [47] Y.Lindell, B. Pinkas, ” Secure Multiparty Computation for PrivacyPreserving Data Mining, “ in the Journal of Privacy and Confidentiality, 2009, 59-98. [48] W.Luo, “An Algorithm for Privacy-preserving Quantitative Association Rules Mining,” in Proceedings of the 2nd IEEE International Symposium, 2006. [49] J.Vaidya, C.Clifton, “Privacy Preserving Association Rule Mining in Vertically partioned data,” in Proceedings of the SIGKDD ’02 Copyright ACM, 2002. [50] J.Vaidya, C.Clifton, “Privacy-Preserving Data Mining: Why, How, and When, “ IEEE Security and Privacy, 2004. [51] S.Verykios, E.Bertino, I.Provenza, Y.Saygin, Y.Theodoridis, “State- of- the -Art in Privacy Preserving Data Mining,” in ACM SIGMOD Record, 33rd ed., Vol 1., 2004, pp. 50-57. [52] S.Wang, Y. Lee, S.Billis, A.Jafari, “Hiding Sensitive Items in Privacy Preserving Association Rule Mining” in IEEE International Conference on Systems, Man and Cybernetics, 2004. [53] A.C. Yao, “ How to generate and exchange secret,” in Proceedings of the 27th IEEE symposium on Foundations of Computer Science. 1986, pp. 162-167. [54] M. Brown, D. Hankerson, J. Lopez, and A. Menezes, “Software Implementation of the NIST Elliptic Curves Over Prime Fields, ” IEEE computer society, 2004. [55] Andrew Burnett, Keith Winters, and Tom Dowling, “A Java implementation of an elliptic curve Cryptosystem- Java programming and practices,” 2002. [56] Mugino Saeki ,“Elliptic curve cryptosystem, ” McGill University, Montreal, February 1997.
Page 411
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Harmonics In Single Phase Motor Drives And Energy Conservation Mustajab Ahmed Khan#1, Dr.A.Arunagiri #2 #
EEET Department, Yanbu Industrial College, Royal Commission ,Kingdom of Saudi Arabia. 1
[email protected] [email protected]
2
Abstract— The modern electric power systems that include non-linear loads may experience power quality problem. Nonlinear loads draw current that cause power quality problems such as harmonic distortion. Harmonic problems are mainly due to substantial increase of non- linear loads such as the Switched mode power supplies(SMPS),Uninterrupted power supplies, Variable speed drives using power electronic devices or microprocessor and power electronic controllers in AC/DC Transmission links. In this paper Electronic regulators, resistive type regulators for single phase induction motors used in fans are analyzed for harmonic analysis and energy saving. A proposed capacitive type regulator is also analyzed and compared with the existing regulators.
Keywords— DAS, Triplens , Harmonics, Regulators,FFT.
1.
INTRODUCTION
The harmonics and power saving schemes study in this paper are based on single phase induction motor in fan loads. A fan regulator is a means of regulating the fan speed in order to achieve the desired air velocity. It is interfaced between the single phase supply and the single phase induction motor which drives the fan blades and effects change in speed . At present the electronic type regulators are increasingly being used in place of the resistive type regulator on account of power savings. compactness, step less speed regulation . However, electronic regulators operate at low power factor and generate high level of harmonics. While the low power factor operation will mean extra burden on the supply system for the supply of reactive power, the harmonics generated cause overheating of the motor which could lead to reduction in fan life. Further, electronic regulator cause a typical humming sound, disturbance in TVs, tape recorder and interference with the telecommunication networks. It is therefore necessary to examine the fan regulators critically from two angles namely, it's energy consumption potential and its side effects. With this in mind, the capacitive type regulator was developed and its performance determined
Velammal College of Engineering and Technology, Madurai
and compared with other regulators like the resistive type and electronic type.
2.
HARMONICS
Harmonics are integral multiples of the fundamental frequency. Devices causing harmonics are present in all industrial, commercial and residential, installations. Harmonics are caused by non-linear loads. A load is said to be non-linear when the current it draws does not have shape as the supply voltage. Devices comprising power electronics circuits are typical non-linear loads. Such loads are increasing frequently and their percentage in overall consumption is growing steadily. Examples include industrial equipment (Welding machines, arc furnaces. Induction furnaces rectifiers),variable-speed drives for Asynchronous and DC motors, office equipment (PCs, photocopy machines, fax machine), household appliances (television sets, microwave ovens, fluorescent lighting, etc.) and UPS. 2.1
Effects of Harmonics:
In distribution systems, the flow of harmonics reduces power quality and consequently causes a number of problems: overloads on distribution systems due to the 1. increase in the rms current. 2. overloads on neutral conductors due to the summing of third-order harmonics and triplens created by single-phase loads. overloads, vibrations and premature ageing of 3. generators, transformers, motors, etc., transformer hum. overloading and premature ageing of capacitors in 4. power factor correction equipment. distortion of the supply voltage, capable of 5. disturbing sensitive loads.
Page 412
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6. disturbances on communications networks and telephone lines. The harmonics most frequently encountered (and consequently the most troublesome) on three-phase distribution systems are the odd-order harmonics (3rd, 5th, 7th, etc.).Utilities monitor harmonic orders 3, 5, 7, 11 and 13. It follows that conditioning of harmonics is imperative up to order 13 and ideally should include harmonics up to order 25. 2.2 Selection of measurement devices:
The conventional resistive type regulator controls the speed by suitably dropping the voltage across the resistance in series with the motor thus causing considerable power loss. 3.2
Electronic regulator:
Digital analyzers, based on recent technology, provide sufficiently accurate measurements for the indicators presented above. Other measurement devices were used in the past.
The solid state fan regulator is triac based. It operates on the principle of the firing angle control of the triac which consumes negligible power. Delayed firing angle of the triac causes chopping of the current to reduce the voltage across the motor and hence reduced the speed. This feature which results in energy conservation. However, the non linear operating nature of the electronic circuitry is the main causes of harmonic generation .
2.2.1
3.3
Functions of Digital Analyzers:
Capacitive type:
The microprocessors used in digital analyzers calculate the values of the harmonic indicators (power factor, crest factor, distortion power, THD), offer a number of additional functions (correction, statistical detection, management of measurements, display, communication, etc.), when they are multi-channel devices, provide simultaneously and nearly in real time the spectral breakdown of voltage and current.
It is a step type regulator like the resistive type . It's power consumption is near to zero at all speeds and there is no waveform distortion. The system operates near unity power factor or slightly leading power factor . This type of regulator can be used for single phase motors of exhaust fans, mixers and grinders .
2.2.2 Operating principle of Digital Analyzers and dataprocessing techniques:
A sample of different types of regulators was chosen for evaluation. The following parameters of the different fan regulators where measured by using the data acquisition system(DAS), software from LabVolt, laptop, and the comparative analyses of the results were done.
Analogue signals are converted into a series of digital values. On the basis of the digital values, an algorithm implementing the Fast Fourier Transform (FFT) calculates the amplitude and the phases of the harmonics over a large number of observation time windows. Most digital analyzers measure harmonics up to the 20th or 25th order for calculation of the THD. Processing of the various values calculated using the FFT algorithm (smoothing, classification, statistics) can be carried out by the measurement device or by external software.
3.4
Methodology of the evaluation:
current drawing by the regulator and the fan 1. combination . power consumption by the regulator . 2. 3. voltage and current harmonics (THD(V),THD(I)) Power factor of the circuit. 4.
On the basis of analysis results, it may be necessary to: Derate any future equipment installed. 1. Quantify the protection and harmonic-filtering 2. solution that must be installed. Compare the values measured to the reference 3. values of the utility (harmonic- distortion limits, acceptable values, reference values).
3.5 3.5.1
Measurement and Analysis: Electronic control of the Single phase motor fan: Figure 1:Electronic type regulator.
PRINCIPLE OF OPERATION OF DIFFERENT 3. TYPES OF REGULATORS 3.1
Resistive type:
Velammal College of Engineering and Technology, Madurai
Page 413
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
α
V1
100.5 110. 7 90 110. 6 75.7 110. 6 60.5 110. 6 45 110. 6 30.2 110. 6 15 110. 6 7 110. 6 0 10.6 1
V2
W2
cosφ THD THD N (V) (I) (rpm)
I1
W1
0.11 9 0.13 1 0.14 3 0.14 3 0.14 7 0.14 7 0.14 7 0.14 6 0.14 9
7.6 67.9 7.45 0.76 0 6 9.65 77.5 0.84 9.6 2 12.3 89.3 12.1 0.91 2 7 6 3 12.3 89.3 12.1 0.91 2 7 6 3 15.3 104. 15.2 0.98 0 9 0 2 15.9 108. 15.7 0.99 1 4 3 16.1 109. 15.9 0.99 5 4 7 8 16.2 110. 16.0 0.99 3 1 7 9 16.4 109. 16.4 0.99 6 9 0 9
2.9
66.8 730
3
62.4 850
3
54
1200
2.9
43.5 1500
2.6
31.3 1715
3.6
19.9 1805
4.2
10.3 1835
3.6
6.3 1860
3.3
4.9 1880
Table 1: Measurements of the Electronic type regulator. (V1-Supply Volts,I1-Input Current ,W1-Input Power, W2-Motor Power,V2-Motor Volts)
Figure 3. Current harmonic spectrum for (I1)( α=45 deg)
Figure 4. Voltage harmonic spectrum(obtained from DAS).
3.5.2 Resistive type:
Figure 2.Current and voltage waveform.( α=45deg) Figure 5: Resistive type regulator.
Velammal College of Engineering and Technology, Madurai
Page 414
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Table 2. Measurements of the Resistive type regulator
V1
I1
110 109.7 110.3 110.7 110.7
W1
0.1 0.11 0.12 0.13 0.15
V2
Cos(φ) %THD %THD N (V) (I) (rpm) 6.25 0.998 1.6 5.1 722 7.8 0.998 1.6 4.5 870 9.8 0.997 1.1 3.9 1100 11.95 0.997 2.2 4.3 1360 16.64 0.987 2.8 4.1 1880
W2
10.8 64 11.85 71 13.34 81.24 14.61 90.7 16.63 110.6
Figure 8:plot of THD(V) Vs Speed 3.5.3 Capacitor regulator for fan:
Figure 9: plot of THD(I) Vs Speed
Figure 6 : Capacitive type regulator
V1
I1
W1
V2
W2
120 120 120 120 120 120
0.09 0.124 0.133 0.136 0.141 0.152
5.45 10.65 12.33 12.72 14.46 17.2
61 86 94 94 102.8 113
5.45 10.65 12.33 12.72 14.46 17.2
Cos(φ) %THD (V) 0.5 1.5 0.675 1.4 0.778 1.5 0.781 1.7 0.857 1.2 0.942 1.6
%THD (I) 5.6 4.1 3.8 5.6 3.4 4.7
N(rpm) 690 1278 1450 1545 1700 1940
Table 3: Measurements of the capacitive type regulator
Figure 10: plot of input Power Vs Speed 4
Figure 7 :plot of power factor Vs Speed
Velammal College of Engineering and Technology, Madurai
CONCLUSION
The results of the evaluation are summarized as : It is found the power loss in the resistive type is of the order of 5 Watts per fan . Capacitive or electronic type can have power saving of about 4.5 watts. per unit. As the fan load is huge, the overall power saving will be enormous. Capacitive type regulator has low level of harmonic in both voltages (1-2%) and current which is less than 5%. The electronic regulator has a level of THD about 5% for voltage and (6to 60)% for current .Capacitive and electronic type regulators have better energy saving potential, but an electronic regulators is a source of
Page 415
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
harmonic. Capacitive type regulators has no noise problems which exist in electronic regulators at low speed. REFERENCES [57] Francis De La Rosa, “Harmonics and Power Systems”,2006, CRC Press. [2] Effects of harmonics on equipment, IEEE Transactions, April 1993,Vol 8. [3] George Wakelah,2000” Harmonics in power systems, their causes,effects and mitigation”, Department of electrical & comp.engg-,Mexico University. [4] Roger C Dugan,Electrical power Systems Quality, McGraw-Hill,New York,1996,pp.130-133.
Velammal College of Engineering and Technology, Madurai
Page 416
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Improvement towards efficient OPFET detector Jaya V. Gaitonde, Rajesh B. Lohani Department of Electronics and Telecommunications, Goa University
Goa Engineering College, Farmagudi-Ponda-Goa-India [email protected], [email protected]
Abstract—OPFET(Optical Field Effect Transistor) is an useful device for optical communication and as photo detector. This paper discusses some performance issues of the device under various conditions. Keywords—OPFET, photodetector, switching time, Noise Equivalent Power(NEP), Signal to Noise Ratio(S/N).
I . INTRODUCTION At present, optical fiber communication plays an important role in cable communication technology for wideband, multimedia and high-speed applications [3] [5]. In order to be able to manufacture wireless terminals for optical fiber links at reasonable cost, good agreement must be achieved between the photo-detector and the millimeter-wave circuit, as well as small size and low weight. When a GaAs MESFET device is optically illuminated, absorption phenomena take place at the gate–drain and gate–source regions, which induce both photoconductive and photovoltaic effects. The performance of a GaAs MESFET can be significantly enhanced by scaling down the device geometry. The radiation is allowed to fall on the semitransparent Schottky gate of the device. When light is turned ‘on’ the parameters such as threshold voltage, channel charge, channel current, channel conductance and gate to source capacitance reach the steady state value at a lesser time than that when light is turned off. The device performance is greatly improved by shortening the gate length. The OPFET performance is also dependent on the NEP(Noise Equivalent Power) and signal to noise ratio(S/N)[4]. It is understood that the received optical signal generates electron-hole pair in the semiconductor resulting in the modulation of the channel conductivity due to photoconductive effect and channel conductance through a development of a forward photovoltage due to photovoltaic effect. The electrical parameters such as threshold voltage, drain-source current, gate capacitances and switching response affect the OPFET’s performance. The diffusion process introduces less process-induced damage compared to ion-implantation, which suffers from current reduction due to a large number of defects introduced by ion-implantation process.
Velammal College of Engineering and Technology, Madurai
The photocurrent peak value, peak-time and discharge-time are some of the performance factors. The frequency response of OPFET is dependent upon transit time and the RC time constant of the device. II.
THEORY
The schematic structure of OPFET is as shown in fig. 1a.
Fig. 1 a. Cross-sectional view of an OPFET [2].
The schematic structure of the ion-implanted GaAs OPFET with back illumination is shown in Fig.1b and 1c for the two cases.
Fig.1. b Schematic structure of the device with fiber inserted partially into the substrate [8]. Fig 1 c Schematic structure of the device when the fiber is inserted up to the substrate-active layer interface [8].
Page 417
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.!. d Schematic structure of the opaque gate device [5]. Fig1.e Schematic structure of the device which is illuminated everywhere on the surface [5].
With the improvement on the process technology, a short gate-length MESFET device can be fabricated. The radiations are allowed to fall on the schottky gate as well as between the spacing between the source-gate & gate- drain , allowing more absorbtion. The gate-controlling capability will be reduced by the penetration of the electric field from the sidewall at both sides of the gate, and the threshold voltage of a short gate-length device will be influenced by the drain bias. The drain current Ids has been calculated by numerically integrating the charge in the channel region given by (1) Ids= εoεrZµn/d[Vgs-Vth-V(x)]E(x) where εo and εr are the permittivity of the air and GaAs , d is the gate length, Z is the device width, µn is the field dependent mobility of electrons in the channel length, Vgs is the gate-to-source voltage, V(x) is the potential distribution in the channel, E(x) is the electric field in x direction. The RC time constant has been obtained by RC= Cgs/gm where Cgs is the gate to source capacitance and gm is the transconductance. The frequency response of a GaAs MESFET depends upon the transit time and the RC time constant of the device. The transit time is the finite time required for the carriers to travel from the source to the drain. For short channel devices, the transit time is usually small compared to the RC time constant resulting from the input gate capacitance and the transconductance. In the presence of illumination,
Velammal College of Engineering and Technology, Madurai
the RC time constant is affected significantly by the incident optical power. The responsivity of the device has been obtained from R(λ)= Iph/Popt where Iph is the photoresponse current and Popt is the incident optical power. λ is the operating wavelength. The photocurrent gain M is given by M= Iph/IL where Iph is the photoresponse current and IL is the primary current. The photoresponse current is the difference between the drain current in the illuminated condition and that in the dark condition. The primary current IL is given by IL= qηPoptA/hv where q is the electronic charge, η is the quantum efficiency, Popt is the incident optical power, A is the illuminated device area, h is the Planck’s constant and v is the frequency of incident radiation. The threshold voltage Vth of an optically biased MESFET is given by Vth= Vth0 – sech( k1Lg/2)A1s where Vth0 is the threshold voltage for a long channel optically biased MESFET given by Vth0=Vbi-Vop-q/e[(Nd-Rtp’/b)(b2/2)-Qtn’((b+1/α)exp(-αb)1/α)] where A1s is the first term of the Fourier coefficient for excess sidewall potential at the source side of the gate, Nd is the uniform donor impurity concentration in the channel, R is the surface recombination rate, b is the thickness of the active layer, Q is the incident photon flux density, tn’ is the lifetime of the electron, tp’ is the lifetime of the hole, α is the absorption coefficient per unit length, k1 is the eigen value of the Green’s function in the gate region, and Lg is the gate length. Vbi is the built-in potential at the Schottky gate contact, and Vop is the photovoltage developed at the Schottky junction due to illumination. The switching time of an OPFET depends on the active channel thickness and corresponding impurity flux density of the diffusion process. The switching time t is computed for different active layer thickness expressed by the following equation t=[L(QBn-Vop-Vgs)1/2]/[Vs[(qNdavga2/2ε)1/2-(QBn-VopVgs)1/2]] (2) Finally, the switching time t as a function of impurity flux density Qdiff can be expressed by following equation: t=[L(QBn-Vop-Vgs)1/2]/[Vs[(Qdiffq(D1t1)1/2/2ε(Л)1/2)1/2(QBn-Vop-Vgs)1/2]] (3) where q is the electronic charge, L is the channel length, Ndavg is the average channel doping concentration, a is the active layer thickness, Qdiff is the impurity flux density during diffusion process, D1 is the diffusion constant for diffusion process, t1 is the diffusion time, Vgs is the gate to source voltage, QBn is the indium tin oxide Schottky barrier height, Vop is the photo induced voltage, Vs is the saturation velocity.
Page 418
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
When a GaAs MESFET device is optically illuminated, absorption phenomena take place at the gate–drain and gate–source regions, which induce both photoconductive and photovoltaic effects. In front illumination, the radiation is allowed to fall on the semi-transparent Schottky gate of the device. For opaque gate, the light is incident through the gate-drain and gate-source spacings. In the case of back illumination, two cases are considered: one in which the fiber is inserted partially into the substrate and the other, in which the fiber is inserted upto the active layer—substrate interface. The later case represents improved absorption in the active layer of the device. In Fig.1b ,the fiber is inserted partially into the substrate so that the absorption takes place in both substrate and active region. In Fig1c, the fiber is inserted up to the junction of the substrate and the active layer where photo-absorption takes place in the active region only. The drain-source current flows along the x-direction and the illumination is incident along the y-direction of the device. Electron-hole pairs are generated due to absorption of photons in the neutral substrate region, the active layer-substrate depletion region, the neutral channel region and the schottky junction depletion region. The optically generated electrons move toward the channel and contribute to the drain-source current when a drain-source voltage is applied while the holes move in the opposite direction. When these holes cross the junction a photo-voltage is developed. This voltage being forward biased reduces the depletion width of both the junctions.
The current is enhanced due to absorption more in the active channel region.
Fig. 2. switching time τ versus Active layer thickness for dark and optical illumination condition.
Equation (2) is used to compute the switching time τ as a function of the device active layer thickness a for dark and optical illumination conditions. The result illustrated in Fig. 2 shows that the switching time under dark condition abruptly decreases with respect to increasing device active layer thickness. The switching time is also reduced under optical illumination condition. Therefore it is observed that the transition from dark to illuminated condition of the OPFET significantly changes the switching time for certain device active layer thicknesses.
The drain-source current is significantly enhanced for the device when the fiber is inserted up to the active layer substrate junction than the case where finite substrate effect is taken into account. Under back illumination the number of carriers generated at the active layer-substrate depletion region is more than that in the gate depletion region. For opaque gate, two photo-voltages are developed: one across the Schottky junction due to generation in the side walls of the depletion layer below the gate and the other across the channel-substrate junction due to generation in the channel-substrate depletion region. There is no generation of electron-hole pairs just below the gate. When the device is illuminated over the entire surface, i.e. on the gate and between the spacing of gate-source and gate-drain, two photo-voltages are developed, one at the Schottky junction and one at the active-layer substrate junction.
III.
Fig 3 switching time τ versus impurity flux density .
The characteristics of switching time as a function of impurity flux density have been determined by using the equation (3) and the plot is illustrated in Figure3. The switching time is shown to be abruptly decreasing initially for increasing impurity flux densities and still decreasing but less dramatically with respect to continued increase in impurity flux density. The plot indicates that the switching characteristics change with low, moderate and high impurity deposition on the silicon during diffusion process. Most of the optically generated carriers should originate within the Schottky depletion region (gate depletion) for both high quantum efficiency and high switching speed, and in addition, the depletion region must not be so wide
RESULTS AND DISCUSSIONS
Velammal College of Engineering and Technology, Madurai
Page 419
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
that the transit time of the carriers limits the frequency response. It is also seen from the two graphs that the switching time required for back illumination is less compared to other cases. This is because for the back illumination case, there is enhanced absorption in the active region. IV.
Microwave and optical technology letters / vol. 26, no. 4, August 20 2000 [8]. Nandita Saha Roy and B. B. Pal “Frequency-Dependent OPFET Characteristics Improved Absorption under Back Illumination.” Journal of Lightwave Technology, Vol. 18, No. 4, April 2000.
with
[9]. Nandita Saha Roy, B. B. Pal, and R. U. Khan “Analysis of GaAs OPFET with Improved Optical Absorption under Back Illumination.” IEEE Transactions on Electron Devices, Vol. 46, No. 12, December 1999.
CONCLUSION
Some methods of increasing the performance of OPFET are studied. It is found that the performance of the device is increased by reducing the switching time with the increase in the active layer thickness and the impurity flux density. Also, the device with back illumination, having reduced switching time may be considered as an useful device for the design of high speed optical detector and radio frequency optical switch in communication. ACKNOWLEDGEMENT Authors thanks Dr. B.B. Pal , ITBHU, and Mr. Vivek Kamat, DTE Goa for providing constant help and encouragement and necessary guidance. REFERENCES [1] Shubha Rani Saxena, R. B. Lohani, R. U. Khan, B. B. Pal. “Generalized dc model of GaAs optical field effect transistor considering ion-implanted profile.” [2]. S.N. Chattopadhyay, N. Motoyama, A.Rudra, A.Sharma, S. Sriram, C.B. Overton and P. Pandey. “Optically controlled silicon MESFET modeling considering diffusion process.” Journal of semiconductor technology and science, vol.7, no.3, September, 2007. [3]. K. Balasubadra1, A. Arulmary, V. Rajamani, K. Sankaranarayanan. “Two dimensional numerical modeling and simulation of a uniformly doped GaAs MESFET photodetector.” Journal of Optical Communications 29 (2008) 4 [4]. P. Chakrabarti, Vinayak Jha, Pankaj Kalra, and Gaurav Gupta. “ Noise modeling of an optically controlled MESFET(OPFET).” Microwave and optical technology letters / vol. 33, no. 2, April 20 2002 [5] Nandita Saha Roy, B. B. Pal, and R. U. Khan, “Frequency-Dependent Characteristics of an Ion-Implanted GaAs MESFET with Opaque Gate Under Illumination”, Journal of lightwave technology, vol. 18, no. 2, February 2000. [6]. M.A. Alsunaidi_, M.A. Al-Absi “ Influence of electrode parameters on the performance of optically controlled MESFETs.” M.A. Alsunaidi, M.A. Al-Absi / Optics & Laser Technology 40 (2008) 711–715 [7]. M. Madheswaran, V. Rajamani, and P. Chakrabarti“ Quasi-twodimensional simulation of an ion-implanted GaAs MESFET photodetector.”
Velammal College of Engineering and Technology, Madurai
Page 420
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Enhancing Temporal Privacy and SourceLocation Privacy in WSN Routing by FFT Based Data Perturbation Method R.Prasanna Kumar#1, T.Ravi*2 #
Department of Computer Science and Engineering, Anna University Jothipuram, Coimbatore 1
[email protected]
*
KCG College of Technology Karapakkam, Chennai 2
[email protected]
Abstract— Wireless Sensor Network (WSN) is an emerging
technology that shows great promise for various futuristic applications both for mass public and military. The sensing technology combined with processing power and wireless communication makes it lucrative for being exploited in abundance in future. The inclusion of wireless communication technology also incurs various types of security threats. Although the content of sensor messages describing “events of interest” may be encrypted to provide confidentiality, the context surrounding these events may also be sensitive and therefore should be protected from eavesdroppers. An adversary armed with knowledge of the network deployment, routing algorithms, and the base-station (data sink) location can infer the temporal patterns of interesting events by merely monitoring the arrival of packets at the sink, thereby allowing the adversary to remotely track the spatiotemporal evolution of a sensed event. One of the most notable challenges threatening the successful deployment of sensor systems is privacy. Although many privacyrelated issues can be addressed by security mechanisms, one sensor network privacy issue that cannot be adequately addressed by network security is sourcelocation privacy. Keywords— WSN, FFT, temporal privacy, source location privacy, perturbation
INTRODUCTION Many data mining applications deal with privacy sensitive data. Financial transactions, health-care records, and network communication traffic are some examples. Data mining in such privacy-sensitive domains is facing growing concerns. Therefore, we need to develop data mining techniques that are sensitive to the privacy issue. This has fostered the development of a class of data
Velammal College of Engineering and Technology, Madurai
mining algorithms that try to extract the data patterns without directly accessing the original data and guarantees that the mining process does not get sufficient information to reconstruct the original data. This paper considers a class of techniques for privacy preserving data mining by value perturbation perturbing the data while preserving the underlying probabilistic properties. It explores FFT based perturbation-based approach , a well-known technique for masking the data using transformation noise. This approach tries to preserve data privacy by adding transformation noise, while making sure that the noise still preserves the “signal” from the data so that the patterns can still be accurately estimated. 1. TEMPORAL AND SOURCE LOCATION PRIVACY BY FFT: A BRIEF OVERVIEW 1.1 DATA PERTURBATION
Data perturbation approaches fall into two main categories, which we call the probability distribution approach and the value perturbation approach. The probability distribution approach replaces the data with another sample from the same (estimated) distribution or by the distribution itself. On the other hand, the value perturbation approach perturbs the values of data elements or attributes directly by some additive or multiplicative noise before it is released to the data miner. Some randomized methods for categorical data may also be classified under this category. One of the main problems of the traditional additive perturbation and multiplicative perturbation is that they perturb each data element independently, and therefore the similarity between attributes or observations which are considered as vectors in the original data space is not well preserved. Many distance/similarity based data mining applications are thus hurt. 1.2 LOCATION PRIVACY
Page 421
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Location privacy is an important security issue. Loss of location privacy can enable subsequent exposure of identity information because location information enables binding between cyberspace information and physical world entities. For example, web surfing packets coming out of a home in a Mesh network enable an eavesdropper to analyse the surfing habits of one family if the source location of those packets can be determined. In a wireless sensor network, location information often means the physical location of the event, which is crucial given some applications of wireless sensor networks. A wireless sensor network can be a low duty cycle network. Often, traffic has a strong correlation with a certain event at certain time. This gives big advantages to an eavesdropper since he does not need sophisticated techniques to discriminate traffic among different events. In this paper, we study the source location privacy problem under the assumption of one single source during a specific period. However, we need to point out that such a scenario can happen in a real wireless sensor network. 1.3 TEMPORAL PRIVACY Temporal privacy amounts to preventing an adversary from inferring the time of creation associated with one or more sensor packets arriving at the network sink. In order to protect the temporal context of the packet’s creation, it is possible to introduce additional, random delay to the delivery of packets in order to mask a sensor reading’s time of creation. Although delaying packets might increase temporal privacy, this strategy also necessitates the use of buffering either at the source or within the network and places new stress on the internal store-and-forward network buffers. 1.4 FAST FOURIER TRANSFORM A fast Fourier transform (FFT) is an efficient algorithm to compute the discrete Fourier transform (DFT) and its inverse. We use a Fast Fourier Transform (FFT) based method for data perturbation, and compare it with the Singular Value Decomposition (SVD) based method. The experimental results show that the FFT based method can obtain similar performance as SVD based method in preserving privacy as well as maintaining utility of the data, however, the computational time used by the FFT based method is much less than the SVD based method. We conclude that the FFT based method is a very promising data perturbation method. The most well known FFT algorithms depend upon the factorization of N, but (contrary to popular misconception) there are FFTs with O(N log N) complexity for all N, even for prime N. Many FFT algorithms only depend on the fact that is an Nth primitive root of unity, and thus can be applied to analogous transforms over any finite field, such as number-theoretic transforms.
2. RELATED WORK There exists a growing body of literature on privacy sensitive data mining. These algorithms can be divided into several
Velammal College of Engineering and Technology, Madurai
different groups. One approach adopts a distributed framework. This approach supports computation of data mining models and extraction of “patterns” at a given node by exchanging only the minimal necessary information among the participating nodes without transmitting the raw data. Privacy preserving association rule mining from homogeneous [9] and heterogeneous [19] distributed data sets are few examples. The second approach is based on data-swapping which works by swapping data values within same feature [3]. There is also an approach which works by adding random noise to the data in such a way that the individual data values are distorted preserving the underlying distribution properties at a macroscopic level. The algorithms belonging to this group works by first perturbing the data using randomized techniques. The perturbed data is then used to extract the patterns and models. The randomized value perturbation technique for learning decision trees [2] and association rule learning [6] are examples of this approach. Additional work on randomized masking of data can be found elsewhere [18]. This paper explores the third approach [2]. It points out that in many cases the noise can be separated from the perturbed data by studying the spectral properties of the data and as a result its privacy can be seriously compromised. Our primary contribution is to provide an explicit data perturbation method, based on Fast Fourier Transformation. 3. DATA PERTURBATION METHODS The desired perturbation methods must preserve the privacy, and at the same time, must keep the utility of the data after the perturbation [14]. We propose a FFT based method for data perturbation in this section, and also review the SVD based method. FFT BASED DATA PERTURBATION 3.1 We propose a FFT based method for data perturbation in this section, and also review the SVD based method. In order to perform FFT (Fast Fourier Transform), a given matrix A, whose size is M x N, must be first transformed so that the width and height are an integer power of 2. This can be achieved in two ways, scaling the matrix up to the nearest integer power of 2 or zero padding to the nearest integer power of 2. The second option was chosen in this paper to facilitate comparison with the original matrix. Suppose the A M N. padded new matrix is A A with size of M M xN There is a DC (Dominant Component) corresponding to zero frequency in Fourier Transform, which represents the average value across the whole matrix. The DC is located at F(O, O). However, in order to perform the filtering more easily on the transformed matrix, it is more easy to move the DC to the center of the transformed matrix ie. F(M/2, N/2). This can be achieved by multiplying each entry (x, y) of the original A A with size matrix by eπi(x±Y) . Now the new matrix matrix is A M N of M M xN N. After the DC is centred, the forward Fast Fourier Trans form is performed to get the transformed matrix F. With the transformed matrix F, low-pass filtering or high-pass filtering
Page 422
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
is performed to distort the transformed matrix in the frequency domain to get a filtered matrix FFF. Then, the inverse Fast Fourier Transform is performed on the filtered matrix FFF to B obtain a distorted matrix B. Because this distorted matrix B B is bigger in size than the original given matrix A, we only pick B B, which is denoted the up-left corner of the distorted matrix B by B so that the resulting matrix has the same size M x N as the original matrix. 3.2 SVD BASED DATA PERTURBATION Singular Value Decomposition (SVD) is a popular method in data mining and information retrieval. SVD based methods have been proven to be efficient in keeping both data privacy and data utility when used as data perturbation methods , however, the time complexity of performing SVD perturbation is o(n3). This limits the size of the data that SVD based methods can be applied on. 4. PERFORMANCE METRICS In this section, we describe the performance metrics used in this paper to compare the data perturbation methods.
4.1 PRIVACY PRESERVATION METRICS This measure is given by Var(X -Y) where X represents a single original attribute and Y the distorted attribute. This measure can be made scale invariant with respect to the variance of X as S = Var(X - Y)/Var(x). The above measure to quantify privacy is based on how closely the original values of a modified attribute can be estimated. We use PM to denote privacy metric in this paper, where PM is the average of the S for all attributes in a data.
4.2 DATA PERTURBATION METRICS RP is used to denote the average change of rank for all the attributes. RK represents the percentage of elements that keep their ranks of magnitude in each column after the perturbation. The metric CR is defined to represent the change of rank of the average value of the attributes. Similarly as RK, CK is defined to measure the percentage of the attributes that keep their ranks of average value after the perturbation. 4.3 UTILITY METRICS The data utility metrics assess whether a data keeps the performance of data mining techniques after data perturbation, e.g., whether the distorted data can maintain the accuracy of the data mining techniques such as classification, clustering, etc. In this paper, we choose the accuracy in Support Vector Machine (SVM) classification as the data utility metric.
test whether the classification algorithm can keep its performance on distorted data. The above data perturbation methods are also compared on their ability to protect the privacy of data. COMPARISON OF PERTURBATION METHODS
5.1
FFT
AND
SVD
BASED
DATA
Table I shows the performance of two data perturbation methods: FFT and SVD. The performance of these two methods varies with different Rank, CutOff or Cutratio parameters. Here we choose Rank to be 10 for SVD, Cutoff = 0.95 and Cutratio = 0.9 for FFT. The parameters are chosen so that the SVM classification can obtain the highest accuracy. The two methods have similar capability in preserving data privacy. They have the same value for PM, CP, and CK. Both of the methods obtain the perfect value for PM, and CK, which means they are powerful in keeping the privacy of data. FFT is a little better than SVD for RP, and SVD is a little better than FFT in RK. The accuracy of classifying the original data is 0.8125. We can get the same accuracy by using FFT. The privacy accuracy of applying SVD is slightly lower than FFT, which is 0.8114. Next we compare the time cost in perturbing the original data. It costs 8.79 seconds for SVD but only 3.24 seconds for FFT. As we have shown in previous sections, FFT excels SVD in time cost. 5.2 INFLUENCE OF CUTRATIO AND CUTOFF IN FFT In this section, we analyze the influence of parameters on the performance of FFT based data perturbation method. We use two parameters for FFT, CutOff and CutRatio. Figure 1 shows the influence of cut ratio on the some of the privacy and utility measures. Here we keep the parameter CutOff to be 0.95. CP, CK and PM are not affected by CutRatio for the data in use and CK and PM always maintain the perfect value. With the decrease of CutRatio, RK steadily goes down and the Accuracy of classification reduces too. However, RP increases and obtains the highest value when CutRatio = 0.7 and decreases afterwards. If we keep the value of CutRatio unchanged and decrease CutOff , the Cutoff are similar to what we have described. 5.3 INFLUENCE OF RANK IN SVD Figure 2 illustrates the influence of Rank on the performance in the SVD-based methods. Like with FFT, CP, CK and PM are not affected by rank to make sense of data. and PM are not affected by Rank for the data in use. With the decrease of Rank in SVD, Accuracy and RK decreases while RP increases. This means using smaller Rank can improve the ability of SVD in preserving privacy, as the data is more distorted. However, using smaller Rank may also sacrifice the utility of the data as the accuracy of classification may decrease.
5. EXPERIMENTS AND RESULTS We have conducted experiments to test the performance of the SVD based and FFT based data perturbation methods. SVM classification are applied on the original data as well as the data distorted by each method to
Velammal College of Engineering and Technology, Madurai
Page 423
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
with Singular Value Decomposition (SVD) based perturbation method. The experiments show that FFT based method is similar to SVD based method in preserving privacy and keeping utility of data. However, the computational time used by the FFT based method is much less than the SVD based method. We conclude that the FFT based method is a very promising data perturbation method.
REFERENCES [1] Hillol Kargupta and Souptik Datta Qi Wang and Krishnamoorthy Sivakumar On the Privacy Preserving Properties of Random Data Perturbation Techniques [2] R. Agrawal and R. Srikant. Privacy-preserving data mining.
Accuracy VS Cut ratio
Accuracy VS Rank
Cut ratio Vs RP
Rank Vs RP
Cut ratio Vs Rk
Rank Vs Rk
Fig 1: Influence of parameters in FFT SVD
Fig 2: Influence of parameters in
XIII. TABLE I COMPARISON OF FFT AND SVD BASED DATA PERTURBATION METHODS.
Data
P M
RP
RK
CP
C K
SV D FFT
1
918.982 6 921.536 5
0.003 6 0.010 0
5.2308 2 5.2308 2
2
In Proceeding of the ACM SIGMOD Conference on Management of Data, pages 439–450, Dallas, Texas, May 2000. ACM Press. [3] V. Estivill-Castro and L. Brankovic. Data swaping: Balancing privacy against precision in mining for logic rules. In Proceedings of the first Conference on Data Warehousing and Knowledge Discovery (DaWaK-99), pages 389 – 398, Florence, Italy, 1999. Springer Verlag. [4] A. Evfimevski, J. Gehrke, and R. Srikant. Limiting privacy breaches in privacy preserving data mining. In Proceedings of the ACM SIMOD/PODS Conference, San Diego, CA, June 2003. [5] A. Evfimevski, R. Srikant, R. Agrawal, and J. Gehrke. Privacypreserving mining of association rules. In Proceedings of the ACM SIKDD Conference, Edmonton, Canada, 2002. [6] S. Evfimievski. Randomization techniques for privacy preserving association rule mining. In SIGKDD Explorations, volume 4(2), Dec 2002. [7] S. Janson, T. L. , and A. Rucinski. Random Graphs. Wiley Publishers, 1 edition, 2000. [8] D. Jonsson. Some limit theorems for the eigenvalues of a sample covariance matrix. Journal of Multivariate Analysis,
0
Accurac y 0.8114
Tim e 8.79
0
0.8125
3.24
XIV. CONCLUSIONS XV. In this paper we propose a Fast Fourier Transform (FFT) based data perturbation method and compare its performance
Velammal College of Engineering and Technology, Madurai
[9] M. Kantarcioglu and C. Clifton. Privacy-preserving distributed mining of association rules on horizontally partitioned data. In SIGMOD Workshop on DMKD, Madison, WI, June 2002. [10] H. Kargupta, K. Sivakumar, and S. Ghosh. Dependency detection in mobimine and random matrices. In Proceedings of the 6th European Conference on Principles and Practice of Knowledge Discovery in Databases, pages 250–262. Springer, 2002. [11] K. Liu, H. Kargupta, and J. Ryan. Random projection and privacy preserving correlation computation from distributed data. Technical report, University of Maryland Baltimore County, Computer Science and Electrical Engineering Department, Technical Report TR-CS-03-24, 2003.
Page 424
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[12] D. G. Manolakis, V. K. Ingle, and S. M. Kogon. Statistical and Adaptive Signal Processing. McGraw Hill, 2000. [13] M. L. Mehta. Random Matrices. Academic Press, London, 2 edition, 1991. [14] V. S. Verykios, E. Bertino, I. N. Fovino, et al.
State-of-the-art in privacy preserving data mining. SIGMOD Record, 33:50-57, 2004. [15] S. J. Rizvi and J. R. Haritsa. Maintaining data privacy in association rule mining. In Proceedings of the 28th VLDB Conference, Hong Kong, China, 2002. [16] J. W. Silverstein and P. L. Combettes. Signal detection via spectral theory of large dimensional random matrices. IEEE Transactions on Signal Processing, 40(8):2100–2105, 1992. [17] G. W. Stewart. Error and perturbation bounds for subspaces associated with certain eigenvalue problems. SIAM Review, 15(4):727–764, October 1973. [18] J. F. Traub, Y. Yemini, and H. Woz’niakowski. The statistical security of a statistical database. ACM Transactions on Database Systems (TODS), 9(4):672–679, 1984. [19] J. Vaidya and C. Clifton. Privacy preserving association rule mining in vertically partitioned data. In The Eighth ACM SIGKDD International conference on Knowledge Discovery and Data Mining, Edmonton, Alberta, CA, July 2002.
Velammal College of Engineering and Technology, Madurai
Page 425
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Theoretical Nvestigation Of Size Effect On The Thermal Properties Of Nanoparticles K.Sadaiyandi# 1 and M.A.Zafrulla Khan#2 #1
Department of Physics, Velammal College of Engineering and Technology Madurai 625 009, India #2 Department of Physics, Udaya School of Engineering, Vellamodi Ammandivilai, Nagercoil - 629 001, India #1
[email protected], #2 [email protected]
Abstract: In the present work the size dependent Melting Temperature (Tm), Debye Temperature ( θ D ) and Mean Square Displacements (MSDs) of Tungsten nanoparticles are determined. First, the size dependent melting temperatures and Debye temperatures are calculated by following liquid drop model. The size dependent lattice contraction is properly accounted following Qi and Wang. Secondly, the size dependent melting temperatures and Debye temperatures for Tungsten nanoparticles are calculated from the recently reported molecular dynamic simulation studies on specific heat capacities. The calculated values by the two methods are agreeing well, which supports the determination of size dependent melting temperature and Debye temperature by the liquid drop model by accounting lattice contraction. The procedure is extended to determine the MSDs of Tungsten nanoparticles for sizes ranging from 2nm to 30nm at 300K. Key Words: Nanotungsten, melting temperature, Debye temperature, Mean Square Displacement, liquid drop model, lattice contraction.
are high density, hardness, melting temperature, elastic modulus and conductivity in conjunction with the low thermal expansion. The combination of these unique properties explains the diverse applications of tungsten ranging from home lighting to thermonuclear fusion first wall protection [4, 5]. With nanoscale tungsten powders available at a reasonable cost, its usage will increase greatly and a new approach is required to balance the size dependent advantages against the temperature dependent limitations. Therefore, it is of great importance to understand the thermal stability of tungsten nanoparticles for their applications at higher temperatures. The Debye Temperature ( θ D ) of nanocrystals is an essential physical quantity to characterize many material properties such as thermal vibrations of atoms and phase transitions. Also the Einstein temperature and volume expansion coefficient are related with θ D . Hence, if the size
The study of nanocrystalline materials is an active area of research in physics, chemistry and materials sciences [1,2]. While numerous techniques are known to produce nanostructures, it is much more difficult to determine the properties of such small collections of atoms. Different physical properties such as mechanical strength, plasticity, melting, sintering and alloying ability, diffusivity, chemical reactivity and the mode of crystal growth have been found depend upon particle size. Size dependent melting point depression of small particles has been studied for many years both theoretically and experimentally [3]. The technical advantages of the low melting temperatures of small particles are (i) the ability to fuse nanoparticles to form a film at a relatively low temperature, (ii) possibility of soldering at relatively low temperatures using nanoparticles, (iii) possibility of controlling the growth process of nanoparticles by controlling the deposition or substrate temperature. Tungsten along with its alloys and compounds, occupies a unique position in material science. The material properties that make tungsten attractive to the metals industry
effect on θ D is known, the size effect on other related properties can be easily determined. As experimental and theoretical investigations in this direction are very less, it is planned to determine the size effect on the Debye Temperature ( θ D ) and Mean Square Displacements (MSDs) of tungsten nanoparticles. Melting temperature (Tm) of nanogold with particle sizes in the range 20 to 50 Å is studied experimentally and theoretically by Buffat and Borel [6]. They proposed a phenomenological model to determine the melting temperature of particles with different sizes. Also Jiang. et. al [7] introduced a simple model, free of any adjustable parameter, for the size dependent melting of nanocrystals. Recently Nanda et. al. [8] derived an expression for the size dependent melting of low dimensional systems following the liquid drop model [9]. Although the results from [6,7,8] are in good agreement with experiments they have not properly accounted the size dependent lattice contraction. Recently Sadaiyandi [10] calculated the melting temperatures and Debye temperatures of Au, Ag and Al nanoparticles from liquid drop model proposed by Nanda et. al [8] by properly accounting the size dependent lattice contraction. As the idea
Velammal College of Engineering and Technology, Madurai
Page 426
I.
INTRODUCTION
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
proposed by Sadaiyandi [10] is simple and successful, it is planned to determine the melting temperatures of tungsten nanoparticles by the same method. From the calculated values of melting temperatures, the size dependent Debye temperatures and MSDs are determined by the Lindemann theory [11]. Recently Amitava Moitra et al [12] reported the size dependent specific heat capacities of tungsten nanoparticles at different temperatures by molecular dynamic simulation. From the reported values of specific heat capacities the size dependent Debye temperatures and melting temperatures are determined by the idea proposed by Reisland [13]. The calculated values are in good agreement with the values arrived from the first method [10] which supports the determination of melting temperatures of nanoparticles by liquid drop model by considering the lattice contraction.
II.
LIQUID DROP MODEL FOR SIZE DEPENDENT MELTING One of the successes of the liquid drop model lies in providing an intuitive explanation for the phenomenon of spontaneous fission of some nuclei. Atomic clusters and nanoparticles being finite systems, their properties are dominated by the surface atoms, there fore their binding energy can be effectively represented by the volume and surface dependent term as in the liquid drop model. From this point of view the melting of atomic clusters and nanoparticles can be understood by scaling the cohesive energy to the melting temperature. Based on the liquid drop model [9], Nanda et al [8] derived an empirical relation for the size dependent melting temperature of nanoparticles (Tm) in terms of bulk melting temperature (Tmb) as,
Tm = Tmb −
6v 0 γ 0.0005736d
Tm β = 1− Tmb d
(1)
(2)
where v0 is the molar volume, d is the diameter of the particle and γ represents the coefficient of surface energy of the material. Here the particle shape is assumed to be spherical. For a nanomaterial with general shape, the size dependent melting, in general, can be written as
Tm β = 1− Tmb zd
(3)
with z = 1, 3/2, and 3 for nanoparticles , nanowires, and thin films. In equation (3), d represents the diameter in case of nanoparticles and nanowires, whereas it represents the thickness in case of thin films. Similar expressions for size dependent melting for spherical nanoparticles has also been derived from thermodynamic arguments [6, 14] and from a model based on
Velammal College of Engineering and Technology, Madurai
surface phonon instability [15]. Using the known values of vo,
γ and Tmb, the values of β for different elements are estimated by Nanda et al [8] and are tabulated in Table-1. Nanda et al [8] compared their results with the theoretical predictions of Jiang et al [16] and found that both the models [8,16] are consistent for Pb thin films. 1) III.
SIZE DEPENDENT LATTICE PARAMETERS It is proved experimentally and theoretically that the lattice parameters of nanoparticles depend on the particle size [17-21]. For an isolated nanoparticle, the lattice constants are often measured to contract. Montano et al [22] measured the nearest neighbor distance of isolated silver particles with sizes from 2.5 nm to 13 nm and found a noticeable contraction in the nearest neighbor distance. Also, the model developed by Jiang et al [23, 24] based on the Laplace – Young equation supports the size induced lattice contraction. Recently Qi and Wang [25] developed a simple model to account for the size and shape dependent lattice parameters of metallic nanoparticles. They have assumed that a nanoparticle is formed in the three steps. First a particle is taken out from an ideal bulk crystal without changing structure. Secondly the surface tension contracts the particle little elastically and thirdly a nanoparticle is formed in equilibrium. By minimising the sum of the increased surface energy and the elastic energy, Qi and Wang obtained a formula to calculate the lattice parameters of metallic nanoparticles as
Δa 1 =− a 1 + Kd where,
K=
α G γ
(4)
(5)
here, a - is the lattice constant , α – is the shape factor and is equal to one for a spherical particle and G is the shear modulus. Qi and Wang [25] found that the values obtained by them for Au, Ag and Al are in good agreement with the experimental values [26-28]. Also Molecular dynamic simulation studies on the size and shape dependent lattice parameters of unsupported small Pt nanoparticles [29] conclude that for transitional metallic nanoparticles, size is the main factor, which affects the lattice parameter.
IV. SIZE DEPENDENT MELTING TEMPERATURE CONSIDERING LATTICE CONTRACTION In the present work, the size dependent lattice constants for the tungsten nanoparticles are TABLE I SURFACE ENERGY ( γ ), BULK MELTING TEMPERATURE (Tmb), BULK MOLAR VOLUME (vo) AND ( β ) FOR TUNGSTEN [8]
Page 427
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
γ
J/m2 at 298 K 2.753
Tmb K
vo x 10-6 m3
3680
9.53
β
nm
0.772
determined by equation (4). As the model proposed by Qi and Wang [25] assumes that a particle is taken from a bulk crystal without changing the structure, for the formation of nanoparticle, one can imagine that a nanoparticle can have the same structure as that of its bulk counter part. Also the contraction in the lattice constant will lead to the change in molar volume of the nanoparticle [10]. Hence from the corrected values of lattice constants obtained from equation (4), new molar volumes are determined for sizes ranging from 2nm to 30nm. The size dependent melting temperatures for tungsten nanoparticles are determined from equation (1), by incorporating the volume contraction and the results along with the results obtained from specific heat calculations are shown in fig-1.
disturb them and the melting process initiates. Quantitative calculations based on the model are not easy, hence Lindemann offered a simple criterion that the melting might be expected when the root mean square displacement exceeds a certain threshold value ( namely when the amplitude reaches at least 10% of the nearest neighbor distance). Although experiments confirm that the Lindemann criterion is not very accurate, it lends strong support to the idea that the magnitude of vibrational amplitude dominates the phenomenon of melting. Also Lindemann model for vibrational melting is adequate for simplest structures i.e. assemblies of closed packed atoms. Further, there are some discrepancies regarding the Tm dependence of θ D . One opinion is that θ D varies linearly with Tm [31] and the other suggests a square root dependence of θ D on Tm according to Lindemann Criterion of melting [11]. Since there are more supporting evidences for the square root dependence [16], the size dependent θ D and MSDs are determined by the Lindemann theory. Following Lindemann theory, Reisland [13] derived the expressions for the Debye temperature and mean square displacement for a bulk crystal as
M elting temperatu re (K)
4000
⎛ ⎜ T θ D = C Lind ⎜ m2 ⎜ mv 0 3 ⎝
3600 3200 Liquid Drop Model From Sp.ht
2800 2400 2000 0
0.2
0.4
0.6
1/size (nm)
Fig.1 Size dependent melting temperature for nanotungsten. It can be seen that the melting temperatures determined from the idea proposed by Sadaiyandi [10] and the melting temperatures calculated from specific heat data [13] are agreeing well.
u2 =
9h 2 T mk B θ 2D
1
⎞2 ⎟ ⎟ ⎟ ⎠
(6)
(7)
where m is the atomic mass and CLind is the Lindemann constant and is equal to 200 for nonmetals and 137 for metals. T is the temperature at which the MSDs are calculated. Studies on the lattice structure of nanocrystalline samples by means of XRD and Massbauer spectroscopy showed that the lattice structure of the nanometer sized crystallites deviates from the equilibrium state. The deviation may be classified as (i) distorted lattice structures in pure elements and stoichiometric line compounds , and (ii) formation of metastable phase below a critical crystalline size. The lattice distortion in various nanocrystalline materials is manifested by a significant change in lattice parameter [32] and corresponding change in molar volume. Hence the expression for θ D of a bulk crystal can be used for a nanocrystalline material by substituting the size dependent molar volume and melting temperature. Determination of size dependent θ D from specific
V. DEBYE TEMPERATURE AND MEAN SQUARE DISPLACEMENTS The forces between the atoms are reflected in the Debye temperature and it is useful to have this as a reference to characterise a crystal. The Debye temperature is a measure of the vibrational response of the material and therefore intimately connected with properties like the specific heat, thermal expansion and vibrational entropy [30]. Lighter inert gas solids melt below their θ D while the other crystals remain solids above it. The first theory explaining the mechanism of melting in the bulk was proposed by Lindemann [11], who used vibration of atoms in the crystal to explain the melting transition. The average amplitude of thermal vibrations increases when the temperature of the solid increases. At some point the amplitude of vibration becomes so large that the atoms start to invade the space of their nearest neighbors and
heat measurements by Herr et al [33] showed that θ D decreases significantly for nanocrystalline materials. The depressed Debye temperature in nanocrystalline sample implies a decrease in the cohesion of atoms in the nanocrystallites, which agrees well with the measured grain
Velammal College of Engineering and Technology, Madurai
Page 428
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VI. SIZE DEPENDENT MELTING TEMPERATURE AND DEBYE TEMPERATURE FROM SPECIFIC HEAT MEASUREMENTS The specific heat capacity Cv of a solid according to Debye approximation is given as [13], x
4
where
x =
1.3E-02 1.2E-02 1.1E-02 1.0E-02 9.0E-03 8.0E-03 7.0E-03 6.0E-03 0
x
x e C v = 9 Nk B x ∫ x dx 2 0 (e − 1) 3
1.4E-02
M S D x 10 -22 m 2
size dependence of lattice parameter [34]. In the present work the size dependent Debye temperatures are determined from equation (6) by substituting the newly calculated melting temperatures and the size dependent MSDs are determined from equation (7) for tungsten nanoparticles. The results along with the results obtained from specific heat calculations are shown in fig-2 and fig-3.
0.1
0.2
0.3
0.4
0.5
0.6
1/size (nm)
(8)
Fig.3 Size effect on the MSDs of tungsten nanoparticles at 300K. The MSD values are increasing with respect to the reduction in size. Also it can be seen that the size effect is dominant below 10nm.
θ D (T )
(9)
T
where, N- is the Avogadro Number and kB is the Boltzmann constant and θ D (T) is the Debye temperature at a particular temperature. The right hand side of eqn (8) is determined by Simpson’s rule for different values of x (from 1 to 100). For improving accuracy the intervals are taken as smaller. A permanent tabulation is formed connecting the x values and right hand side of eqn.(8).
290
The values of specific heat capacities are collected from [12] for tungsten nanoparticles with sizes 8nm, 10nm and 12nm at different temperatures. The x value corresponding to the collected specific heat at a particular temperature is obtained from the permanent tabulation. Knowing the values of x and T, the corresponding values of θ D (T) are calculated from eqn (9). Now a graph is drawn connecting T and θ D (T). When the curve is extrapolated to meet the Y axis (i.e. T = 0K), the Y intercept gives the Debye temperature of the solid. In order to improve the accuracy, the Y intercept is determined by second order least square fitting method. The fitting curves are shown in fig.4, fig.5 and fig.6 for tungsten with particle size 8nm, 10nm and 12nm.
270
20000
260 Liquid drop model From Sp.Ht.
250 240 230 220 210 200 0
0.1
0.2
0.3
1/Size
0.4
0.5
0.6
Debye Tem perature (K)
Debye Tem perature (K)
280
16000 12000 8000 4000
(nm)
Fig.2 Size dependent Debye temperature for nanotungsten. It can be seen that the Debye temperatures determined from the idea proposed by Sadaiyandi [10] and the Debye temperatures calculated from specific heat data [13] are agreeing well.
0 0
1000
2000
3000
4000
Temperature (K)
Fig.4 Debye temperature for nanotungsten with size 8nm. Points are the calculated values. The solid line is the fitting curve. The fitted value of
θD
is
271K
Velammal College of Engineering and Technology, Madurai
Page 429
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VII.
RESULTS AND DISCUSSION Determination of size dependent
θ D from specific heat measurements by Herr et al [33] showed that θ D decreases significantly for nanocrystalline materials. The depressed Debye temperature in nanocrystalline sample implies a decrease in the cohesion of atoms in the nanocrystallites, which agrees well with the measured grain size dependence of lattice parameter[34]. 20000
D e b y e T e m p e r a tu r e (K )
18000 16000 14000 12000 10000 8000 6000 4000 2000 0 0
1000
2000
3000
4000
fig.2, it can be seen that the values of melting temperatures and Debye temperatures obtained by the two methods are agreeing well, which shows that the size dependent melting temperatures can be determined by liquid drop model proposed by Nandha et al [8] by considering the idea of size dependent lattice contraction proposed by Qi and Wang [25]. Also fig.1 shows that the melting temperature and Debye temperature of Tungsten decreases with the reduction in size which is in agreement with the earlier predictions [33]. Fig.3 shows the size dependent MSDs at 300K for nanotungsten. It shows that the MSD values are increasing with reduction in size and also it can be seen that the size effect on MSDs is predominant in the sizes ranging from 2nm to 10nm. When the size is above 20nm the changes are too small to be neglected. CONCLUSION VIII. The size effect on θ D for tungsten nanoparticles is determined (i) by the liquid drop model considering the size dependent lattice contraction [10] and (ii) from the available specific heat data [12, 13]. The calculated values by the two methods are agreeing well, which supports the determination of size dependent melting temperature and Debye temperature by the liquid drop model by accounting lattice contraction. The procedure is extended to determine the MSDs of Tungsten nanoparticles for sizes ranging from 2nm to 30nm at 300K.
Temperature (K)
Fig.5 Debye temperature for nanotungsten with size 10nm. Points are the calculated values. The solid line is the fitting curve. The fitted value of
θD
is
273.5K.
2) ACKNOWLEDGEMENT The author wishes to acknowledge the management of Velammal College of Engineering and Technology, Madurai and Udaya School of Engineering, Nagercoil for constant encouragement.
18000
3) REFERENCES
Debye Temperature (K)
16000 14000 12000 10000 8000 6000 4000 2000 0 0
1000
2000 Temperature (K)
3000
4000
Fig.6 Debye temperature for nanotungsten with size 12nm. Points are the calculated values. The solid line is the fitting curve. The fitted value of
θD
is
278K.
In the present work the size dependent melting temperature and Debye temperature are determined by liquid drop model and specific heat calculations. From fig.1 and
Velammal College of Engineering and Technology, Madurai
[1] A.S.Edelstein and R.C.Cammarata, Nanomaterials: Synthesis, Properties and Applications ( Inst. Phys., Bristol, 1996) [2] H.Gleiter, Prog. Mater Sci. 33, 223 (1989) [3] M.Takagi, J.Phys.Soc.Jpn. 9, 359 (1954) [4] S.W.H.Yih and C.T.Wang , Tunsten : Sources, Metallurgy, Properties and Applications , (Plenum, New York1979). [5] E.Lassner and W.D.Schubert, Tungsten : Properties, Chemistry, Technology of the Element , (Plenum, New York 1999) [6] Ph.Buffat and J.P.Borel, Phys.Rev.A 13(6), 2287 (1976) [7] Q.Jiang, N.Aya, F.G.Shi, Appl. Phys. A 64, 627 (1997) [8] K.K.Nanda, S.N.Sahu and S.N.Behera, Phys Rev. A 66, 013208 (2002) [9] C.Brechignac, et al., J.Chem. Phys. 101, 6992 (1994) [10] K.Sadaiyandi, Mat. Chem. Phys 115, 703 (2009) [11] F.A.Lindemann Phys.Z. 11, 609 (1910) [12] A.Moitra, S.Kim, J.Houze, B.Jelinek, S.G.Kim, S.J.Park, R.M.German and M.F.Horstemeyer, J.Phys.D: Appl. Phys, 41, 185406 (2008) [13] J.A. Reisland, The Physics of Phonons, Wiley, London1973. [14] P.R.Couchman and W.A.Jesser, Nature(London) 269, 481 (1977) [15] M.Wautelet, J.Phys.D. 24, 343 (1991) [16] Q.Jiang, H.Y.Tong, D.T.Hsu, OK.Okuyama and F.G.Shi Thin Solid Films 312, 357 (1998) [17] R.Lamber, S.Warjen and I.Jaeger Phys. Rev. B. 51, 10968 (1995) [18] X.F.Yu, X.Liu, K.Zhang and Z.Q.Hu J.Phys.C 11, 937 (1999) [19] C.Solliard and M.Flueli Surf. Sci. 156, 487 (1985)
Page 430
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[20] V.I.Zubov, I.V.Mamoatov and J.N.T. Rabelo Nanostructure.Mater. 8, 595 (1997) [21] M.Fukuhara Phys. Lett. A 313, 427 (2003) [22] P.A.Montana, W,Schulze, B.Tesche, G.K.Shenoy and T.I.Motrison Phys. Rev.b 30, 672 (1984) [23] Q.Jiang, L.H.Liang and D.S.Zhao J.Pjys.Chem B 105, 6275 (2001) [24] L.H.Liang, J.C.Li and Q.Jiang Physica B 334, 49 (2003) [25] W.H.Qi and M.P.Wang J.Nanopart. Res. 7, 51 (2005) [26] C.W.Mays, Vermank and D.KuhlmannWilsdorf, Surf. Sci. 12, 134 (1968) [27] H.J. Wassermanand J.S.Vermaak Surf. Sci. 22 164 (1970) [28] J. Woltersdorf, A.S.Kepijiko and E.Pippel. Surf.Sci. 106, 64 (1981) [29]W.H.Qi, B.Y.Huang, M.P.Wang, Z.M.Yin and J.Li, J.Nanopart.Res.11(3) 579 (2009) [30] G.Grimvall. Thermophysical Properties of Materials, (North Holland, Amsterdam, 1956) [31] R.C.G.Killean and E.J.Lisher, J.Phy.C: Solid State Phys. 8, 3510 (1975) [32]K. Lu and Y. H. Zhao, Nanostruct. Mater. 12, 559 (1999) [33] U.Herr, M.Geigl and K.Samwer, Philos. Mag. A77, 641 (1998) [34] B.S.Murty, M.K.Datta and S.K.Babi. Sadhana 28, 23 (2003)
Velammal College of Engineering and Technology, Madurai
Page 431
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Dynamic Key Management to minimize communication latency for efficient group communication Dr.P.Alli 1 , G.Vinoth Chakkaravarthy 2 , R.Deepalakshmi 3 1
Department of Computer Science and Engineering, Velammal College of Engg.&Tech,Tamilnadu , India Department of Computer Science and Engineering, Velammal College of Engg.&Tech,Tamilnadu , India 3 Department of Computer Science and Engineering, Velammal College of Engg.&Tech,Tamilnadu , India 1 [email protected], 2 [email protected], 3 [email protected] 2
Abstract:
Introduction:
Group key management is one of the basic building blocks in securing Group Communication (GC), the majority of research in this area mainly concerned with increasing the security while minimizing cryptographic computation cost. Due to the reliable GC platform, network delay is amplified by necessary acknowledgments between the group members. The bottleneck is shifted from computation to communication latency. In this paper, a new Virtual Binary Tree Technique (VBT) to specific group key agreement is proposed which supports dynamic group membership and handles network failures. VBT communication-efficient and secure against hostile eavesdroppers and various attacks specific to group settings. In VBT, common group key is generated based on the share taken from each participant and transmitted to all participants in the team. The member, who is having the group key, is allowed to communicate with other members in the group. Communication latency occurs while authentication is carried out every time between the team of users is reduced. To make this scheme as a dynamic one, when a new user arrives, a new shared secret key is formed , when a person leaves from the group , their share is taken from the existing shared secret key and new shared secret key is formed based on the contents of existing participants and it is broadcasted. So the person who left from the team is prevented to participate in the communication .Thus VBT assures the confidentiality of the team with minimum communication latency.
Security is an important requirement in reliable group communication over open networks in order to prevent intruder attack. The majority of research in group key management was mainly concerned with increasing the security while minimizing cryptographic computation cost. But heavyweight computation such as large number arithmetic that forms the basis of many modern cryptographic algorithms is the greatest burden imposed by security protocols. However, the continuing increase in computation power of modern workstations speed up the heavyweight cryptographic operations. In contrast, communication latency has not improved appreciably. Due to the reliable group communication platform, network delay is amplified by necessary acknowledgments between the group members [2]. However, the continued advances in computing power have not been matched by a decrease in network communication delay. Thus, communication latency, is increasingly dominating the key setup latency, replacing computation delay as the main latency contributor. Hence, there is a need to minimize the size of messages and especially the number of rounds in cryptographic protocols.
Keywords: communication latency, group key, dynamic membership, member join, member leave, Diffie-Hellman key exchange , confidentiality
Velammal College of Engineering and Technology, Madurai
The bottleneck shift from computation to communication latency prompts us to look at cryptographic protocols in a different light: allowing more liberal use of cryptographic operations while attempting to reduce the communication overhead. The proposed scheme in this paper is efficient to avoid such latency with out minimizing cryptographic operations. We consider a protocol, is formed by a virtual tree based approach that extends the 2-party Diffie-Hellman key exchange and supposes the formation of a secure static group[2].This protocol involves heavy computation and communication requirements: O(n) communication rounds and O(n) cryptographic operations are necessary to establish a shared key in a group of ‘n’ members. We extend it to deal
Page 432
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
with dynamic groups and network communication-efficient manner
failures
in
a
Cryptographic Properties: The following cryptographic properties is assured in the group key formation[2].
Group Key Secrecy: It guarantees that it is computationally infeasible for a passive adversary to discover any group key. Forward Secrecy: It guarantees that a passive adversary who knows a contiguous subset of old group keys cannot discover subsequent group keys. Backward Secrecy: It guarantees that a passive adversary who knows a contiguous subset of group keys cannot discover preceding group keys. Key Independence: It guarantees that a passive adversary who knows any proper subset of group keys cannot discover any other group key not included in the subset. Backward and Forward Secrecy properties assume that the adversary is a current or a former group member. From our definition of shared secret key , it would be more desirable to guarantee that any bit of the group key is unpredictable. Thus group key secrecy guarantees that it is computationally infeasible for a passive adversary to distinguish any group key from random number.
Proposed tree based group key generation: The proposed tree structure for group communication [6] is shown in figure: 1. Here all the nodes will be allotted with a value. The leaf nodes are members Ml,M2.....Mn. A member's value is obtained from root node to its member leaf node. By evaluating the values in the nodes, group key is formed for communication. Then the key is broadcasted across the tree structure to all its nodes. In a proposed scheme called Group Key Management Protocol (GKMP) [1] , the entire group shares the same key called as session-encrypting key (SEK).
Figure 1. Tree for Group key formation The tree has two types of nodes: leaf and internal. Each leaf node is associated with a specific group member. An internal node IN(i) always has two children: another (lower) internal node IN(i-1) and a leaf node LN(i). The exception is IN(1) which is also a leaf node corresponding to M1. Each leaf node LN(i) has a session random ri chosen and kept secret by Mi. The blinded version thereof is bri = α ri mod p. Every internal node IN(j) has an associated secret key kj and a public blinded key (bkey) bkj = α kj mod p. The secret key ki (i > 1) is the result of a Diffie-Hellman key agreement between the node’s two children (k1 is an exception and is equal to ri.), which is computed recursively as follows: ki = (bki-1)ri mod p = (bri)ki-1 mod p = α ri ki-1 mod p if i > 1. The group key with the root node is : K4 = The root (group) key is never used directly for the purposes of encryption. Instead, such special-purpose subkeys are derived from the root key, e.g., by applying a cryptographically secure hash function to the root key. All bkeys (bki) are assumed to be public. Group Key Agreement: A comprehensive group key agreement solution must handle adjustments to group secrets subsequent to all membership change operations in the underlying group communication system. The following membership changes are considered: Single member changes include member join or leave, and multiple member changes include group merge and group partition. Join occurs when a prospective member wants to join a group [6]. Leave occurs when a member wants to leave (or is forced to leave) a group[6].
Member joins: Consider, initially the group has n users {M1,M2…. Mn}. When the group communication system announces the arrival of a new member, both the new member and the prior group members receive this notification simultaneously. The new
Velammal College of Engineering and Technology, Madurai
Page 433
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
member Mn+1 broadcasts a join request message that contains its own bkey bkn+1 Upon receiving this message, the current group’s sponsor Mn refreshes its session random, computes brn, kn, bkn and sends the current tree BT(n) to Mn+1 with all bkeys. Next, each member Mi increments n = n + 1 and creates a new root key node IN(n) with two children: the root node IN(n-1) of the prior tree Ti on the left and the new leaf node LN(n) corresponding to the new member on the right. Note that every member can compute the group key since: • All existing members only need the new member’s blinded session random. The new member needs the blinded group key of the • prior group. Step 1: The new member broadcasts request for join Mn+1
brn+1 = α r n+1
event from the group communication system, each remaining member updates its key tree by deleting the nodes LN(d) corresponding to M d and its parent node IN(d) The nodes above the leaving node are also renumbered. The former sibling IN (d-1) of M d is promoted to replace (former) Md’s parent. The sponsor Ms selects a new secret session random, computes all keys (and bkeys) just below the root node, and broadcasts BT(s) to the group. This information allows all members (including the sponsor) to recompute the new group key. Step 1: Every member • updates key tree as described above, • removes all keys and bkeys from the sponsor node to the root node
C = { M1, …. Mn }
Step 2: Every member updates key tree by adding new member node and • new root node, removes bkn, • The sponsor Mn additionally • generates new share rn and computes brn, kn, bkn broadcasts updated tree BT(n) •
The sponsor Ms additionally generates new share and computes all (keys, bkeys) • • and broadcasts updated tree BT(s) C – { Md }
BT(s)
Ms
Step 2: Every member computes the group key using BT(s)
C U { M n+1 } = {M1 , ….., M n+1 } BT(n) Mn Step 3: Every member computes the group key using BT(n)
Figure 3. Leave operation The leave protocol provides forward secrecy since a former member cannot compute the new key owing to the sponsor’s changing the session random. The protocol also provides key independence since knowledge of the new key cannot be used to derive the previous keys; this is, again, due to the sponsor refreshing its session random. Figure 2.Tree update in join
Member Leaves: From a a group of n members when a member M d (d ≤ n) leaves the group. If d > 1, the sponsor Ms is the leaf node directly below the leaving member, i.e., M d-1. Otherwise, the sponsor is M2. Upon hearing about the leave
Velammal College of Engineering and Technology, Madurai
Performance analysis and communication efficiency: The above VBT key management system is analyzed with an existing key management schemes TGDH .The overhead of the TGDH protocol depends on the tree height, the balancedness of the key tree, the location of the joining tree, and the leaving nodes. The number of modular
Page 434
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
exponentiations for a leave event in the VBT depends on the location of the deepest leaving node.
Map",Information and Communications University (ICU), 584, Hwaam-Dong, Yuseong-gu, Daej eon, 305-732, Korea.
Conclusion:
[2] Yongdae,Perrig and Tsudik, "Group Key Agreement efficient in communication", IEEE transactions on computers, Oct 2003.
The proposed key management scheme VBT instantly generates a key. The key is computed as a function of all current group members participates in the communication. In this scheme, the currently operating valid users only can take part in communication and it ensures the presence of users without much difficulty. This scheme supports all dynamic peer group operations: join, leave. Furthermore, it easily handles cascaded (nested) membership events and network failures. Hence the proposed VBT ensures that the efficient communication between the groups with less delay with out minimizing cryptographic computations.
[3] Yair Amir, Yongdae Kim, Cristina Nita-Rotaru, and Gene Tsudik. “On the performance of group key agreement protocols” Technical Report CNDS-2001-5, [4] David A. McGrew and Alan T. Sherman, " Key Establishment in Large Dynamic Groups Using One- Way Function Trees" Cryptographic Technologies Group, TIS Labs at Network Associates, May 20, 1998. [5] C.Becker and U.Wille. "Communication complexity of group key distribution”. In 5th ACM Conference on Computer andCommunications Security, November 1998.
References: [1] Sang won Leel, Yongdae Kim2, Kwangjo Kiml "An Efficient Tree-based Group Key Agreement using Bilinear
[6] Dr.K.Duraiswamy, S.P.Shantharajah “Key Management and Distribution for Authenticating Group Communication” IEEE transactions on computers,2006
Velammal College of Engineering and Technology, Madurai
Page 435
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Integrated Biometric Authentication Using Finger Print And Iris Matching A.Muthukumar1 and S.Kannan2
1. Faculty, Dept of ECE, Kalasalingam University,[email protected] 2. Faculty, Dept of EEE, Kalasalingam University,[email protected] 1
[email protected]
Abstract A biometric measures an individual's unique physical or behavioral characteristics in order to recognize or authenticate his identity. Biometric identification is made up of two stages: enrolment and verification. Automatic and reliable extraction of minutiae from fingerprint images is a critical step in fingerprint matching. The quality of input fingerprint images plays an important role in the performance of automatic identification and verification algorithms. Error-correcting codes are proposed as a means of correcting iris readings for authentication purposes. This paper presents a technique that combined with a bimodal biometric verification system that makes use of finger print images and iris images. Each individual verification has been optimized to operate in automatic mode and designed for security and authentication access application. 1. INTRODUCTION Fingerprint based identification has been one of the most successful biometric techniques used for personal identification. Each individual has unique fingerprints. A fingerprint is the pattern of ridges and valleys on the finger tip. A fingerprint is thus defined by the uniqueness of the local ridge characteristics and their relationships. Minutiae points are these local ridge characteristics that occur either at a ridge ending or a ridge bifurcation[1]. A ridge ending is defined as the point where the ridge ends abruptly and the ridge bifurcation is the point where the ridge splits into two or more branches. Automatic minutiae detection becomes a difficult task in low quality fingerprint images where noise and contrast deficiency result in pixel configurations similar to that of minutiae. This is an important aspect that has been taken into consideration in this project for extraction of the minutiae with a minimum error in a particular location.
Velammal College of Engineering & Technology, Madurai
Fig 1: Finger print and Iris In this paper, error-correcting codes are proposed as a means of correcting iris readings for authentication purposes. We store a hashed version of an individual's biometric template (along with check digits) to verify their identity. A major advantage of this approach is that it is unnecessary to store the individual's biometric template explicitly, a situation which puts the biometric at risk from dishonest systems operators. Given that the potential theft of biometrics is one of the main objections to their usage, any such increase in security is highly desirable. If a valid message is defined to be a codeword, then a code is the set of all code words. The code distance, d, is the minimum number of bits which differ between any two code words. If d=2t+ 1, (1) then up to 2t errors can be detected and up to t of these corrected; if there are t errors or less, the corrupted codeword is
Page 436
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.1.1.Pre-Processing IRIS
Fin ger prin t
Edge detect ion
Norm alisati on
Norm alisat ion
Frequ ency and Orient ation
Iris Matc hing using Corre i Mask Gener ation and Match
Comb ine Iris and
The input image is segmented from the background which ensures the removal of noise. For this, the whole image is divided into blocks of size 16×16 and the variance of each block is computed. The variance is then compared with the threshold value. If the variance of a block is less than the threshold value, then it is deleted from the original figure. This process is carried out for the whole image. The image obtained from the above step is then normalized to get the desired variance of the given image. The normalized image is given by
Fig 2: Overview of the proposed work automatically closer to one valid codeword than any other, but if there are more than t errors, there is no way of knowing to which codeword one should correct[3]. Therefore, a good code has as large a code-distance as possible, as this enables more errors to be corrected. Each codeword in the code is of the same length, usually represented by n. It is divided into k information bits and (n - k) check digits. Such a code is known as a systematic (n, k) code. As suggested by their name, check digits are used to correct the corrupted codeword to its original value. Finger print with false minutiae removal was studied for enhancement of image[1],[2].Then iris with error correcting codes used for improving the image enhancement and edge detection of the image[3],[4]. 2.PROPOSED WORK
⎧M0 + v(VAR0(I(i, j) - M)2 /(VAR)) ⎪ if I(i, j) > M ⎪ G(i, j) = ⎨ ⎪ M0 - v(VAR0(I(i, j) - M)2 /(VAR)) ⎪⎩ otherwise
where I(i,j) denotes the gray-level value at pixel (i, j), M and VAR denote the estimated mean and variance of I respectively and G(I,j) denotes the normalized gray-level value at pixel (i, j). and VAR0 are the desired mean and variance values respectively. The estimation of the orientation of the image is then carried out as the next step. The whole image is divided into blocks of size 16×16 and the local orientation in the figure is computed by
The whole problem is divided into the following three steps: 1. Pre-Processing 2. Normalization 3. Post-Processing(matching) This fig.2 shows the details of the first step named preprocessing which gives an insight into the process that has been followed by the enhancement of the inputs fingerprint and iris images.
2.1FINGER PRINT MATCHING In this section , image is segmented and it is normalized using gabor filtering are explained. Then in PostProcessing ,how the false minutiae removed was explained.
Velammal College of Engineering & Technology, Madurai
Vx(i,j) = Vy(i,j) =
i +ω/2
j+ω/2
u =i -ω/2
v = j -ω/2
i +ω/2
j+ω/2
∑
∑ 2∂x(u, v)∂y(u, v)
∑ω
∑ω (∂ x(u, v) − ∂ y(u, v))
u =i - /2
2
2
v = j - /2
θ (i, j) = 0.5 tan (Vx(i,j)/Vy(i,j)) −1
where θ(i, j) is the least square estimate of the local ridge orientation at the block centered at pixel (i, j). The angles between the blocks are then smoothened by passing the image through a low pass filter as follows.
X [k ] =
1 ω −1
∑ G(u, v), k = 0,1,...l − 1
ω d =0
Page 437
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 u=i+(d-(ω/2))cosO(i,j)+(k-(l/2))sinO(i,j), v=i+(d-(ω/2))sinO(i,j)+((l/2)-k)cosO(i,j), The following method is adopted for the calculation of the frequency of the local blocks. X-signatures of each block are computed along the direction perpendicular to the orientation angle in each block. The window used for this purpose is of size 16×32. The frequency is then computed by the distance between the peaks obtained in the X-signatures. The window for this is given by the formula
X [k ] =
1 ω −1
∑ G(u, v), k = 0,1,...l − 1
ω d =0
u=i+(d-(ω/2))cosO(i,j)+(k-(l/2))sinO(i,j),
2.1.3Post-Processing The minutiae points obtained in the above step may contain many spurious minutiae. This may occur due to the presence of ridge breaks in the given figure itself which could not be improved even after enhancement. This results in false minutiae points which need to be removed. These unwanted minutiae points are removed in the post-processing stage. False minutiae points will be obtained at the borders as the image ends abruptly. These are deleted using the segmented mask. As a first step, a segmented mask is created. This is created during segmentation carried out in the stage of preprocessing and contains ones in the blocks which have higher variance than the threshold and zeros for the blocks having lower variance. Finally after enhancing and false removal of the minutiae, then finger print is matched with the template to give the person is authenticated or not. This is first matching output of this paper.
v=i+(d-(ω/2))sinO(i,j)+((l/2)-k)cosO(i,j),
2.2.IRIS MATCHING
In general, the frequency of image constitutes has a certain frequency for the hole image and hence the above step can be omitted if the global frequency of the given figure is known.
In this section , image is segmented and it is normalized using gabor filtering are explained. Then in PostProcessing ,how the error correcting are added and how the matching was occurs was explained.
As the next step, each block is filtered along the direction of the orientation angle using the value of the frequency obtained for each block. A Gabor filter is used for this process and a suitable value of local variances is taken for carrying out the process of filtering. A Gabor filter takes care of both the frequency components as well as the spatial coordinates. The inputs required to create a Gabor mask are frequency, orientation angle and variances along x and y directions. Filtering is done for each block using the local orientation angle and frequency. Pre-processing of the image is completed by the steps as mentioned and the enhanced image is obtained. 2.1.2.Normalization The next step after enhancement of the image is the extraction of minutiae. The enhanced image is binarised first in this step. The skeleton of the image is then formed. The minutiae points are then extracted by the following method. The binary image is thinned as a result of which a ridge is only one pixel wide. The minutiae points are thus those which have a pixel value of one (ridge ending) as their neighbor or more than two ones (ridge bifurcations) in their neighborhood. This ends the process of extraction of minutiae points.
Velammal College of Engineering & Technology, Madurai
2.2.1.Pre-Processing & Normalization In this paper, an improved means of matching iris readings was proposed; however, prior to this stage, the segmentation and image processing stages must take place and in this paper, use of Daugman's concept was strongly recommended algorithm for this purpose. Thus, an Iris Code and corresponding mask are produced, but we do not subsequently employ his technique of Hamming Distance measurements for iris matching. The 256-byte mask contains the locations of bits to be omitted from the reading because of occlusions by eyelids, eyelashes and so on. The enrolment reading is strictly controlled by an operator; occlusions which occur in the enrolment mask are extremely likely to re-occur in subsequent readings. Therefore, bit positions included in the enrolment mask are automatically ignored in all further readings for this user, and we "reduce" the iris reading to its valid bits only. This reduced iris data acquired at enrolment becomes the k information bits of a codeword. Check digits are calculated in the usual manner, and the codeword is then hashed and stored in the user's file. Here also first the input image is segmented from the background which ensures the removal of noise with various filters of Prewitt, Laplace and canny operator A Gabor filter is used for this process and a suitable
Page 438
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 value of local variances is taken for carrying out the process of filtering. A Gabor filter takes care of both the frequency components as well as the spatial coordinates. The inputs required to create a Gabor mask are frequency, orientation angle and variances along x and y directions. Filtering is done for each block using the local orientation angle and frequency. Pre-processing of the image is completed by the steps as mentioned and the enhanced image is obtained. Normalization is done by same process which done in finger print process.
Fig.3 Original Fingerprint Image
2.2.2Post-Processing If we choose to work with a Reed-Solomon code over GF(28), this has a maximum length of 255 octets. The ground field of 28 lends a natural ease to dealing with octetbased calculations, an obvious advantage when working with bytes of data. A further advantage is that it is of characteristic two, which means that the operations of addition and subtraction are identical. This greatly simplifies arithmetic in the field, allowing us to make use of the XOR (exclusiveOR) function. Let the code length, n, be equal to 255, the maximum value it can attain in this code. The number of information symbols, k, cannot be greater than 200, as we know that this is the maximum number of valid bytes in a good Caucasian image. In addition, the check digits are effectively unhidden and thus could be accessed by an unscrupulous systems operator. As such, we must ensure that their number does not exceed n/2, as otherwise this would allow the codeword to be deciphered. Finally after finding the results of finger print and iris matching which will given to xor operation to prove that both finger print and iris will be come from same person in order to authenticate the user.
Fig.4 Normalized Image
Fig.5 Enhanced Image-Gabor Filtering 3. RESULTS AND DISCUSSIONS The original image of the finger print is given as input which was shown in Fig.3. Then the image was normalized using Gabor filtering and it was enhanced which was shown in fig.4,5.After enhancing the image, the false minutiae was removed and the given image is matched with stored image for matching which is shown in Fig.6,7. Fig.6 Minutiae obtained after deleting spurious minutiae at the borders
Velammal College of Engineering & Technology, Madurai
Page 439
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.7Final extracted Minutiae after applying the windowing technique The original image of the Iris is given as input which was shown in Fig.8. Then the image was converted into gray scale image which is normalized using Gabor filtering with various operators like laplacian , prewitt and it was enhanced which was shown in fig.9-13.After enhancing the image, the error detecting codes are combined with the binary image and that image is matched with stored image for matching..
Fig.10 Prewitt filter
Fig.11 Laplacian filter
Fig.12 Normalized image Fig.8 Original Image
Fig.12 Enhanced image using edge detection Fig.9Gray scale image
Velammal College of Engineering & Technology, Madurai
Page 440
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 4.CONCLUSION The main benefit of this method is its fast running speed. The method identifies the unrecoverable corrupted areas in the fingerprint and removes them from further processing. By treating the enrolment iris reading as a codeword, and the verification reading as a corrupted codeword, we have shown that it is possible to reliably match iris readings using ReedSolomon codes. Such an approach provides greater security than matching algorithms currently in use, as it only requires that a hashed version of the iris reading be stored, rather than an explicit template.
5.REFERENCES [1] D.H. Ballard, “Generalizing the Hough Transform to Detect Arbitrary Shapes,” Pattern Recognition, vol. 13, no. 3, pp. 111-112,1997 [2] A. Ross, J. Shah, and A. Jain, “Towards Reconstructing Fingerprints from Minutiae Points,” Proc. SPIE Conf. Biometric Technology for Human Identification II, pp. 68-80, 2005 [3] J. Daugman. “High confidence visual recognition of persons by a test of statistical independence” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.15,no.11, 1148-1161, Nov 1993. [4] J. Daugman. “How iris recognition works”, IEEE Transactions on Circuits and Systems for Video Technology, vol.14 no.1, 21-30,Jan 2004. [5] A. K. Jain, L. Hong, S. Pantanki and R. Bolle, “An Identity Authentication System Using Fingerprints”, Proc of the IEEE, vol.85, no.9,1365-1388, 1997
Velammal College of Engineering & Technology, Madurai
Page 441
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A New Frame Work For Analyzing Document Clustering Algorithms Mrs. J. Jayabharathy* and Dr. S. Kanmani** *Senior Lecturer, Department of Computer Science & Engineering **Professor & Head, Department of Information Technology Pondicherry Engineering College, Puducherry, India [email protected] [email protected]
ABSTRACT Nowadays all paper documents are in electronic format, because of quick access and smaller storage. So, it is a major issue to retrieve relevant documents from the larger database. Clustering documents to relevant groups is an active field of research finding various applications in the fields of text mining, topic tracking systems, intelligent web search engines and question answering systems. This paper proposes a frame work for comparing the existing document clustering algorithm. From the frame work, the details regarding the algorithms, their capabilities, evaluation metrics, data set and performance of the various methods are analyzed. From the analyses it is revealed that majority of the researches consider vector space model for document representation, F-measure, Isim & Esim , Precision and Recall are the frequently used metrics. Document clustering still has future scope in incremental document clustering, semi-supervised clustering, ontology based clustering, topic detection and document summarization. KEY WORDS— Text mining – Document clustering – unsupervised learning – Semi-supervised learning – ontology. INTRODUCTION Information extraction plays a vital role in today’s life. How efficiently and effectively the relevant documents are extracted from World Wide Web is a challenging issue[16]. As today’s search engine does just string matching, documents retrieved may not be so relevant according to user’s query. A good document clustering approach can assist computers in organizing the document corpus automatically into a meaningful cluster hierarchy for efficient browsing and navigation, which is very valuable for overcoming the deficiencies of traditional information
Velammal College of Engineering & Technology, Madurai
retrieval methods. If documents are well clustered, searching within the group with relevant documents improves efficiency and reduces the time for search. A good document clustering algorithm should have high intra-cluster similarity and less inter- cluster similarity. i.e the documents with the clusters should be more relevant compared to the documents of other clusters[12]. Document Clustering is an unsupervised approach, in which documents are automatically grouped into predefined number of groups based on their content. Document clustering methods are categorized as partitioning, hierarchical, densitybased, grid-based, model-based and etc. In this paper we mainly concentrate on partitional and hierarchical based clustering methods. This paper gives detailed view about the various document clustering algorithms with respect to the similarity measures, data set and performance metrics. The following section discusses about the procedure of general document clustering algorithm, section 3 gives the detailed survey of document clustering algorithms. Section 4 gives the overview of the proposed framework for comparing the existing document clustering algorithms. Section 5 concludes the paper and discusses about the challenges in document clustering. A GENERAL CLUSTERING
PROCEDURE
FOR
DOCUMENT
Document clustering problem could be expressed as follows: Let N be the total number of documents to be clustered. The set D={D1, D2, ….DN} be the set of N documents [25]. The documents are grouped into non-overlapping clusters C= { C1, C2, ….Ck} where k denotes the number of groups formed. C1U C2U….U Ck = D, Assume Ci ≠ Φ and Ci ∩ Cj = Φ where i ≠ j. Widely used document representation methods are VectorSpace Model (VSM), Multiword term, Character N-gram
Page 442
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
representation are some the existing representation models. VSM is identified as the most widely representation for documents. Using VSM, the document d is represented by the term-frequency (tf ) vector dtf = (tf1; tf2; : : : ; tfm), where tfi is the frequency of the ith term in d. Currently, the TF-IDF (Term Frequency times Inverse Document Frequency) model is the most popular model. In the TF-IDF model, a term ti is weighted by its frequency in the document tfi times the logarithm of its inverse document frequency, i.e., tfi * log2(N/n), where N is the number of documents in the collection and n is the number of documents where the term ti occurs at least once. Using the TF-IDF model, terms that appear too rarely or too frequently are ranked lower than the other terms. Widely e-documents suffer from high dimensionality; document representation [15] is itself a challenging issue. Usually document sources are of unstructured format, transforming these unstructured documents to structured format requires some preprocessing steps. Figure 1. Represents the sequence of steps involved in document clustering. Commonly used pre-processing steps are Tokenization: Tokenization is the process of mapping sentences from character strings into strings of words. For example, the sentence “This is a Survey Paper” could be tokenized into \This, \is, \a, \Survey, \Paper. Data cleaning: Non-textual information, such as HTML tags, punctuation marks and digits, are removed from the document. Stopwords removal: Stop words are typical frequently occurring words that have little or no discriminating power, such as \a", \about", \all", etc., or other domain-dependent words. Stopwords are often removed. Stemming: Removes the affixes in the words and produces the root word known as the stem [17]. Typically, the stemming process will be performed so that the words are transformed into their root form. For example connected, connecting and connection will be transformed into connect. Most widely used stemming algorithms are Porter [21], Paice stemmer [ 20], Lovins [19], S-removal[18]
Velammal College of Engineering & Technology, Madurai
Figure 1 General Procedure for Document Clustering Document Clustering methods [3] are categorized as partitioning, hierarchical, graph-based, model-based and etc. Partitional Clustering – It divides the data objects into nonoverlapping subsets (clusters) such that each data object is in exactly one subset. Hierarchical clustering is a set of nested clusters organized as a hierarchical tree. In this approach, a hierarchy is built bottom-up by iteratively computing the similarity between all pairs of clusters and then merging the most similar pair. Graph based model – This algorithm transforms the problem so that graph theory can be used [4]. In this approach, the objects (documents or words) to be clustered can be viewed as a set of vertices. Two vertices are connected with an undirected edge of positive weight based on certain measurement. In the clustering process, a set of edges, called edge separator, are removed so that the graph is partitioned into k pair-wise disjoint sub-graphs. Objective of the partitioning is to find such separator with a minimum sum of edge weights. Clustering algorithms are based on certain similarity measures [2]. There are many different ways to measure how similar are the two documents. Similarity measures are highly dependent on the choice of terms to represent text documents. Accurate clustering requires a precise definition of the closeness between a pair of objects, in terms of either the pair-wise similarity or distance. Cosine similarity and Jaccard correlation coefficient are widely used similarity or
Page 443
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
distance measures. Euclidean distance and entropy are being applied to calculate pair-wise distance between objects in clustering. CLUSTERING ALGORITHMS This section gives a detailed survey of recent work in document clustering. Benjamin C.M Fung et al [1] elaborated Frequent itemset based hierarchical clustering [22] technique for addressing the problems in hierarchical document clustering. This technique uses frequent item sets to construct clusters and to organize clusters into a topic hierarchy. They have experimentally evaluated FIHC and compared with Hierarchical Frequent Term based clustering (HFTC) proposed by Beil, Ester & Xu, [23] and Unweighted Pair Group Method with Arithmetic Mean (UPMA) proposed by Kaufman & Rousseuw, [24]. This paper concludes that FIHC outperforms well with respect to time, accuracy and scalability. A quantitative and qualitative method was introduced by Yi Peng et al [5] which extract phrases/noun from the documents for clustering. Two hypotheses are designed as follows 1. Instead of single word, this approach uses noun/phrases and words representation and used an optimization method from CLUTO tool kit to group the documents. 2. More meaningful subject headings of document collection could be obtained by merging the similar, smaller clusters instead of generating larger clusters directly. The data set considered for clustering is Data Mining and Knowledge Discovery Publications. This data set was clustered as 8-way and 100-way category using CLUTO tool kit and Intra-cluster similarity and Inter-Cluster similarity are considered as performance metrics. The author concludes that the clustering documents using phrase/word representation generates more meaningful appropriate clusters compared to single word representation and merging smaller and similar documents yields more relevant documents. A framework named SS-NMF (Semi Supervised Non-negative Factorization) which incorporates user defined constraints for document clustering was proposed by Yanhua Chen et al [6], users are able to provide supervision for clustering in terms of must-link and cannot link-link pairwise constraints. An efficient iterative algorithm had been considered for computation of symmetric tri-factorization of the document-document similarity matrix to infer the document clusters. MEDLINE Database, Reuters-21578 Text Collection and Foreign Broadcast Information Service Data are the data sets utilized. This SS-NMF frame work has been compared with supervised and semi-supervised
Velammal College of Engineering & Technology, Madurai
clustering algorithms k-means, kernel k-means, Spectral Normalized Cuts, Non-negative Factorization (NMF), Semisupervised kernel k-means (SS-KK), semi – supervised spectral clustering with normalized cuts (SS-SNC). Confusion Matrix and Accuracy as evaluation metrics and concludes that SS-NMF works better than other supervised and unsupervised algorithms. Jiangtao Qiu [7] et al proposed a Topic Oriented Semantic Annotation algorithm. The author considers semisupervised document clustering approach. User’s need should be represented as multi-attribute topic structure. Using the topic structure, the next step is to compute topic semantic annotation for each document and topic semantic similarity between documents. Ontology based Document topic semantic annotation (ODSA) algorithm was proposed for computing topic semantic annotation for each document. They have also proposed a dissimilarity function using which the dissimilarity matrix have to be constructed. The dissimilarity of two documents is their topic-semantic dissimilarity in topic oriented document clustering. Then the documents are clustered using proposed Optimizing Hierarchical clustering algorithm which is based on agglomerative hierarchical document clustering algorithm. They have also proposed a new cluster evaluation function DistanceSum. In this paper, a comparison between ODSA and Term Frequency and Inverse Document Frequency (TFIDF) representation have been experimented and proved that ODSA takes less time for constructing dissimilarity matrix when compared to TFIDF. They have also assessed the quality of the cluster using the F-measure as the performance metric, i.e. the clustering algorithm based on ODSA and TFIDF have been experimented and judged. The discovery of Topic by Clustering Keywords without any prior knowledge was proposed by Christian Wartena and Rogier Brussee [8]. Identification of most informative keyword is done by using natural probability distributions and the documents are clustered according to various similarity functions like Cosine similarity, JensenShannon divergence distance between document distribution and term distribution. A collection of 8 Wikipedia topics are considered as data set. Considering F-measure as the performance metrics, the experiments show that distribution of terms associated with keyword gives better results. An Active Learning Framework for Semisupervised document clustering with language modeling [9] proposed by Ruizhang Huang and Wai Lam incorporates user defined constraints for clustering. This paper mainly deals with semi-supervised document clustering. This framework concentrates on pair-wise constraints for user provided information i.e. user can provide hierarchical constraints for
Page 444
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
indicating the parent to child relationship between documents. Using these constraints the clustering processes learns actively. Also they have devised a gain function for selecting document pairs. Instead of document represented as bag of words they include term-term dependence relationship using term co-occurrences. Term- Term dependency is used for estimating the terms related to the same cluster. A collection of documents from 20-newsgroup corpus, Reuters RCV1 corpus TDT3 (Topic detection and Tracking) are taken as dataset for experiments. NMI (Normalized Mutual Information) and Pair-wise F-measure are considered as performance evaluation metrics and compares the proposed framework with Active semi-supervision for pair-wise constrained clustering algorithm and concludes that the proposed approach outperforms. Document Clustering Description Extraction and its Applications [10] is a problem of labeling the clustered documents. This paper mainly focuses on automatic labeling concept based on machine learning. Clustering problem transformed into classification for labeling the clusters. Considering two Benchmark models Baeline1 and Baseline1 and SVM (support vector machine), Multiple Linear Regression (MLR) and Logistical regression models (Logit) machine learning approaches are considered for labeling. More than 2,000 academic documents from Information centre for Social Sciences of RUC were taken as the dataset for experiments. Considering Precision, recall and F1 value as performance metrics, the author justifies that SVM outperforms better compared to the other models. Baseline1 model behaves worst compared to five models. Yanjun Li et al [13] propose two clustering algorithms, Clustering based on Frequency of Words (CFWS) and Clustering based on Frequency of Word Meaning Sequence (CFWMS) instead of considering bag of words representation. Word is the one which is simply represented in the document, whereas word meaning is the concept which expressed by synonyms of word forms. Word meaning sequence is the sequence of frequency of the word occurring in the documents. Authors have addressed many issues document representation, dimensionality reduction, labeling the clusters, overlapping of clusters. They preferred Generalized suffix tree for representing the frequency words in each documents. Dimensionality is reduced because of word meaning representation instead of words. CFWS is evaluated to be better algorithm compared to bi-secting kmeans and FIHC with respect of accuracy, overlapping
Velammal College of Engineering & Technology, Madurai
cluster quality and self labeling features. CFWMS achieves better accuracy when compared to CFWS, and modified Bisecting k-means using background knowledge (BBK)[11] Another significant topic in document clustering is document summarization, gives only the main points of the original document. This helps the user to grasp the contents of the entire documents in a moment. Ontology Enchanced clustering based summarization of medical documents[16] was proposed by Kokilavani and Balasubramanie. They have chosen MESH descriptor set for matching string and used kmeans algorithm for clustering. The summary generated was evaluated using Precision and Recall as metrics. Khaled B. Shaban [11] proposed a new frame work namely a Semantic Approach for Document clustering, which adopts meaning based representation of text, and their use in measuring similarities between documents using the knowledge representation structure. Author has used Semantic Graph Model (SGM) for document representation, which is similar to graph based representation. SGM is based on syntactic and symantic information extracted from the documents. For comparative analysis the author takes Vector Space Model and evaluated the performance with respect to F-measure, Entropy, overall similarity and concludes that keywords alone cannot give better cluster quality. He has also proposed a similarity measure precisely for semantic based mining. FRAME WORK FOR ANALYZING THE EXISTING DOCUMENT CLUSTERING TECHNIQUES This study mainly highlights the recent research work in the field of document clustering. This paper mainly focuses about the proposed frame work for comparing various document clustering techniques. The comparative study is based on the survey, which is made by analyzing the existing algorithms, considering the topics like Document clustering technique, Data set used for experiments, Performance metrics and summarization about the performance of the proposed technique. Column 1 in this frame work states the title of the survived papers. Frame work, algorithm and techniques which are discussed in the existing papers are stated in column 2. The 3rd column gives the details of the data set which is considered for conducting the experiments. Metrics considered by the authors for performance evaluation are given in column 4. The concise details about the performance of the algorithms are listed in column 5.
Page 445
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PROPOSED FRAME WORK PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
PERFORMANCE
Recent Trends in Data Mining (DM): Document clustering of DM Publications – 2005
K-means Implemented in CLUTO - Toolkit
1436 papers from different data mining publications
Human Inspection
Forms 47 clusters according to the category
Hierarchical Document Clustering - 2006
Frequency Itemsetbased Hierarchical clustering (FIHC)
100k documents
F-measure
Improved Accuracy
A Hybrid Strategy for Clustering Data Mining Documents- 2006
A qualitative method which extract noun/phrases instead of words from the documents.
Books, thesis & Publications of Data mining and knowledge discovery 1626 papers.
Isim – Intracluster s Similarity Esim – intercluster similarity
Clustered into 8 and 100 categories and concluded that 100way clustering gives quality clustering- Human inspection
Incorporating User Provided constraints into Document clustering's - 2007
SS-NMF : semi supervised nonnegative matrix factorization framework
MEDLINE Database, Reuters-21578 Text Collection and Foreign Broadcast Information Service Data
Confusion matrix Accuracy metric
Concludes that SSNMF outperforms well when compared with semi supervised and un supervised clustering methods
Topic Oriented Semi- Supervised Document Clustering - 200 7
Ontology based Document topic semantic annotation (HOWNET ontology knowledge base)
Chinese web pages about “Li Ming” – same name different personality football player who belongs different clubs
Time analysis for computing Dissimilarity matrix and Dimensionality analysis
Compares with TFIDF w.r.t time (construction of dissimilarity matrix ) And dimensionality analysis (less)
A document Clustering and Ranking system for Exploring MEDLINE
CLUTO tool kit
Society of surgical oncology Bibliography– 10 categories of
Document clustering is a part of Information retrieval in this
This framework retrieves reduced set of articles because of clustering, topic extraction and ranking methods
Bisecting K – means clustering algorithm
Velammal College of Engineering & Technology, Madurai
Scalability
and
Page 446
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
citations - 2007
Spectroscopy
cancer
paper
Text Document Clustering based on Frequent word meaning sequences
Clustering based on Frequent Word Sequences(CFWS) and Clustering based on Frequent Word Meaning Sequences (CFWMS) – CLUTO kit
9 categories from Reuters data set, Classic data set and 200 documents from search engine results.
F-measure Purity
CFWS & CFWMS – has better accuracy compared to Bisecting K-means and Frequent Itemset based Hierarchical clustering(FIHC)
Similarity Measures for Text Document Clustering
Similarity measure – Euclidian, Jaccard coefficient,Cosine simiarity, Pearson correlation coefficient, Average Kullback – Leibler Divergence
20 newsgroups, Classic – academic papers, wap – web pages, web knowldege base - webkb
Purity Entropy
Except Euclidian other S.M are quite similar. Jaccard and Pearson are slightly better with higher purity value
Research Field Discovery Based on Text Clustering
Newman Fast clustering algorithm for network model
Research proposals of National Natural Science and Foundation – 3423 relevant proposals
Modularity
Gives insight of the year wise research field .
Topic Detection by Clustering Keywords
Identifies most informative keyword by Probability , Clustering using Bisecting K-means algorithm
Wikipedia articles of various categories like architecture, popmusic etc
F-measure
Implemented using 3 similarity measures like Cosine, Jensen shanon divergence for term and document distribution. Jensen shanon for term gives better result
Velammal College of Engineering & Technology, Madurai
PERFORMANCE
Page 447
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
PERFORMANCE
An Active Learning Framework for Semi-supervised document Clustering with language modeling
DevisedGain Function for selecting documents pairs, Language Model, Term-Term dependence instead of bag of words Designed a Framework – Active Semi
20 newsgroups, Reuters – newswire stories And TDTTopic detection and detection project
F-measure
Proposed Active Semi gives better clustering performance
Document Clustering Description Extraction and its Applications
Uses Machine Learning approach like SVM, Multiple Linear Regression Model and Two Benchmarks – Baseline1 and 2 for Cluster Description
Documents from Information Center for Social Science of RUC – 2000 documents
Precision Recall
SVM performs better compared to other 4 Machine Learning methods
A Semantic Approach for Document Clustering
A Framework designed – Parses document syntactically, Semantically converts to Semantic Graph Model Algorithm – KNearest Neighbor, HAC, Single Pass Clustering
Reuters Set
F-measure Entropy
Quality of Clustering outperforms well compared to VSM
Ontology Enhanced Clustering Based Summarization of Medical Documents
K- means and MesH ontology descriptor for matching sentences for summarization
PubMed documents
Precision Recall
MesH descriptor based summarization gives better summarization
Velammal College of Engineering & Technology, Madurai
Data
NMI (Normalized mutual information)
Page 448
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
PERFORMANCE
Text Document Clustering Based on Neighbors
3 methods : 1. new method for selecting initial cluster centroid 2. Similarity – links and cosine 3. Heuristic function for spiting the cluster
Reuters and MEDLINE
F-measure Purity
Compared with K-Means, Bisecting K-Means and with proposed techniques. Gives better results.
A Comparison of Document Clustering Technique - 2000
Similarity Measure – Centroid, UPGMA and Inter cluster similarity
Reuters, web AceA Web Agent for Document Categorization and Exploration
F-measure
Compares Bi-secting kmeans is better than UPGMA and K-means.
From the study it is inferred that more than 70% of the document clustering algorithms are based on Partitional and hierarchical based algorithms. It is understood that k-means for document clustering dominates in late 1990’s. Later Bisecting k-means algorithm produces good cluster quality compared to k-means. Recent researches mainly concentrates on Ontology based document clustering
Velammal College of Engineering & Technology, Madurai
techniques. It is also inferred that supervised document clustering techniques drives more importance than normal clustering algorithms. Nowadays we have broad scope in supervised document clustering, topic detection and in ontology based document clustering. From the survey it is understood that most of the document clustering algorithms are implemented in CLUTO tool kit.
Page 449
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
P recision & R ecall
35
C onfusion matrix 30 25
Inter&Intra clus similarity
20
E ntrophy & Modularity NMI
15 Misclassification index
10
P urity 5 F ‐measure 0 ME T R IC S
Figure 2 Percentage of Metrics considered for documents clustering algorithms The graph in Figure 2 states the various metrics considered in document clustering in X-axis and percentage of references in Y-axis. More than 30% of the document clustering algorithms concentrates on Intra and Inter cluster similarity measure. 30% of the authors prefer Precision and Recall which gives the quality about the document relevancy. 20% of the document clustering algorithm considers Fmeasure and Purity as performance metrics. Around 10%, the qualities of some clustering techniques are evaluated by Human intervention. 5% to 10% researchers consider (Normalized mutual information) NMI, Modularity, Confusion matrices as performance metrics. Figure 3 shows about the choice of data sets for document clustering with the percentage of usage. From the frame work it is analyzed that majority of the researcher’s attention has been drawn by 20 newsgroups and MEDLINE data set for experiments. Data documents from Reuters which has categories like ORGs, PEOPLE, PLACES etc has received next importance. Wikipedia document sets receive next priority. PubMed and other data sets receive next
Velammal College of Engineering & Technology, Madurai
precedence. From the analyzes it is understood that recent researchers mainly concentrate on semi-supervised.
Page 450
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
35 30 25
R euters
20
20News group
15
Medline
10
W kipedia
5
C las s ic
0
P ubmed
D ata s ets for experiments Figure 3 Data sets for documents clustering algorithms
CONCLUSION In this paper a frame work to compare existing document clustering algorithms was proposed, which gives the summarization of recent research work on various algorithms in document clustering. This frame work precisely states the details about the algorithms, data sets, metrics and performance. Document clustering still has scope in various issues like incremental document clustering, semi-supervised clustering, topic detection and in document summarization.
REFERENCES 1.
2. 3. 4.
5.
Benjamin C. M. Fung, Ke Wang, Martin Ester, “Hierarchical Document Clustering”, The Encyclopedia Of Data Warehousing and Mining,2005. Anna Huang, “Similarity Measures for Text Document Clustering,”, NZCRSC, 2008. Jiawei Han and Micheline Kamber. “Data Mining: Concepts and Techniques., Mogan Kaufmann”, San Francisco, CA, USA, 2001. Eui-Hong Han, George Karypis, Vipin Kumar, and Bamshad Mobasher. “Clustering based on association rule hypergraphs”, In Research Issues on Data Mining and Knowledge Discovery, 1997. Yi Peng, Gang Kou, Yong Shi and Zhengxin Chen, “ A Hybrid Strategy for Clustering Data Mining
Velammal College of Engineering & Technology, Madurai
6.
7.
8.
9.
10.
11.
Documents”, Sixth IEEE International Conference on Data Mining – Workshop (ICDMW’06). 2006. Yanhua Chen, Manjeet Rege, Ming Dong, Jing Hua, “Incorporating User Provided constraints into Document clustering's”, - Proc. of IEEE International Conference on Data Mining, 2007. (Regular Paper). Jiangtao Qiu, Changjie Tang, “Topic Oriented SemiSupervised Document Clustering”, SIGMOD2007 Ph.D. Workshop on Innovative Database Research 2007(IDAR2007). Christian Wartena, Rogier Brussee, “Topic Detection by Clustering Keywords”, IEEE -19th International Conference and Expert System Application,2008 Ruizhang Huang, Wai Lam, “An Active Learning Framework for Semi-supervised document Clustering with language modeling”, Data & Knowledge Engineering,Elesiver 2008. Chenzhi Zhang, Huilin Wang, Yao Liu, Hongjiao Xu, “Document Clustering Description Extraction and its Applications”, Computer Processing of Oriental Languages. Language Technology for the Knowledge-based Economy, 22nd International Conference,ICCPOL 2009, Springer 2009. Khaled B. Shaban, “A Semantic Approach for Document Clustering”, Journal of Software 2009.
Page 451
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
12.
13.
14.
15. 16.
17. 18.
M. Steinbach, G. Karypis, and V. Kumar. “A comparison of document clustering Techniques”, In Proceedings of Workshop on Text Mining, 6th ACM SIGKDD International Conference on Data Mining (KDD’00), pages 109–110, August 20–23. Yanjun Li, Soon M. Chung, John D. Holt, “Text Document Clustering based on Frequent word meaning sequences”, Data Knowledge Engineering, Elesiver 2007. A.Hotho, S. Staab, and G. Stumme, “Ontology improves tect document clustering”, Proceedings of 3rd IEEE International Conference on Data Mining”,2003. Pp 551-554. Document Representation and Dimension Reduction for Text Clustering A. A. Kogilavani, Dr. P. Balasubramanie, “Ontology Enhanced Clustering Based Summarization of Medical Documents”, International Journal of Recent Trends in Engineering, 2009 WB Frakes, CJ Fox , “Strength and Similarity of Affix Removal Stemming Algorithms”, ACM SIGIR Forum, 2003 Harman, D. "How Effective is Suffixing." Journal of the American Society for Information Science 42 (1), 1991, 7-15.
Velammal College of Engineering & Technology, Madurai
19. 20. 21. 22 .
23.
24. 25.
Lovins, J. B. "Development of a Stemming Algorithm." Mechanical Translation and Computational Linguistics 11, 1968, 22-31. Paice, Chris D. "Another Stemmer." SIGIR Forum 24 (3), 1990, 56-61. Porter, M. F. "An Algorithm for Suffix Stripping."Program 14, 1980, 130-137. Fung B, Wnag K & Ester .M, “Hierarchical Document Clustering using Frequent itemsets”, SIAM International Conference on Data Mining, SDM ’03.2003. Pp 59-70. Beil F, Ester M & Xu. X , “ Frequent Term Based Test Clustering”, International Conference on Knowledge Discovery and Data Minig, KDD’02. 2002.pp 436-442. Kaufman, L & Rousseeuw P.J, “ Finding groups in data: An introduction to cluster analysis”, John Wiley & Sons, New York. 1990. Zhaogu Xuen, Li chen, Yanzhangg Dang, Jiang Yu, “ Research Field Discovery Based on text clustering”, IEEE 2008.
Page 452
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Optimization of Tool Wear in Shaping Process by Machine vision system Using Genetic Algorithm S.Palani¹ G.Senthilkumar² S.Saravanan³ J.Ragunesan³ 1. Professor, Mechanical department, Mount Zion Engineering College, Pudukkottai. [email protected]
2. Lecturer, Mechanical department, Velammal Engineering College, Madurai. [email protected]
3. Student, Mechanical department, Sudharsan Engineering College, Pudukkoattai. [email protected]
3. Student, Mechanical department, Sudharsan Engineering College, Pudukkoattai. [email protected]
Abstract Optimization has significant practical importance particularly for operating in machines. In order to increase the accuracy of finishing product the tool must be in good condition always as much as possible. The cutting tool must be in good condition when go for high precision of components i.e. the tolerance is very closer there the optimized cutting conditions were needed. To achieve good condition of tool, the machining parameters like speed, feed, depth of cut and average gray intensity level should be optimized. This paper aims to find out the safety cutting conditions for minimizing the tool wear by applying the optimized input parameters using genetic algorithm technique. Introduction The growing demand for product quality and economy necessity have forced incorporation of monitoring the process parameters in automated manufacturing systems. The greatest limitation of automation in the machining operation is tool wear. If the tool wear increases then the tool life will be minimum. So the tool wear has been optimized by selecting the optimal cutting parameters such as speed and depth of cut. The main machining parameters which are to be considered as variables of the optimization are speed and depth of cut, the required output is minimum tool wear. The optimization of tool wear is very important in modern machining processes. The cutting tool is the most critical part of the machining system, and recent advances in tool technology mean that many traditionally difficult materials
Velammal College of Engineering & Technology, Madurai
can now be more readily machined. Added to his, the desire for faster, more flexible manufacturing systems to improve productivity places even greater demands on cutting tools. Reliable wear data can be used not only for adaptive process control and optimization. Tool wear monitoring can also be used to determine the optimum conditions for a given machining process, such as finding the most economical cutting speed. The optimum set of these three parameters are determined for a particular job-tool combination of High Speed Steel-Mild Steel during which optimizes the tool wear. Many graphical and analytical techniques were initially used for optimization. Some of them are overcome by numerical optimization methods which may be either simplex or gradual based methods. These and other similar numerical optimization techniques have been found to give reasonably accurate results only for simply problems. But in the case of highly nonlinear optimization problems, the numbers of iterations are required to get optimum point, thus consuming significant computation time. So an alternative method for optimization is probabilistic methods. Genetic Algorithm is searches from a population and gives a set of reasonably good solution to the user. Computer vision technique Computer vision techniques have many particular advantages to offer when used for tool wear optimization. The measurement is non-contacting. And the magnification can also be obtained based on the power of the lens used. Hence clarity of the image obtained is more. Computer vision techniques were used to include a parameter affecting the
Page 453
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
tool wear i.e. the gray levels. With the given speed, feed, and depth of cut, shaping operation is performed on work piece, and the image of this tool is captured through the machine vision system. The image is imported to the computer and image processing techniques, the average gray level is found out. Genetic algorithm A genetic algorithm (GA) is a search technique used in computing to find exact or approximate solutions to optimization and search problems. Genetic algorithm is a global population search technique based on the operations of natural genetics and mimics the natural biological process. Genetic algorithm first encodes all variables into a finite bit binary string called as chromosomes. Chromosomes represent a possible solution of chromosomes are formed. Each chromosome is decoded and is evaluated according to the fitness function.
Fig: Structure of Genetic algorithm Input parameters
Table: Data’s for computing the regression equation
The following equation called as the regression equation and it is used as fitness function in Genetic programming. TW = - 0.779 - 0.000733 S + 12.5 F + 1.44 D + 0.00221 AG + 0.00131 sd - 20.8 fd
Velammal College of Engineering & Technology, Madurai
Page 454
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Training parameters
Comparison of measured tool wear with GA predicted tool wear
Experimental results
The result shows the average error of the prediction of tool wear in shaping using GA is 2.761 % i.e. the accuracy is 97.239% Optimized cutting conditions
Velammal College of Engineering & Technology, Madurai
Page 455
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Conclusion In this research an adaptive Genetic algorithm has been used to predict the optimum cutting conditions of a tool and tool wear in machining operation by employing the GA learning algorithm. The following conclusions can be drawn from this research. The reason for introducing the Genetic algorithm network technique in the present study is that the machining process is complex and uncertain in nature; the present machining theories are not adequate for the purpose of practical problem.
References 1. Alegre.E, Alaiz.R, Barreiro.J & Ruiz.J, Assessment and visualization of machine tool wear using computer vision.(2008) 2. C.Flix Prasad, S.Jayabal & U.Nataraajan.Optimization of tool wear in turning using genetic algorithm. (2007) 3. C.Bradly & Y.S.Wong. Surface texture indicators of tool wear-A machine vision approach.(2001) 4. Sukhomay pal, P.Stephan Heyns, Burkhard H. Freyer, Nico J.Theron & Surjya K.Pal. Tool wear monitoring and selection of optimum cutting conditions with progressive tool wear effect and input uncertainties.(2009) 5. H. H. Shahabi & M. M. Ratnam. Assessment of flank wear and nose radius wear from workpiece roughness profile in turning operation using machine vision. (2008) 6. Rick Riolo and Bill Worzel. Genetic Programming Theory& Practice.(2006) 7. Wen-Tung chein & Chung-shay Tsai. The investigation on the prediction of tool wear and determination of optimum cutting conditions in machining of stainless steel.( 2003)
8. D.Venkatesh, KKannan & R.Saravanan. A Genetic Algorithm based model for the optimization of machining process.(2009) 9. R.Saravanan, PAsokan & MSachidhanandham. Multi objective GA approach for optimization of tool wear.(2002)
Velammal College of Engineering & Technology, Madurai
Page 456
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Comparative Study Of Various Topologies And Its Performance Analysis Using Wdm Networks P. Poothathan1, S. Devipriya2 , S. John Ethilton3 P.Poothathan1 ,Department of Physics, Velammal College of Engg. Tech., Madurai-9, India. [email protected]. S.Devipriya2 , Department of Physics, Velammal College of Engg. Tech., Madurai-9, India. [email protected] S.John Ethilton3 ,Department of Physics, Velammal College of Engg. Tech., Madurai-9, India. [email protected]
ABSTRACT Optical wavelength division multiplexing (WDM) networking technology has been identified as a suitable candidate for future wide area network (WAN) environments, due to its potential ability to meet rising demands of high bandwidth and low latency communication. Networking protocols and algorithms are being developed to meet the changing operational requirements in future optical WANs. Simulation is used in the study and evaluation of such new protocols, and is considered a critical component of protocol design. In this paper, we construct an optical WDM network simulation tool, which facilitates the study of switching and routing schemes in WDM networks. This tool is designed as an extension to the network simulator ns2. In this work, the effectiveness of the proposed approach has been verified through numerical example and simulated results for various network scenarios, such as ring, mesh and interconnection topologies and we analyse the blocking probability of distributed light path establishment in wavelength-routed WDM networks with no wavelength conversion. We discuss the basic types of connection blocking: i) due to the dimension of the network. ii) due to offered network load. iii) due to network load in Erlang and Packet delay due to load in Erlang. Key words: Optical Networking, WDM network, topology, simulator, blocking probability, delay.
I. INTRODUCTION Optical fiber communications was mainly confined to transmitting a single optical channel until the late 1980s. Because the fibre attenuation was involved, this channel required periodic regeneration, which included detection, electronic processing, and optical retransmission. Such regeneration causes a high-speed optoelectronic bottleneck and can handle only a single wavelength. After the new generation
Velammal College of Engineering and Technology, Madurai
amplifiers were developed, it enabled us to accomplish highspeed repeaterless single channel transmission. A WDM system enables the fiber to carry more throughputs. By using wavelength-selective devices, independent signal routing also can be accomplished. Two common network topologies can use WDM, namely, the mesh and the ring networks. Each node in the mesh has a transmitter and a receiver, with the transmitter connected to one of the central passive mesh’s inputs and the receiver connected to one of the mesh’s outputs. WDM networks can also be of the ring variety. Rings are popular because so many electrical networks use this topology and because rings are easy to implement for any network geographical configuration. In an ideal WDM network, each user would have its own unique signature wavelength. Routing in such a network would be straightforward. This situation may be possible in a small network, but it is unlikely in a large network whose number of users is larger than the number of provided wavelengths. In fact, technologies that can provide and cope with 20 distinct wavelengths are the state of the art. There are some technological limitations in providing a large number of wavelengths, for instance: due to channel broadening effects and non-ideal optical filtering, channels must have minimum wavelength spacing. Wavelength range, accuracy, and stability are extremely difficult to control. Therefore, it is quite possible that a given network may have more users than available wavelengths, which will necessitate the reuse of a given set of wavelengths at different points in the network. The same wavelength could be reused by any of the input ports to access a completely different output port and establish an additional connection. This technique increases the capacity of a WDM network. The recent emerge of high-bit rate IP networking application is creating the need for on demand provisioning of wavelength routing with service - differentiated offering within transport layer. To fulfill these requirements, different WDM
Page 457
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 optical network architecture has been proposed. For servicedifferentiated offering network, an accurate engineering of WDM span design is needed. So the factors such as additive nature of signal degradation, limited cascade ability of optical components and traffic dependent signal quality should be taken into account for accurate WDM span design.
probability with network dimension, network load and delay etc. The following topologies have been studied in our project:
In order to demonstrate that our approach performs better than that reported in the literature and to investigate the performance of algorithms, we must resort to simulation studies based upon the ns-2 Network Simulator. Not able to find a suitable simulator that could support our proposed DWP algorithm, we designed and developed a simulator to implement routing and wavelength assignment in all-optical networks for various topologies such as ring , mesh and interconnected rings.
1.RING TOPOLOGY
• Ring • Mesh and • Interconnected rings.
Ring networks are less connected and consequently need less messaging, which in turn generally assures faster convergence towards destination, and hence decreases computation time. We considered rings for topologies where varying network size would not significantly vary the performance (i.e., for bidirectional rings only two paths exist independently of the ring size).
WDM optical networks have gained prime importance due to the rapid growth of internet and the ever increasing demand for voice and video transmission. By allowing several channels to be routed on the same fiber on different wavelengths, the capacity of each link is increased tremendously. However this also calls for more efficient planning before provisioning lightpaths. The recent advent of high bit rate IP network applications is creating the need for on demand provisioning of wavelength routed channels with service differentiated offerings within the transport layer. To fulfill these requirements different optical transport network architectures have been proposed driven by fundamental advances in WDM technologies. The availability of ultra longreach transport and all optical switching has enabled the deployment of all optical networks. While being attractive for their transparent and cost effective operation all optical networks require accurate engineering of WDM spans to meet the requirements of dynamic wavelength routing. The additive nature of signal degradation, limited cascade ability of optical components and traffic dependent signal quality are some of the reasons that make the provisioning of on demand wavelength channels a challenging task.
1.1 Nodes Ring Topology Here, we study a 9-node bi-directional ring topology as shown in Figure1.1, of which the properties per wavelength and network element are shown. As it can be seen from the results, the method DWP, being a multipath strategy, yields superior results for the given service. This is because DWP method yields several feasible paths, out of which the best can be selected, in contrast to the non-DWP is checked on remaining requirements (transmission degradation, reliability).
To overcome the problems of analog WDM design, electronic regeneration is deployed at optical switching nodes in the so called opaque optical networks. However electronic regeneration can also impose limitations on the wavelength routing, such as delay accumulation, reliability reduction and increase in the operational cost. These issues become particularly critical if service requirements force multidimensional optimization such as maximum reliability and minimum transmission degradation. We analyze the blocking
Velammal College of Engineering and Technology, Madurai
Page 458
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 1.2 Snapshot of the simulation for 9 node ring topology
Fig. 1.2 shows the 9 node WDM network ring topology generated by the topology generator along with a snapshot of the simulation run. The bandwidth and the number of wavelengths on each multi-wavelength link are configured by the specified values. The propagation delay of each link is generated at random by the topology generator. The exponential session traffic pairs are randomly distributed amongst all nodes, which are displayed by the packet flows in nam. In the snapshot the update events of the virtual topology are captured and displayed in chronological order in the annotation panel of the main window by the event monitor. For instance, at 7.009009s, a light path is created for the traffic session 0, from node 0 to node 5 and the lightpath is established on the shortest path (path1) between source and destination without wavelength conversion.
2. MESH TOPOLOGY
the shortest path according to a certain constraint is two hops, we include only paths up to three hops (“plus one hop”). The path length limitation as used here reduces the algorithm convergence time, by avoiding messages handling information about paths with hop lengths larger than the limit. Note that the number of hops is an intrinsic variable of the adapted Bellman-Ford algorithm and is therefore advantageous to use, but additionally also a limitation on the used constraints, possibly combined by some cost-function, might be used to further reduce the number of candidate paths. In this example, we assume only nodes and links as network elements. All nodes have the same properties per wavelength along the network. Transmitters and receivers are modeled as ideal. Due to the full connectivity, compared to the 9-node ring network, this network is analyzed for higher loads. Also we assume that reliability to be constant over time. Here, two versions of DWP are used, one which uses a reduced set of feasible paths, where DWP is used for wavelength allocation only, and whiles the other considers all feasible paths, i.e. where DWP is used for routing and wavelength allocation.
We next consider the example of a nine-node mesh network, where a reduced set of pre-routed paths is introduced. To limit the number of candidate feasible paths we define a hop limit that automatically adapts to different connection requests and network loads.
2.2 Snapshot of the simulation for the 9 node mesh topology
3. INTERCONNECTED RING TOPOLOGY
2.1.Mesh Topology
Therefore we use a limit relative to the number of hops of the shortest path found, according to a single constraint. For example, if a limit is defined in terms of number of hops and if
Velammal College of Engineering and Technology, Madurai
As an example of strong practical importance, we study interconnection of networks, where, as specifically shown in Figure 3.1, six ring networks are interconnected (denoted as “6 x 5”). We believe that interconnected networks will be increasingly important to study, since they will have to deal with a variety of network technologies and architectures, originating from different vendors and service providers. To analyze interconnected networks with differentiated
Page 459
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 transmission impairments and service requirements, we consider an example where networks are designed with four different link types and the three link Parameters that are considered for routing: transmission degradation, reliability and delay.
4. PERFORMANCE ANALYSIS In order to demonstrate that our approach performs better than that reported in the literature and to investigate the performance of algorithms, we must resort to simulation studies based upon the ns-2.0 Network Simulator. Not able to find a suitable simulator that could support our proposed DWP algorithm, we designed and developed a simulator to implement routing and wavelength assignment in all-optical networks for various topologies such as ring , mesh and interconnected rings. The simulator accepts input parameters such as the number of nodes in the network, link information with weight, number of wavelengths per fiber, connection requests.
3.1 6x5-nodes network of interconnected rings Figure 3.1 shows the distribution of different link types (Type A, B, C, and D) along the example networks. . For the 6x5 network, we assumed five-node rings using links of Type A only. They are interconnected by two node disjoint links of Type B. Two link types are used for their long-haul interconnection: Type C for connections to the next closest neighbors, and Type D for connections diagonally
Some of the calls may be blocked because of the unavailability of free wavelength on links along the route from the source to the destination. The ratio of the total number rejected requests (blocked calls) to the total number of light path requests in the network is defined as the blocking probability. The output of the simulator is the blocking probability for the specified parameters along with the detailed information of connections. All these parameters can be initialized before running the simulations to obtain results for a given selection of parameters. Extensive simulations are then carried out for every combination of parameters of interest and the results are obtained .
4.1 Blocking probability Vs Number of nodes 4.1 Blocking probability Vs Number of nodes 3.2 Snapshot of the simulation for the 6x5 nodes interconnected rings topology
Velammal College of Engineering and Technology, Madurai
When we analysis the call blocking probability with size of the network (Number of nodes), the blocking probability becomes almost constant with number of nodes. But the graph
Page 460
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 shows that DWP methods are useful for mesh networks, while they bring less benefit for rings. With increasing number of nodes, the blocking probability decreases for mesh, and increases for ring networks. This is due to that DWP can use one alternative path only in bi-directional ring structures, which is N-n hops long (being number of nodes, being length of the shortest path), with which, if selected, even more network resources get allocated.
The call blocking probabilities are obtained as a function of the traffic load arriving at the network for all topologies. Here, the traffic arrival rate, traffic holding time and other conflicts are kept constant. Then the graph is plotted between the load and the blocking probability. There is no blocking probability for low load arriving at the network for all network topologies. When the arrival load to the system will be increased, leading to a remarkable increase in the blocking occur in the Performance Analysis of Constraint Based WDM Networks using DWP Algorithm network due to the offered load.
4.2 Delay due to Network size The time delay is obtained as a function of the network size for various network topologies such as ring and mesh. Here, the traffic load and other conflicts such as traffic arrival rate and traffic holding time are kept as constant. Then graph is plotted between the number of load and delay time. In the simulation, we observed that, at low network size, the delay time I s low for ring network and higher for mesh network. As the network size increases the delay time increases for ring network than the mesh network.
4.3 Blocking probability Vs Load The mesh network shows lower blocking probability for higher loads compared with other network topologies. 4.4 Blocking due to conflicts (Traffic arrival rate & Traffic holding time)
Figure 4.2 Time delay Vs Number of nodes
This is due to the fact that with larger ring networks, larger amounts of network resources are allocated and cause delay of future calls whenever a service is accommodated and also DWP can use one alternative path only in bi-directional ring structures while it can use more candidate paths in mesh structure. 4.3 Blocking due to Load
Velammal College of Engineering and Technology, Madurai
4.4 Blocking probability Vs network load in Erlang
Page 461
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 The call blocking probabilities are obtained as a function of load in Erlang for the network for all topologies. Here, the traffic load and other conflicts are kept constant. Then the graph is plotted between the load in Erlang traffic and the blocking probability. At low arrival rates, primarily the load on the network causes blocking; however, as the load in Erlang increases by increasing the traffic arrival rate while the offered load remains constant, the blocking due to load doesn’t increase by much. However, the blocking due to conflicting connection requests increases as expected and becomes the dominant source of blocking for networks with low loads. We also observe that, in the simulation, the blocking due to load increases slightly as the arrival rate increases. Although the offered load remains constant, the actual network utilization is increasing, since connections, which are blocked still reserve network resources for a short period of time, leading to a slight increase in blocking probability. Furthermore, during the connection setup process when resources are being reserved, the reserved resources will go unused for a short period of time before the connection can begin transmitting data. This resource reservation overhead will be higher when the connections are established for shorter time duration, and the number of connections being established is higher. Thus, as the arrival rate increases, the overall load in the network will tend to increase.
4.5 Delay due to Conflicts Load in Erlang
Then graph is plotted between the delay time and the load in Erlang. The load in Erlang is increased by increasing either traffic arrival rate or traffic holding time. In the simulation, we observed that, the delay time increases with load in Erlang. This is due to the fact that, larger amounts of network resources are allocated and cause delay of future calls whenever a service is accommodated.
5. CONCLUSION In this paper, we proposed a new approach to constraintbased path selection for dynamic routing and wavelength allocation in optical networks based on WDM. Our approach considered service-specific path quality attributes, such as physical layer impairments, reliability, policy and traffic conditions. Here we considered dynamic routing method. Although this method requires a long set of time, it is more efficient than other methods term of blocking probability. To validate the network modeling, we presented the detail of the revised OWns architecture, which is designed to fulfill the key characteristic of WDM network and we implemented the behavior of DWP algorithm for dynamic routing and wave length assignment problem for various networks such as ring, mesh and interconnected topologies with no conversion and we compared the behavior of blocking probability and packet delay for various network topologies. The study of blocking and delay is very useful to know about the suitable topology for various network sizes, traffic offered load and load due to conflicting factors of a network.
The time delay is obtained as a function of the load in Erlang for various network topologies such as ring and mesh. Here, the traffic load and other conflicts are kept as constant.
4.5 Blocking probability Vs Packet delay
Velammal College of Engineering and Technology, Madurai
BIBLIOGRAPHY 1) A. Jukan, and Gerald Franzl, “Path Selection Methods With Multiple Constraints in Service-Guaranteed WDM Networks” IEEE/ACM TRANSACTIONS ON NETWORKING, vol 12,pp.59-71(2004). 2) A. Jukan and H. R. van As, “Service-specific resource allocation in WDM networks with quality constraints,” IEEE J. Select/AREAS COMMUNICATION, vol 18, pp. 2051– 2061(2000). 3) J. P .Jue and Gaoxi Xiao, “Analysis of Blocking Probability for Connection Management Schemes in Optical Networks”, IEEE/ ACM TRANSACTIONS ON NETWORKING,vol 17, pp.1546-1550(2005). 4) Ashwin Sridharan and Kumar N. Sivarajan, “Blocking in All-Optical Networks”,IEEE/ACM TRANSACTIONS ON NETWORKING, vol. 12,pp.384- 396 (2004).
Page 462
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 5) Paramjeet Singh,Ajay K. Sharma, Shaveta Rani, “Routing and wavelength assignment strategies in optical networks”www.sciencedirect.com/ OPTICAL FIBER TECHNOLOGY,vol.13, pp. 191–197(2007). 6) Vinh Trong Le, Xiaohong Jiang, Yasushi Inoguchi, Son Hong Ngoa, Susumu Horiguchib, “A novel dynamic survivable routing in WDM optical networks with/without sparse wavelength conversion” www.sciencedirect.com/ OPTICAL SWITCHING AND NETWORKING, vol.3,pp. 173–190(2006). 7) C. Siva Rain Murthy and Mohan Guruswamy Chapter 4 Wavelength Rerouting Algolithms," WDM OPTICAL NETWORKS Concepts De-sign and Algorithms", PHI 2002. 8) R. Ramaswami, et al., Optical Networks: A Practical Perspective, Morgan Kaufmann Publishers Inc., 2002. Mr. P. Poothathan, Senior Lecturer, Department of Physics, Velammal College of Engg. & Tech., Madurai, completed his M.Sc. Physics from Madurai Kamaraj University. He received his M.Tech. degree in Optoelectronics & Optical communication from Rajiv Gandhi Techincal Univerisity, Bhopal. He is having a rich 14 years of teaching experience. He published many papers in various national and international conferences. His areas of interest are optical communication and Nano-computing. Ms. S. Devipriya, she received her M.Sc., and M.Phil degrees in Physics from the Pondicherry University, India, She is currently working as a Lecturer in the Department of Physics at Velammal College of Engg. & Tech., Madurai. She is having a teaching experience of 5 years and her research areas of interest are nonlinear Physics and Nanotechnology. She is an University rank holder in II position in her M.Sc. Physics. Dr. S. John Ethilton, Assistant Professor & Head, Department of Physics, Velammal College of Engg. & Tech., Madurai. He completed his M.Sc. Physics from Madurai Kamaraj University. He obtained his doctoral degree from Manonmaniam Sundaranar University, Tirunelveli. He is having totally 10 years of teaching experiences. He is having 3 international publications and 1 national publication. His research interests are Nanotechnology and fuel cells.
Velammal College of Engineering and Technology, Madurai
Page 463
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Modeling of Cutting Parameters for Surface Roughness in Machining M. Aruna#1, P. Ramesh Kumar#2 #
Department of Mechanical Engineering, Velammal College of Engineering and Technology, Madurai, India. 1
2
[email protected] [email protected]
Abstract— In the present-day research, Response Surface Methodology is used to investigate the relationships and parametric interactions among the three controllable variables, namely cutting speed, feed rate and depth of cut, on the Surface roughness. Experiments are conducted on EN 24 steel with Ceramic inserts. To study the proposed second-order polynomial model for Surface roughness, the central composite experimental design is used to estimate the coefficients of the three factors, which are believed to influence the Surface roughness in Machining process. The response was modeled using a Response Surface Model based on experimental results. The significant coefficients are obtained by performing Analysis of Variance (ANOVA) at 5% level of significance. It was found that Cutting speed and feed rate have significant effect on the Surface roughness. This methodology is very effectual, needs only 20 experiments to levy the conditions, and model sufficiency was very satisfactory. Keywords— Response Surface Methodology (RSM), ANOVA, Central Composite Design (CCD)
I. INTRODUCTION There is a heavy demand for the advanced materials with high strength, high hardness, high temperature resistance and high strength to weight ratio, in the present-day’s technologically advanced industries like, Automobile, Aeronautics, Nuclear, Gas Turbine Industries etc. This necessity leads to evolution of advance materials like High strength alloys, Ceramics, Fibre-reinforced composites etc. In machining of these materials, conventional manufacturing processes are increasingly being replaced by more advanced techniques, and it is difficult to attain good surface finish and close tolerance. The appropriate range of feeds and cutting speeds, which provide a satisfactory tool life, is very limited [1,2]. Machinability of material provides an indication of its adaptability to be manufactured by a machining process. In general, machinability can be defined as an optimal combination of factors such as low cutting force, high material removal rate, good surface integrity, accurate and consistent workpiece geometrical characteristics; low tool feed rate and good curling of chip and breakdown of chips [3].
Velammal College of Engineering & Technology, Madurai
Therefore, it is troublesome to establish a model that can accurately predict the performance by correlating the process parameters. The optimum processing parameters are very much essential to boost up the production rate by a large extent and to shrink the machining time, since the advance materials are costlier, when compared to ordinary materials. Improving the surface quality in the advance materials is still a challenging problem. Little research has been reported about machining on EN 24 steel using Surface Response Methodology. In this work, surface response approach is used for development of a model and analysis of surface roughness, with cutting speed, feed rate and depth of cut as input parameters. A Central Composite Design (CCD) for combination of variables and response surface method (RSM) has been used to analyze the effect of the three parameters [4,5]. The ceramic cutting tools can be operated at higher cutting speeds than carbides and cermets for increased productivity and material removal rates. Experiments are conducted on hardened EN 24 steel with the hardness 40 HRC using toughened alumina cutting tool and Ti[C,N] mixed alumina ceramic tool. Flank wear models are developed and tool life analysis is done using the models [6]. Also, Experiments are conducted on S.G.Iron workpiece in a high speed lathe. The machinability of the S.G.Iron with ZTA cutting tool was evaluated in terms of the surface finish of the turned workpiece, flank wear of the cutting tool insert and cutting force during machining. The surface finish increases with cutting speed and the flank wear increases with cutting time and cutting speed [7]. The Taguchi and Shainin experimental design processes are also available for obtaining optimum cutting parameters. But still, it seems that the RSM approach would yield much better results. [8].
II. EXPERIMENTATION A number of experiments were conducted to study the effects of various machining parameters on Machining process. These studies have been undertaken to investigate the effects of cutting speed, feed rate and depth of cut on
Page 464
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 surface roughness. In this work, EN 24 steel is considered as work piece material and its composition is given inTable1. TABLE. 1 COMPOSITION OF EN 24 STEEL
C
Si
Mn
Cr
Mo
Ni
0.35-0.45 0.01-0.35 0.45-0.7 0.9-1.4 0.2-0.35 1.25-1.75
Designation: 40 Ni 2 Cr 1 Mo 28
In these methods, there is a possibility that the experiments will stop with fairly few runs and decide that the prediction model is satisfactory. Experiments have been carried out on the machining set up and the data were collected with respect to the influence of the predominant process parameters on surface roughness. 20 experiments are conducted and the value of surface roughness with design matrix is tabulated in the Table.3. TABLE. 3 PLANNING MATRIX OF THE EXPERIMENTS WITH THE OPTIMAL SURFACE ROUGHNESS
III. RESPONSE SURFACE METHODOLOGY Response surface methodology (RSM) is a collection of mathematical and statistical techniques that are useful for modelling and analysis of problems in which output or response influenced by several variables and the goal is to find the correlation between the response and the variables. It can be used for optimising the response [9,10,11]. It is an empirical modelization technique devoted to the evaluation of relations existing between a group of controlled experimental factors and the observed results of one or more selected criteria. A prior knowledge of the studied process is thus necessary to achieve a realistic model. The first step of RSM is to define the limits of the experimental domain to be explored. These limits are made as wide as possible to obtain a clear response from the model. The cutting speed, feed rate and depth of cut are the machining variables, selected in this investigation. The different levels retained for this study are depicted in Table 2. In the next step, the planning to accomplish the experiments by means of Response Surface Methodology (RSM) using a Central Composite Design (CCD) with three variables, eight cube points, four central points, six axial points and two centre points in axial, in total 20 runs are carried out. Total number of experiments conducted with the combination of machining parameters is presented in Table 2. The levels of significant are depicted in this Table. The Central Composite Design is used, since it gives a comparatively accurate prediction of all response variables related to quantities measured during experimentation. CCD offers the advantage that certain level adjustments are allowed and can be used in two-step chronological response surface methods [12].
Run
A
B
C
Ra (µm)
1
1.00000
-1.00000
1.00000
0.678
2
-1.00000
1.00000
-1.00000
0.589
3
0.00000
0.00000
0.00000
1.366
4
1.68179
0.00000
0.00000
0.510
5
0.00000
0.00000
0.00000
0.866
6
0.00000
0.00000
0.00000
1.051
7
1.00000
-1.00000
-1.00000
0.483
8
0.00000
0.00000
-1.68179
0.716
9
1.00000
1.00000
-1.00000
0.983
10
0.00000
0.00000
0.00000
0.516
11
-1.68179
0.00000
0.00000
0.561
12
0.00000
0.00000
0.00000
0.983
13
-1.00000
-1.00000
-1.00000
0.489
14
0.00000
0.00000
1.68179
0.733
15
0.00000
1.68179
0.00000
0.689
16
0.00000
-1.68179
0.00000
0.438
17
-1.00000
1.00000
1.00000
0.923
TABLE. 2 DIFFERENT VARIABLES USED IN THE EXPERIMENTATION AND THEIR LEVELS
18
1.00000
1.00000
1.00000
0.899
19
0.00000
0.00000
0.00000
0.466
20
-1.00000
-1.00000
1.00000
0.456
Variable
Cutting Speed (m/min) Feed rate (mm/rev) Depth of cut (mm)
Coding
A B C
Level 1 2 3 (-1) (0) (1) 80 110 140 0.04 0.06 0.08 0.4 0.45 0.5
Velammal College of Engineering & Technology, Madurai
The mathematical model is then developed that illustrate the relationship between the process variable and response. The behaviour of the system is explained by the following empirical second-order polynomial model.
Page 465
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2
ŷ = b x +b x + b x + b x + b x + b x +b x +b x .… 2
0 0
1 1
2 2
3 3
12 12
13 13
23 23
11 11
(1)
Analysis of variance (ANOVA) for the adequacy of the model is then performed in the subsequent step.
TABLE. 4 ANOVA TABLE FOR SURFACE ROUGHNESS, ESTIMATED REGRESSION
Source Regression
DF
Seq SS
Adj SS
Adj MS
F
P
9
0.49487
0.494868
0.054985
0.76
0.652
Linear
3
0.24668
0.246681
0.082227
1.14
0.378
Square
3
0.23974
0.239742
0.079914
1.11
0.390
Interaction
3
0.00844
0.008445
0.002815
0.04
0.989
Residual Error
10
0.71925
0.719246
0.071925
Lack-of-Fit
5
0.13928
0.139282
0.027856
0.24
0.928
Pure Error
5
0.57996
0.579963
0.115993
Total
19
1.21411
F value indicates that the variation of the process parameters makes a big change on the surface roughness and p denotes its TABLE. 5 COEFFICIENTS FOR SURFACE ROUGHNESS
Term
Coef
SE Coef
T
P
Constant
0.87157
0.10938
7.968
0.000
A
0.03663
0.07257
0.505
0.625
B
0.12522
0.07257
1.726
0.115
C
0.03226
0.07257
0.445
0.666
A*A
- 0.09969
0.07065
-1.411
0.189
B*B
-0.08979
0.07065
-1.271
0.233
C*C
-0.03286
0.07065
0.465
0.652
A*B
0.01925
0.09482
0.203
0.843
A*C
-0.02375
0.09482
-0.250
0.807
B*C
0.01100
0.09482
0.116
0.910
R-Sq = 40.76% R-Sq(adj) = 0.00%
The ANOVA Table for the curtailed quadratic model for surface roughness is shown in Table 4. The F ratio is calculated for 95% level of confidence. The value which are less than 0.05 are considered significant and the values greater than 0.05 are not significant. Table 4 indicates that the model is adequate since P values of lack-of-fit are not significant and F-statistics is 0.24. This implies that the model could fit and that it is adequate. From Table 5, the reduced model results indicate that the model is significant (R2 and adjusted R2 are 40.76% and 0.0%, respectively). The R2 value indicates that the machining parameters explain 40.76% of variance in surface roughness. This value indicates that the presented model fits the data well. The p-value shows that the model, linear terms and squared terms are non significant at α-level of 0.005 influence on surface roughness. The results predicted by regression model are compared with experimental measurements. Larger
Velammal College of Engineering & Technology, Madurai
percent contribution on surface roughness. Also, owing to the p-value of interaction is 0.928(>0.05), one can easily deduce that the interactions of distinct design variables are not significant. In other words, the most dominant design variables have minimum interaction with others in the current text. The final response equation for surface roughness is given as follows. Surface roughness=0.720+0.0366A+0.125B+0.0323
(C2)
The final model tested for variance analysis (F-test) indicates that the adequacy of the test is established. For analysing the data, the checking of goodness of fit of the model is very much required. The model adequacy checking includes the test for significance of the regression model, test for significance on model coefficients, and test for lack of fit. For this purpose, analysis of variance (ANOVA) is performed. The fit summary recommended that the quadratic model is statistically significant for analysis of surface roughness. This implies that the proposed model is adequate to illustrate the pattern of surface roughness. IV. RESULT AND DISCUSSION The effect of the machining parameters (cutting speed, feed rate and depth of cut) on the response variable surface roughness has been evaluated by conducting experiments. The results are put into the Minitab software for further analysis. Figure.1 shows the estimated response surface for surface roughness in relation to the process parameters of cutting speed and feed rate. It can be seen from the figure, the surface roughness tends to increase, significantly with increase in feed rate and maximum surface roughness is obtained at medium speeds.
Page 466
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 Also in Fig. 1, in the 3D graphics, it can be seen that when the cutting speed is low and the feed rate is high the surface roughness is in its maximums. The 3D surface graphs for the surface roughness which have curvilinear profile in accordance to the quadratic model are obtained and the adequacy of the model is thus verified.
Surface Plot of SURFACE ROUGHNESS vs B, A Hold Values C 0
In the Fig. 2, between the cutting speed and feed rate, with their elliptical shape, there is a strong, positive, second degree relationship. V. CONCLUSIONS The present study develops surface roughness models for three different parameters namely cutting speed, feed rate and depth of cut for machining process of EN 24 steel using response surface method. The second-order response models have been validated with analysis of variance. It is found that all the three machining parameters and some of their interactions have significant effect on surface roughness considered in the present study. With the model equations obtained, a designer can subsequently select the best combination of design variables for achieving optimum surface roughness. This eventually will reduce the machining time and save the cutting tools. References
0.75
[1]
FA CE ROUGHNESS 0.50 0.25
1 0
0.00 -2
-1
-1
0 A
[2]
B
-2
1
[3]
[4] Fig. 1 Surface plot of surface roughness vs B, A
Fig. 2 Example of of anSURFACE unacceptable low-resolution Contour Plot ROUGHNESS vs B, A image
1.5 Fig. 3 Example of an image with
1.0
B
0.5
SURFACE ROUGHNESS < 0.2 0.2 – 0.4 acceptable resolution 0.4 – 0.6 0.6 – 0.8 > 0.8
[5]
[6] [7]
[8]
Hold Values C 0
0.0
[9]
-0.5 [10]
-1.0 [11]
-1.5 -1.5 -1.0 -0.5 0.0 A
0.5
1.0
1.5
Fig. 2 Contour plot of surface roughness vs B, A
Velammal College of Engineering & Technology, Madurai
[12] [13]
L.N. Lopez del acalle, J. Perez, J.I. Llorente and J.A. Sanchez, “Advanced cutting conditions for the milling of aeronautical alloys”, Journal of Materials Processing Technology, vol 100, No. 1-3, pp. 111, 2000. E. Brinksmeier, U. Berger and R. Jannsen, “High speed milling of Ti6Al-4V for aircraft application”, First French and German Conference on High Speed Machining, Conf. Proceeding, Metz, pp. 295-306, 1997. M.Y. Noordin, V.C. Venkatesh, S. Sharif, S. Elting and A. Abdullah, “Application of response surface methodology in describing the performance of coated carbide tools when turning AISI 1045 steel”, “Journal of Materials Processing Technology”, vol. 145, No. 1, pp. 4658, 2004. R. Snoyes, F. Van Dijck, “Investigations of EDM operations by means of thermo mathematical models”, Annals of CIRP 20 (1), pp. 35, 1971. A. Senthil Kumar, A. Raja Durai and T. Sornakumar, “Machinability of hardened steel using alumina based Ceramic cutting tools”, International Journal of refractory Metals and Hard Materials, vol. 21, pp 109-117, 2003. T. Sornakumar, R. Krishnamurthy and C.V. Gogularathnam, “Machining performance of ZTA cutting tool”, 12th ISME conference, vol. 2, 2001. A.J.Thomas, and J.Antony, “A Comparative analysis of the Taguchi and DOE techniques in an aerospace environment”, International Journal of Productivity and Performance Management, vol. 54, pp-56, 2001. D. Mandal, S.K. Pal, and P. Saha, “Modeling of electrical discharge machining process using back propagation neural network and multiobjective optimizations using non-dominating sorting genetic algorithm-II”, Journal of Materials Processing Technology, vol. 186, 154-162, 2007. K. Wang, Hirpa L. Gelgele, Yi Wang , Qingfeng Yuan and Minglung Fang.,” A hybrid intelligent method for modeling the EDM process”, International Journal of Machine Tools & Manufacture, vol. 43, pp.995–999, 2003. P. J. Wang and K.M. Tsai, “Semi-empirical model on work removal and tool wear in electrical discharge machining”, Journal of Materials Processing Technology, vol. 114, Issue 1, pp. 1-17, 2001. K. Palanikumar, “Modeling and analysis for surface roughness in machining glass fiber reinforced plastics using response surface methodology”, Materials and Design, vol. 28, pp. 2611–2618, 2007. D. C. Montgomery, “Design and Analysis of Experiments (second ed.)”, Wiley, New John Wiley and Sons, New York, 1984. L. Robert Mason, F. Richard Gunst, Dallas, Texas, James L. Hess. “Statistical Design and Analysis of Experiments with applications to Engineering and Science (Second Edition)”, A John Wiley & sons publication, 2003.
Page 467
Proceedings of the International Conference On COMPUTERS, COMMUNICATION & INTELLIGENCE
22nd & 23rd July 2010 Organised by
Velammal College of Engineering & Technology
Viraganoor, Madurai – 625 009, India
International Conference on Computers, Communication & Intelligence 22nd & 23rd July 2010 CONFERENCE ORGANIZATION Chief Patron : Shri. M.V. Muthuramalingam, Chairman Organising Chair : Dr. N. Suresh Kumar, Principal Organising Secretaries : Dr. P. Alli and Dr. G. Manikandan
INTERNATIONAL ADVISORY COMMITTEE MEMBERS Dr. Kimmo Salmenjoki, Seinajoki University of Applied Sciences, Finland Dr. Fiorenzo Fianceschini, Polytechnico di Torino, Italy Dr. Namudri Kamesh, University of North Texas, USA Dr. Henry Selvaraj, University of Nevada, USA Dr. Kasim Mousa Al Aubidy, Philadelphia University, Jordan Dr. Jiju Antony, University of Strathclyde, UK Dr. Lorno Uden, University of Staffordshire, UK Dr. A. Vallavaraj, Caledonian College of Engineering, Sultanate of Oman Dr. Paulraj Murugesa Pandiyan, University of Malaysia, Perlis Dr. Arunagiri, Yanbu Industrial College, Kingdom of Saudi Arabia. Dr. Raja Sooriya Moorthi, Carnegie Melon University, USA Dr. Angappa Gunasekaran, University of Massachusetts, USA Dr. Sridhar Arjunan, RMIT University , Australia,
NATIONAL ADVISORY COMMITTEE MEMBERS Cmdr. Suresh Kumar Thakur, NRB, DRDO, India Dr. M. Mathirajan, Anna University, India Dr. Chitra T. Rajan, PSG Tech, India Dr. G. Arumugam, MKU, India Dr. S. Ibrahim Sadhar, Wins Infotech Pvt,. Ltd, India Dr. T. Devi, Bharathiar University, India Dr. L. Ganesan, AC Tech, India Dr. T. Purushothaman, GCT, India Dr. R. Murugesan, MKU, India Dr. B. Ramadoss, NIT, India Dr. S. Mercy Shalini, TCE, India Dr. Kannan Balasubramanian, MSEC, India Dr. K. Muneeswaran, MSEC, India Dr. K. Ramar, NEC, India Mr. Jegan Jothivel, Cisco Networking Academy, India
PAPER ID
PAPER TITLE
SESSION ID
PAGE NO. IN PROCEEDINGS
SESSION 1 AI002
Study of similarity metrics for genomic data using go-ontology
S1-01
87 - 94
AI005
Hybrid PSO based neural network classifier and decision tree for brain MRI mining Gap: genetic algorithm based power estimation technique for behavioral circuits Human action classification using 3d star skeletonization and rvm classifier Enhanced knowledge base representation technique for intelligent storage and efficient retrieval using knowledge based markup lFace detection using wavelet transform and rbf neural network
S1-02
95 - 100
S1-03
101 - 107
S1-04
108 - 115
S1-05
154 - 157
S1-06
303 - 306
Automated test case generation and performance analysis for GUI application The need of the hour – nosql technology for next generation data storage Intelligent Agent based Data Cleaning to improve the Accuracy of WiFi Positioning System Using Geographical Information System (GIS) Designing Health Care Forum using Semantic Search Engine & Diagnostic Ontology Optimization of Tool Wear in Shaping Process by Machine vision system Using Genetic Algorithm Framework for Comparison of Association Rule Mining using Genetic Algorithm A New Method For Solving Fuzzy Linear Programming With TORA
S1-07
178 - 187
S1-08
188 - 192
S1-09
24 - 30
S1-10
77 - 81
S1-11
453 - 456
S1-12
14 - 20
S1-13
237 - 239
S2-01
116 - 122
S2-02
316 - 319
COMN018
Relevance vector machine based gender classification using gait appearance features An energy efficient advanced data compression and decompression schemes for wsn Active noise control: a simulation study
S2-03
320 - 325
AI013
A survey on gait recognition using hmm model
S2-04
123 - 126
COMN022
Human motion tracking and behavior classification using multiple cameras
S2-05
131 - 134
AI006 AI007 COMP002 AI009 COMP016 COMP017 COMP115 COMP102 COMP135 COMP111 COMP114
SESSION 2 AI008 COMN013
COMP007
Adaptive visible watermarking and copy protection of reverted multimedia data
S2-06
168 - 173
COMP026
Hiding sensitive frequent item set by database extension
S2-07
357 - 362
COMP027
Integrated biometric authentication using finger print and iris matching Improvement towards efficient OPFET detector
S2-08
436 - 441
S2-09
417 - 420
S2-10
326 - 329
S2-11
369 - 376
COMN034
Texture segmentation method based on combinatorial of morphological and statistical operations using wavelets High Performance Evaluation of 600-1200V, 1-40A Silicon Carbide Schottky Barrier Diodes and Their Applications Using Mat L b based shape matching for trademarks retrieval Phase
S2-12
149 - 153
COMP146
The Medical Image segmentation
S2-13
212 - 215
S3-01
307 - 315
S3-02
158 - 167
S3-03
464 - 467
S3-04
174 - 177
S3-05
334 - 336
S3-06
442 - 452
Secure Multiparty Computation Based Privacy Preserving Collaborative Data Mining Towards Energy Efficient Protocols For Wireless Body Area Networks A cascade data mining approach for network anomaly Detection system Rule Analysis Based On Rough Set Data Mining Technique
S3-07
66 - 70
S3-08
207 - 211
S3-09
377 - 384
S3-10
291 - 296
On the Investigations of Design, Implementation, Performance and Evaluation issues of a Novel BD-SIIT Stateless IPv4/IPv6 T Role l of IPv6 over Fiber (FIPv6): Issues, Challenges and its The Impact on Hardware and Software. Entrustment based authentication protocol for mobile systems.
S3-11
260 - 269
S3-12
270 - 277
S3-13
389 - 392
COMP103 COMN028 COMP118
SESSION 3 AI014 COMP006 COMP142 COMP008 COMP013 COMP032 COMP038 COMP133 COMP119 COMP124 COMP128 COMP129 COMP137
An clustering approach based on functionality of genes for microarray data to find meaningful associations i i web based personalization of e-learning courseware using Semantic concept maps and clustering Modeling of Cutting Parameters for Surface Roughness in Machining A web personalization system for evolving user profiles in dynamic web sites based on web usage mining techniques and agent h l actionable knowledge within the organization using rough Creating set computing A new frame work for analyzing document clustering algorithms
SESSION 4 COMP022
Exploiting parallelism in bidirectional dijkstra for shortest-path computation Cld for improving overall throughput in wireless networks
S4-01
351 - 356
S4-02
46 - 49
S4-03
127 - 130
S4-04
6 - 13
S4-05
135 - 138
S4-06
139 - 148
S4-07
421 - 425
COMN027
Congestion management routing protocol in mobile adhoc networks Performance improvement in ad hoc networks using dynamic addressing Hierarchical zone based intrusion detection system for mobile adhoc networks. Implementing High Performance Hybrid Search Using CELL Processor Enhancing temporal privacy and source-location privacy in wsn routing by fft based data perturbation method Mixed-radix 4-2 butterfly fft/ifft for wireless communication
S4-08
203 - 206
COMP035
NTRU - public key cryptosystem for constrained memory devices
S4-09
55 - 59
COMP036
A novel randomized key multimedia encryption algorithm secure against several attacks Denial Of Service: New Metrics And Their Measurement
S4-10
60 - 65
S4-11
363 - 368
Fpga design of application specific routing algorithms for network on chip Selection of checkpoint interval in coordinated checkpointing protocol for fault tolerant open-mpi
S4-12
330 - 333
S4-13
216 - 223
S5-01
240 - 245
S5-02
343 - 350
S5-03
21 - 23
COMP024
Latest Trends and Technologies in Enterprise Resource Planning – ERP Integrating the static and dynamic processes in software development Content management through electronic document management system A Multi-Agent Based Personalized e-Learning Environment
S5-04
397 - 401
COMP109
Architecture Evaluation for Web Service Security Policy
S5-05
284 - 290
COMP110
Harmonics In Single Phase Motor Drives And Power Conservation.
S5-06
412 - 416
COMP030
Identification in the e-health information systems
S5-07
402 - 405
COMP138 COMN005 COMN020 COMN023 COMN024 COMN026
COMP037 COMP012 COMP019
SESSION 5 COMP116 COMP020 COMP023
COMP033
S5-08
297 - 302
S5-09
426 - 431
COMN033
A Robust Security metrics for the e-Healthcare Information Systems Theoretical Investigation Of Size Effect On The Thermal Properties Of Nanoparticles An efficient turbo coded ofdm system
S5-10
193 - 198
COMP018
Compval – a system to mitigate sqlia
S5-11
337 - 342
COMP126
Fault Prediction Using Conceptual Cohesion in Object Oriented System A Framework for Multiple Classifier Systems Comparison (MCSCF) A Comparative Study of Various Topologies and its performance analysis using WDM Networks
S5-12
256 - 259
S5-13
31 - 40
S5-14
457 - 463
S6-01
246 - 255
S6-02
224 - 229
S6-03
41 - 45
S6-04
1-5
COMP130
MRI Mammogram Image Segmentation using N Cut method and Genetic Algorithm with partial filters Localized Cbir for indexing image database
S6-05
278 - 283
COMP021
Particle swarm optimization algorithm in grid computing
S6-06
50 - 54
COMP149
Advancement in mobile technology Using BADA Enhancing the Life Time of Wireless Sensor Networks Using Mean Measure Mechanism Cloud Computing And Virtualization
S6-07
199 - 202
S6-08
82 - 86
S6-09
230 - 236
S6-10
432 - 435
COMP101
Dynamic Key Management to minimize communication latency for efficient group communication Towards Customer Churning Prevention through Class Imbalance
S6-11
71 - 76
COMP120
Membrane Computing - an Overview
S6-12
385 - 388
COMP148
Privacy Preserving Distributed Data Mining Using Elliptic Curve Cryptography Modeling A Frequency Selective Wall For Indoor Wireless Environment
S6-13
406 - 411
S6-14
393 - 396
COMP147
COMP127 COMP150
SESSION 6 COMP117 COMP112 COMP123 COMP125
COMP107 COMP113 COMP121
COMP149
A New Semantic Similarity Metric for Handling all Relations in WordNet Ontology Simplification of diagnosing disease through microscopic images of blood cells Efficient Apriori Hybrid Algorithm For Pattern Extraction Process
Paper Index Sl. No 1.
2. 3. 4. 5.
6. 7. 8. 9. 10. 11.
12. 13. 14. 15.
16. 17.
18. 19. 20.
Title MRI Mammogram Image Segmentation using NCut method and Genetic Algorithm with partial filters A.Pitchumani Angayarkanni Performance Improvement in Ad Hoc Networks Using Dynamic Addressing S.Jeyanthi & N.Uma Maheswari Framework for Comparison of Association Rule Mining using Genetic Algorithm K.Indira & S.Kanmani Content Management through Electronic Document Management System T.Vengattaraman, A.Ramalingam & P.Dhavachelvan Intelligent Agent based Data Cleaning to improve the Accuracy of WiFiPositioning System Using Geographical Information System (GIS) T.Joshva Devadas A Framework for Multiple Classifier Systems Comparison (MCSCF) P.Shanmugapriya & S.Kanmani Efficient Apriori Hybrid Algorithm For Pattern Extraction Process J.Kavitha, D.Magdalene Delighta Angeline & P.Ramasubramanian CLD for Improving Overall Throughput in Wireless Networks Dr. P. Seethalakshmi & Ms. A. Subasri Particle Swarm Optimization Algorithm In Grid Computing Mrs.R.Aghila, M.Harine & G.Priyadharshini NTRU - Public Key Cryptosystem For Constrained Memory Devices V.Pushparani & Kannan Balasubramaniam A Novel Randomized Key Multimedia Encryption Algorithm Security Against Several Attacks S. Arul Jothi Secure Multiparty Computation Based Privacy Preserving Collaborative Data Mining J.Bhuvana & Dr.T.Devi Towards Customer Churning Prevention through Class Imbalance M.Rajeswari & Dr.T.Devi Designing Health Care Forum Using Semantic Search Engine & Diagnostic Ontology Prof.Mr.V.Shunmughavel & Dr.P.Jaganathan An Enhancing the Life Time of Wireless Sensor Networks Using Mean Measure Mechanism P.Ponnu Rajan & D.Bommudurai Study of Similarity Metrics for Genomic Data Using GO-Ontology V.Annalakshmi,R. Priyadarshini &V. Bhuvaneshwari Hybrid PSO based neural network classifier and decision tree for brain MRI mining Dr.V.Saravanan & T.R.Sivapriya GAP: Genetic Algorithm based Power Estimation Technique for Behavioral Circuits Johnpaul C. I, Elson Paul & Dr. K. Najeeb Human Action Classification Using 3D Star Skeletonization and RVM Classifier Mrs. B. Yogameena, M. Archana & Dr. (Mrs) S. Raju Abhaikumar Relevance Vector Machine Based Gender Classification using Gait Appearance Features Mrs. B. Yogameena, M. Archana & Dr. (Mrs) S. Raju Abhaikumar
Page No. 1-5
6-13 14-20 21-23 24-30
31-40 41-45 46-49 50-54 55-59 60-65
66-70 71-76 77-81 82-86
87-94 95-100
101-107 108-115 116-122
21. A Survey on Gait Recognition Using HMM Model M.Siva Sangari & M.Yuvaraju 22. Congestion Management Routing Protocol In Mobile ADHOC Networks A. Valarmathi1 & RM. Chandrasekaran 23. Human Motion Tracking And Behaviour Classification Using Multiple Cameras M.P.Jancy & B.Yogameena 24. Hierarchical Zone Based Intrusion Detection System for Mobile Adhoc Networks. D G Jyothi & S.N Chandra shekara 25. Implementing High Performance Hybrid Search Using CELL Processor Mrs.Umarani Srikanth 26. Phase Based Shape Matching For Trademarks Retrieval B.Sathya Bama, M.Anitha & Dr.S.Raju 27. Enhanced Knowledge Base Representation Technique for Intelligent Storage and Efficient Retrieval Using Knowledge Based Markup Language A. Meenakshi, V.Thirunageswaran & M.G. Avenash 28. Semantic Web Based Personalization Of E-Learning Courseware Using Concept Maps And Clustering D.Anitha 29. Adaptive visible watermarking and copy protection of reverted multimedia data S.T.Veena & Dr.K.Muneeswaran 30. A Web Personalization System for evolving user profiles in Dynamic Web Sites based on Web Usage Mining Techniques and Agent Technology G.Karthik, R.Vivekanandam & P.Rupa Ezhil Arasi 31. Automated Test Case Generation and Performance Analysis for GUI Application Ms. A.Askarunisa & Ms. D. Thangamari 32. The Need Of The Hour - NOSQL Technology for Next Generation Data Storage
123-126 127-130 131-134 135-138 139-148 149-153 154-157
158-167
168-173 174-177
178-187 188-192
K.Chitra & Sherin M John
33. An Efficient Turbo Coded ofdm system Prof. Vikas Dhere 34. Advancement In Mobile Technologyusing Bada V.Aishwarya, J.Manibharathi & Dr.S.Durai Raj 35. Mixed-Radix 4-2 Butterfly FFT/IFFT For Wireless communication A.Umasankar & S.Vinayagakarthikeyan 36. Towards Energy Efficient Protocols For Wireless Body Area Networks Shajahan Kutty & J.A. Laxminarayana 37. The Medical Image Segmentation Hemalatha & R.Kalaivani 38. Selection of a Checkpoint Interval in Coordinated Checkpointing Protocol for Fault TolerantOpen MPI P.M.Mallikarjuna Shastry & K. Venkatesh 39. Simplification Of Diagnosing Disease Through Microscopic Images Of Blood Cells Benazir Fathima, K.V.Gayathiri Devi, M.Arunachalam & M.K.Hema 40. Cloud Computing And Virtualization R. Nilesh Madhukar Patil & Mr. Shailesh Somnath Sangle 41. A New Method For Solving Fuzzy Linear Programming With TORA S. Sagaya Roseline , A. Faritha Asma & E.C. Henry Amirtharaj 42. Latest Trends And Technologies In Enterprise Resource Planning – Erp B.S.Dakshayani
193-198 199-202 203-206 207-211 212-215 216-223
224-229 230-236 237-239 240-245
43. A New Semantic Similarity Metric for Handling all Relations in WordNet Ontology K.Saruladha, Dr.G.Aghila & Sajina Raj 44. Fault Prediction Using Conceptual Cohesion in Object Oriented System V.Lakshmi, P.V.Eswaripriya, C.Kiruthika & M.Shanmugapriya 45. On the Investigations of Design,Implementation, Performance and Evaluation issues of a Novel BD-SIIT Stateless IPv4/IPv6 Translator J.Hanumanthappa, D.H.Manjaiah & C.V.Aravinda 46. The Role of IPv6 over Fiber (FIPv6): Issues, Challenges and its Impact on Hardware and Software. J.Hanumanthappa, D.H.Manjaiah & C.V.Aravinda 47. Localized CBIR for Indexing Image Databases D.Vijayalakshmi & P. Vijayalakshmi 48. Architecture Evaluation for Web Service Security Policy B.Joshi.vinayak ,Dr.D.H. Manjaiah ,J. Hanumathappa & Nayak.Ramesh.Sunder 49. Rule Analysis Based On Rough Set Data Mining Technique P.Ramasubramanian, V.Sureshkumar & P.Alli 50. A Robust Security metrics for the e-Healthcare Information Systems Said Jafari, Fredrick Mtenzi, Ronan Fitzpatrick & Brendan O’Shea 51. Face Detection Using Wavelet Transform And Rbf Neural Network M.Madhu, M.Moorthi, S.Sathish Kumar & Dr.R.Amutha 52. An Clustering approach based on Functionality of Genes for Microarray data to find meaningful associations M.Selvanayaki & V.Bhuvaneshwari 53. An Energy Efficient Adavanced Data Compression And Decompression Schemes For Wsn G.Mohanbabu#1, Dr.P.Renuga#2 54. Active Noise Control: A Simulation Study Sivadasan Kottayi & N.K. Narayanan 55. Texture Segmentation Method Based On Combinatorial Of Morphological And Statistical Operations Using Wavelets V.Vijayapriya & Prof.K.R.Krishnamoorthy 56. FPGA Design Of Routing Algorithms For Network On Chip R.Anitha & Dr.P.Renuga 57. Creating Actionable Knowledge within the Organization using Rough set computing Mr.R.Rameshkumar, Dr.A.Arunagiri, Dr.V.Khanaa & Mr.C.Poornachandran 58. COMPVAL – A system to mitigate SQLIA S. Fouzul Hidhaya & Dr. Angelina Geetha 59. Integrating the Static and Dynamic Processes in Software Development V. Hepsiba Mabel, K. Alagarsamy & S. Justus 60. Exploiting Parallelism in Bidirectional Dijkstra for Shortest-Path Computation R.Kalpana, Dr. P.Thambidurai, R. Arvind kumar, S. Parthasarathi & Praful Ravi 61. Hiding Sensitive Frequent Item Set by Database Extension B. Mullaikodi & Dr. S.Sujatha 62. Denial Of Service:New Metrics And Their Measurement Dr.KannanBalasubramanian & P.Kavithapandian 63. High Performance Evaluation of 600-1200V, 1-40A Silicon Carbide Schottky Barrier Diodes and Their Applications Using Mat Lab K.Manickavasagan 64. A Cascade Data Mining Approach for Network Anomaly Detection System C. Seelammal
246-255 256-259 260-269
270-277
278-283 284-290 291-296 297-302 303-306 307-315
316-319 320-325 326-329
330-333 334-336 337-342 343-350 351-356 357-362 363-368 369-376
377-384
65. Membrane Computing - an Overview R.Raja Rajeswari & Devi Thirupathi 66. Entrustment Based Authentication Protocol For Mobile Systems. R.Rajalakshmi & R.S.Ponmagal 67. Modeling A Frequency Selective Wall For Indoor Wireless Environment. Mrs. K.Suganya, Dr.N.Suresh Kumar & P.Senthil Kumar 68. A Multi-Agent Based Personalized e-Learning Environment T. Vengattaraman, A. Ramalingam, P. Dhavachelvan & R.Baskaran 69. Identification in the E-Health Information Systems Ales Zivkovic 70. Privacy Preserving Distributed Data Mining Using Elliptic Curve Cryptography M.Rajalakshmi & T.Purusothaman 71. Harmonics In Single Phase Motor Drives And Energy Conservation Mustajab Ahmed Khan & Dr.A.Arunagiri 72. Improvement towards efficient OPFET detector Jaya V. Gaitonde & Rajesh B. Lohani 73. Enhancing Temporal Privacy and Source-Location Privacy in WSN Routing by FFT Based Data Perturbation Method R.Prasanna Kumar & T.Ravi 74. Theoretical nvestigation of size effect on the thermal properties of nanoparticles K.Sadaiyandi & M.A.Zafrulla Khan 75. Dynamic Key Management to minimize communication latency for efficient group communication Dr.P.Alli ,G.Vinoth Chakkaravarthy & R.Deepalakshmi 76. Integrated Biometric AuthenticationUsing Fingerprint and IRIS Matching A.Muthukumar & S.Kannan 77. New Framework for Analyzing Document Clustering Algorithms Mrs. J. Jayabharathy & Dr. S. Kanmani 78. Optimization of Tool Wear in Shaping Process by Machine vision system Using Genetic Algorithm S.Palani, G.Senthilkumar, S.Saravanan & J.Ragunesan A Comparative Study of Various Topologies and its performance analysis using WDM Networks 79. P. Poothathan, S. Devipriya & S. John Ethilton 80. Modeling of Cutting Parameters for Surface Roughness in Machining M. Aruna & P. Ramesh Kumar
385-388 389-392 393-396 397-401 402-405 406-411 412-416 417-420 421-425
426-431 432-435
436-441 442-452 453-456
457-463 464-467
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
MRI Mammogram Image Segmentation Using Ncut Method And Genetic Algorithm With Partial Filters (1)
S.Pitchumani Angayarkanni M.C.A,M.Phil,Ph.d Lecturer,Department of Computer Science, Lady Doak College, Madurai [email protected]
ABSTRACT: Cancer is one of the most common leading deadly diseases which affect men and women around the world. Among the cancer diseases, breast cancer is especially a concern in women. It has become a major health problem in developed and developing countries over the past 50 years and the incidence has increased in recent years. Recent trends in digital image processing are CAD systems, which are computerized tools designed to assist radiologists. Most of these systems are used for automatic detection of abnormalities. However, recent studies have shown that their sensitivity is significantly decreased as the density of breast increases. In this paper , the proposed algorithm uses partial filters to enhance the images and the Ncut method is applied to segment the malignant and benign regions , futher genetic algorithm is applied to identify the nipple position followed by bilateral subtraction of the left and the right breast image to cluster the cancerous and non cancerous regions. The system is trained using Back Propagation Neural Network algorithm. Computational efficiency and accuracy of the proposed system are evaluated based on the Frequency Receiver Operating Characteristic curve(FROC). The algorithm are tested on 161 pairs of digitized mammograms from MIAS database. The Receiver Operating Characteristic curve leads to 99.987% accuracy in detection of cancerous masses. Keywords: Filters, Normalized Cut, Segmentation, BPN, Genetic Algorithm and FROC.
INTRODUCTION: Breast cancer is one of the major causes for the increased mortality among women especially in developed countries. It is second most common cancer in women. The World Health Organization’s International estimated that more than 1,50,000 women worldwide die of breast cancer in year. In India, breast cancer accounts for 23% of all the female cancer death followed by cervical cancer which accounts to 17.5% in India. Early detection of cancer leads to significant improvements in conservation treatment. However, recent studies have shown that the sensitivity of these systems is significantly decreased as the density of the breast increased
Velammal College of Engineering and Technology, Madurai
while the specificity of the systems remained relatively constant. In this work we have developed automatic neuron genetic algorithmic approach to automatically detect the suspicious regions on digital mammograms based on asymmetries between left and right breast image. One of the major tool used for early detection of breast cancer is mammography. Mammography offers high quality images at low radiation doses and is the only widely accepted imaging method for routine breast cancer screening. Although mammography is widely used around the world for breast cancer detection, there are some difficulties when mammography is used for diagnosing breast cancer. One of the difficulties with mammography is that mammograms generally have low contrast compared with normal breast structure, and thus make it difficult for radiologists to interpret them. Studies show that the interpretation of mmaograms by radiologists could result in high rate of falsepositive and false-negative. This difficulty has caused high proportion of women without cancers to undergo breast biopsies and miss the breast treatment time. Several solutions were proposed in the past to increase accuracy and sensitivity of mammography and reduce unnecessary biopsies. Double reading of mmamograms is one of the solutions and has been advocated to reduce the proportion of missed cancers. The basic idea for double reading is to read the mammograms by two radiologists. However this solution is both costly and time consuming.Instead CAD has drawn attention from both computer scientists and radiologists in the interpretation of mammograms. CAD which integrates computer science, image processing , pattern recognition and artificial intelligence technologies can be defined as a diagnosis that is made by a radiologist who uses the output from a computerized analysis of medical images as a “second opinion” in detecting lesions and in making diagnostic decisions. It has been proven that this kind of system can improve the accuracy of breast diagnosis for early prediction of breast cancer. Computer aided breast cancer detection system is especially useful
Page 1
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
when the radiologist become tired of screening mammograms. In the CAD System for breast cancer, the detection of abnormal regions , such as calcification, mass and architectural distortion is the central task and the performance of a CAD system will depend on the performance of the detection of these abnormalities. There have been many proposed algorithms for detection of these abnormalities. In this paper we have introduced the detection of microcalcifications. As one of the early signs of breast cancer , mirocalcifications are tiny granule like deposits of calcium, which appear as small bright spots of mmaograms. Their size varies from 0.1 mm to 1mm. “Cluster: of MCs is defines as a group of three to five MCs within regions. Generally microcalcification clusters are important indication of possible cancer. This algorithm effectively and automatically detect MCs . 2. ALGORITHM DESIGN: There are four steps involved in the algorithm for the detection MCCs which is shown in the figure.
Thus, the product of the image matrix, which is usually very large because it represents the initial image (pixel table), by the filter yields a matrix corresponding to the processed image. 2.1.1 HIGH PASS FILTER: It allow high frequency areas to pass with the resulting image having greater detail resulting in a sharpened image. The boundary information of the enhanced image was extracted for visual evaluation. A high-pass (laplacian) filter was used for this purpose.
Figure 2: Mammogram Image enhanced using high pass filter
The table coefficients determine the properties of the filter. The following is an example of a 3 X 3 filter:
2.1.2) LOW PASS FILTER: Low pass filtering, otherwise known as "smoothing", is employed to remove high noise from a digital image. Noise is often introduced during the analog-to-digital conversion process as a side-effect of the physical conversion of patterns of light energy into electrical patterns . There are several common approaches to removing this noise: • If several copies of an image have been obtained from the source, some static image, then it may be possible to sum the values for each pixel from each image and compute an average. This is not possible, however, if the image is from a moving source or there are other time or size restrictions. • If such averaging is not possible, or if it is insufficient, some form of low pass spatial filtering may be required. There are two main types: • reconstruction filtering, where an image is restored based on some knowledge of the type of degradation it has undergone. Filters that do this are often called "optimal filters". • enhancement filtering, which attempts to improve the (subjectively measured) quality of an image for human or machine interpretability. Enhancement filters are generally heuristic and problem oriented One of the most important problems in image processing is denoising. Usually the procedure used for denoising, is dependent on the features of the image, aim of processing and also post-processing algorithms [5]. Denoising by lowpass filtering not only reduces the noise but also blurs the edges.
Velammal College of Engineering and Technology, Madurai
Page 2
Parti al Filter
Feature Extraction using NCut Segmentation
Genetic Algorith m
Multilay ered BPN
Fig 1: Flow Chart of Algorithm
2.1 PARTIAL FILTER FOR IMAGE ENHANCEMENT: A filter is a mathematical transformation (called a convolution product) which allows the value of a pixel to be modified according to the values of neighbouring pixels, with coefficients, for each pixel of the region to which it is applied. The filter is represented by a table (matrix), which is characterized by its dimensions and its coefficients, whose centre corresponds to the pixel concerned. 1
1
1
1
4
1
1
1
1
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Spatial and frequency domain filters are widely used as tools for image enhancement. Low pass filters smooth the image by blocking detail information. Mass detection aims to extract the edge of the tumor from surrounding normal tissues and background, high pass filters (sharpening filters) could be used to enhance the details of images.
method
the
Microcalcifications
are
clustered. Figure 4: After Normalized Cut Segmentation
Figure 3: Mammogram Image Enhanced Using Low Pass filter
2.2 IMAGE SEGMENTATION: The goal of image segmentation is to cluster pixels into salient image regions, i.e., regions corresponding to individual surfaces, objects, or natural parts of objects. In this we apply Normalized Cut method of segmentation to cluster microcalcification regions. Finally we outline the normalized cut approach of Shi and Malik [13].Here we seek a partition F and G = V − F of the affinity weighted,undirected graph (without source and sink nodes). In order to avoid partitions where one of F or G is a tiny region, Shi and Malik propose the normalized cut criterion, namely that F and G should minimize.
The computational efficiency 12.563 seconds on the 160x160 image. 2.3 GENETIC ALGORITHM: A partial filtering absed normalized cut method is used to generate a image to separate the breast and the non breast region . The GA enhances the breast border . Border detector detects the edges in the binary images , where each pixel takes on either the intensity value of zero for a non border pixel or one for border pixel. Each pixel in the binary map corresponds to an underlying pixel in the original image . In this proposed system , kernel is extracted from border points as a neighborhood array of pixels of the size 3*3 window of binary image. The binary kernels are considered population strings for GA. The corresponding kernels are extracted from gray level mammogram image using spatial coordinate points and the sum of the intensity values are considered as the fitness value . After identifying initial population and the fitness value , the genetic operator can be applied to generate a new population. Reproduction operator produces new string for crossover. Reproduction is implemented as linear search through roulette wheel with slots weighted in proportion to kernel fitness values. In this function, a random number multiplies the sum of population fitness called as stopping point.
Figure 5: GA
Note any segmentation technique can be used for generating proposals for suitable regions F, for which N(F, V − F) could be evaluated. Indeed, the SMC approach above can be viewed as using S and T to provide lower bounds on the terms L(F, V ) and L(G, V ) (namely L(S, V ) and L(T, V ), respectively), and then using the S-T min cut to globally minimize L(F,G) subject to S C F and T C G. Using this
Velammal College of Engineering and Technology, Madurai
Page 3
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.4 GENERATING THE ASYMMETRIC IMAGE: After the images were aligned, bilateral subtraction was performed [47,48] by subtraction was performed by subtracting the digital matrix of the left breast image from the digital matrix of the right breast image. Microcalcification in the right breast image have positive pixel values in the image obtained after subtraction, while microcalcification in the left breast image have negative pixel values in the subtracted image. As a result, two new images were generated: one with positive values and the other with negative values. The most common gray level was zero, which indicated no difference between the left and right images. Simple linear stretching of the two generated images to cover the entire available range of 1024 gray levels was then calculated. The difference between corresponding pixels contains important information that can be used to discriminate between normal and abnormal tissue. The asymmetry image can be thresholded to extract suspicious regions. To generate FROC curve, the asymmetry image is thresholded using ten different intensity values ranges from 50-150. Figure 6 shows a asymmetry image and connected regions extracted based on thresholding to obtain a progressively larger number of high difference pixels.
Figure 6
Figure 7a) Steps involved in automated Classification using Ant Colony Optimization
2.6. ROC CURVE: Finally the technique was evaluated on the mammograms randomly selected from the non-suspicious section of the data base. The method outlined small regions in 5 out of the 15 non suspicious mammograms. The areas identified were generally very small compared to those in abnormal mammograms
Asymmetric images
Two different techniques are used in the interpretation of mammogram. The first technique consists of systematic search of each mammogram for visual pattern symptomatic tumors. Such as, a bright, approximately circular blob with hazy boundary might indicate the presence of a circumscribed mass. The second technique, the asymmetric approach , consists of systematic comparison of corresponding regions in the left and the right breast. 2.5. BPN TRAINING: In addition, a backpropagation artificial neural network (BPANN) was also developed and evaluated on the same data. The parameters for ANN training were published before. Figure 5 compare the ROC curves for the LGP and the BPANN algorithms respectively. The BP-ANN yielded an ROC area index of Az=0.88±0.01. Our GP approach achieved a statistically significantly better performance with Az=0.91±0.01.
Velammal College of Engineering and Technology, Madurai
Figure 8 Lesion Areas detected for Abnormal and Non-Suspicious cases (large image extracts). [Figures (a) and (b) are presented at different ordinate scales]
Fig 8(a) shows the extracted areas for the abnormal lesions. (Image sequence 54 - 87 are stellate lesions and 74 to 100 are regular masses). We first establish whether these represent two different populations, by applying a Mann-Whitney (Wilcoxon rank sum) non-parametric test, since it is unrealistic to presume any specific underlying distribution. Median values are 450 and 1450 pixels respectively which produce a confidence level of 85% that the two data sequences emanate from distinct populations. Since this is not significant at normally acceptable levels we can compare the abnormals as a single distribution against the nonsuspicious set, Fig 8(b). Using the same test, median values of 5500 and 10 pixels for the two distributions are established, giving a confidence level of greater than 97.5%
Page 4
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
that the two distributions are different, suggesting that our PROTOCOLS ARE AN EFFECTIVE METHOD OF AREA DETECTION. CONCLUSION: The proposed algorithms are tested on 161 pairs of digitized mammograms from Mammographic Image analysis Society(MIAS) database. A free response receiver operating characteristic (FROC) curve is generated for the mean value of the detection rate for all the 161 pairs of mammograms in the MIAS database, to evaluate the performance of the proposed method. There is no doubt that for the immediate future mammography will continue to play a major role in the detection of breast cancer. The ultimate objective of this thesis was to identify tumor or masses in breast tissue. Since hamartomas consists of normal breast tissue with abnormal proportions and the first step was try to identify the different tissue type in mammography with normal breast tissue. The important features have been extracted from the Normalized cut method of the each sub image using various statistical techniques. The Genetic algorithm has been implemented and the breast border was identified from the clustered image. The tests that were carried out using a set of 117 tissues samples, 67 benign and 50 malignant. The result analysis has given a sensitivity of 99.8%, a specificity of 99.9% and an accuracy above 99.9%, which means encouraging results. The preliminary results of this approach are very promising in characterizing breast tissue. REFERENCES: [1] Bosch. A.; Munoz, X.; Oliver.A.; Marti. J., Modeling and Classifying Breast Tissue Density in Mammograms, Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference on Volume 2, Issue , 2006 Page(s): 1552 – 15582. [2] Dar-Ren Chena, Ruey-Feng Changb, ChiiJen Chenb, Ming-Feng Hob, Shou-Jen Kuoa, ShouTung Chena, Shin-Jer Hungc, Woo Kyung Moond, Classification of breast ultrasound images using fractal feature, ClinicalImage, Volume 29, Issue4, Pages 234-245. [3] Suri, J.S., Rangayyan, R.M.: Recent Advances in Breast Imaging, Mammography,and Computer-Aided Diagnosis of Breast Cancer. 1st edn. SPIE (2006) [4] Hoos, A., Cordon-Cardo, C.: Tissue microarray pro.ling of cancer specimens and cell lines: Opportunities and limitations. Mod. Pathol. 81(10), 1331–1338 (2001) [5] Lekadir, K., Elson, D.S., Requejo-Isidro, J., Dunsby, C., McGinty, J., Galletly, N.,Stamp, G., French, P.M., Yang, G.Z.: Tissue characterization using dimensionality reduction and .uorescence imaging. In: Larsen, R., Nielsen, M., Sporring, J. (eds.) MICCAI 2006. LNCS, vol. 4191, pp. 586–593. Springer, Heidelberg (2006).
[6] A. Papadopoulos, D. I. Fotiadis, and A. Likas,“An Automatic microcacalcification Detection System Based On a Hybrid Neural Network Classifier,” Artificial Intelligence in Medicine,vol. 25, pp. 149-167, 2002. [7] A. Papadopoulos, D. I. Fotiadis, and A. Likas, “Characterization of Clustered microcalcifications in Digitized Mammograms Using Neural Networks and Support Vector Machine,” Artificial Intelligence in Medicine,vol. 34, pp. 141-150, 2005. [8] R. Mousa, Q. Munib, and A. Moussa, “Breast Cancer Diagnosis System based in Wavelet Analysis and Fuzzy-Neural,” Expert Systems with Applications, vol. 28, pp. 713-723, 2005.
Velammal College of Engineering and Technology, Madurai
Page 5
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Performance Improvement in Ad Hoc Networks Using Dynamic Addressing S.Jeyanthi#1, N.Uma Maheswari*2 #
Lecturer, Computer Science Department PSNA College of Engg & Tech,Dindigul,Tamilnadu,India 1
[email protected]
*
Assistant Professor PSNA College of Engg & Tech,Dindigul,Tamilnadu,India 2
[email protected]
Abstract Dynamic addressing refers to the assignment of IP addresses automatically. In this paper we propose the scalable routing in ad hoc networks. It is well known that the current ad hoc protocol do not scale to work efficiently in networks of more than a few hundred nodes. Most current adhoc routing architectures use flat static addressing and thus, need to keep track of each node individually, creating a massive overhead problem as the network grows. In this paper, we propose that the use of dynamic addressing can enable scalable routing in adhoc networks. We provide an initial design of a routing layer based on dynamic addressing, and evaluate its performance. Each node has a unique permanent identifier and a transient routing address, which indicates its location in the network at any given time. The main challenge is dynamic address allocation in the face of node mobility. We propose mechanisms to implement dynamic addressing efficiently. Our initial evaluation suggests that dynamic addressing is a promising approach for achieving scalable routing in large adhoc and mesh networks.
Keywords — Adhoc networks, Flat static addressing, Dynamic addressing, Unique permanent identifier. I. Introduction
Adhoc
networking technology has advanced tremendously but it has yet to become a widely deployed technology. Ad hoc networks research seems to have downplayed the importance of scalability. In fact, current ad hoc architectures do not scale well beyond a few hundred nodes. Existing Ad Hoc Routing Layers do not support several hundred nodes and lack of scalability. It uses flat static addressing. It creates a massive Routing overhead. It increases searching time (not optimal solution). The easy-touse, self-organizing nature of ad hoc networks make them attractive to a diverse set of applications. Today, these are usually limited to smaller deployments, but if we can solve
Velammal College of Engineering and Technology, Madurai
the scalability problem, and provide support for heterogeneous means of connectivity, including directional antennas, communication lasers, even satellites and wires, ad hoc and mesh-style networking is likely to see adoption in very large networks as well. Large-scale events such as disaster relief or rescue efforts are highly dependent on effective communication capabilities. Such efforts could benefit tremendously from the use of self-organizing networks to improve the communications and monitoring capabilities available. Other interesting candidate scenarios are community networks in dense residential areas, large scale, long-range networks in developing regions, and others, where no central administrator exists, or where administration would prove too costly. The current routing protocols and architectures work well only up to a few hundred nodes. Most current research in ad hoc networks focus more on performance and power consumption related issues in relatively small networks, and less on scalability. The main reason behind the lack of scalability is that these protocols rely on flat and static addressing. With scalability as a partial goal, some efforts have been made in the direction of hierarchical routing and clustering [1] [2] [3]. These approaches do hold promise, but they do not seem to be actively pursued. It appears to us as if these protocols would work well in scenarios with group mobility [4], which is also a common assumption among cluster based routing protocols. We examine that whether dynamic addressing is a feasible way to achieve scalable adhoc routing. By”scalable” we mean thousands up to millions of nodes in an ad hoc or mesh network. With dynamic addressing, nodes change addresses as they move, so that their addresses have a topological meaning. Dynamic addressing simplifies routing but introduces two new problems: address allocation, and address lookup. As a guideline, we identify a set of properties that a scalable and efficient solution must have: Localization of overhead: a local change should affect only the immediate neighborhood, thus limiting the overall overhead incurred due to the change.
Page 6
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Lightweight, decentralized protocols: To avoid the responsibility at any individual node, and to keep the necessary state to be maintained at each node as small as possible. Zero-configuration: To remove the need for manual configuration beyond what can be done at the time of manufacture. Minimal restrictions on hardware: Omni directional link layers do not scale to large networks. Localization technologies, such as GPS, may limit protocol applicability. We present a complete design including address allocation, routing and address lookup mechanisms, and provide thorough evaluation results for the address allocation and routing components. First, we develop a dynamic addressing scheme, which has the necessary properties mentioned above. Our scheme separates node identity from node address, and uses the address to indicate the node’s current location in the network. Second, we study the performance of a new routing protocol, based on dynamic addressing, through analysis and simulations. The address allocation scheme uses the address space efficiently on topologies of randomly and uniformly distributed nodes, empirically resulting in Average routing table size< 2 log2 n Where n is the number of nodes in the network. We describe a new approach to routing in ad hoc networks, and compare it to the current routing architectures. However, the goal is to show the potential of this approach and not to provide an optimized protocol. We believe that the dynamic addressing approach is a viable strategy for scalable routing in ad hoc networks. II. Related Work In most common IP-based ad hoc routing protocols [5] [7] [8], addresses are used as pure identifiers. Without any structure in the address space, there are two choices: either keep routing entries for every node in the network, or resort to flooding route requests throughout the network upon connection setup. However, this approach can be severely limiting as location information is not always available and can be misleading in, among others, non-planar networks. For a survey of ad hoc routing, see [9]. In the Zone Routing Protocol (ZRP) [10] and Fisheye State Routing (FSR) [11], nodes are treated differently depending on their distance from the destination. In FSR, link updates are propagated more slowly the further away they travel from their origin, with the motivation that changes far away are unlikely to affect local routing decisions. In ZRP is a hybrid reactive/ proactive protocol, where a technique called border casting is used to limit the damaging effects of global broadcasts. In multilevel-clustering approaches such as Landmark [12], LANMAR [3], L+ [13], MMWN [1] and Hierarchical State Routing (HSR) [2], certain nodes are elected as cluster
Velammal College of Engineering and Technology, Madurai
heads. These cluster heads in turn select higher level cluster heads, up to some desired level. A node’s address is defined as a sequence of cluster head identifiers, one per level, allowing the size of routing tables to be logarithmic in the size of the network, but easily resulting in long hierarchical addresses. In HSR, for example, the hierarchical address is a sequence of MAC addresses, each of which is 6 bytes long. A problem with having explicit cluster heads is that routing through cluster heads creates traffic bottlenecks. In Landmark, LANMAR and L+, this is partially solved by allowing nearby nodes route packets instead of the cluster head, if they know a route to the destination. Our work is, as far as we know, the first attempt to use this type of addressing in ad hoc networks. Tribe [14] is similar to DART at a high level, in that it uses a two phase process for routing: first address lookup, and then routing to the address discovered. However, the tree-based routing strategy used in Tribe bears little or no resemblance to the area based approach in DART. Tree-based routing may under many circumstances suffer from severe traffic concentration at nodes high up in the tree, and a high sensitivity to node failure. III. Overview of Network Architecture In this section, we present sketch of network architecture shown in figure 1, which could utilize the new addressing scheme effectively. In our approach, we separate the routing address and the identity of a node. The routing address of a node is dynamic and changes with node movement to reflect the node’s location in the network topology. DART
Cluster creation
Address allocation Mapping Distributed lookup table
Routing Figure.1 Overall system design
Page 7
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The identifier is a globally unique number that stays the same throughout the lifetime of the node. For ease of presentation, we can assume for now that each node has a single identifier. When a node joins the network, it listens to the periodic routing updates of its neighboring nodes, and uses these to identify an unoccupied address. The joining node registers its unique identifier and the newly obtained address in the distributed node lookup table. Due to mobility, the address may subsequently be changed and then the lookup table needs to be updated. When a node wants to send packets to a node known only by its identifier, it will use the lookup table to find its current address. Once the destination address is known the routing function takes care of the communication. The routing function should make use of the topological meaning that our routing addresses possess. We start by presenting two views of the network that we use to describe our approach: a) the address tree, and b) the network topology. Address Tree: In this abstraction, we visualize the network from the address space point of view. Addresses are l bit binary numbers, al-1, . . . , a0. The address space can be thought of as a binary address tree of l + 1 level, as shown in figure 2. The leaves of the address tree represent actual node addresses; each inner node represents an address sub tree a range of addresses with a common prefix. • Level 0 sub tree is a single address. • Level 1 sub tree has a 2 bit prefix and can contain up to two leaf nodes. • Level 2 sub tree containing addresses [100] through [111].
each sub tree of the address tree are enclosed with dotted lines. Note that the set of nodes from any sub tree in figure 2 induces a connected sub graph in the network topology in figure 3.
Figure.3 A network topology with node addresses assigned.
The nodes that are close to each other in the address space should be relatively close in the network topology. More formally, we can state the following constraint. Prefix Sub graph Constraint: The set of nodes that share a given address prefix form a connected sub graph in the network topology. This constraint is fundamental to the scalability of our approach. Intuitively, this constraint helps us map the virtual hierarchy of the address space onto the network topology. The longer the shared address prefix between two nodes, the shorter the expected distance in the network topology. Finally, let us define two new terms that will facilitate the discussion in the following sections. A Level-k sub tree of the address tree is defined by an address prefix of (l-k) bits, as shown in figure 2. For example, a Level0 sub tree is a single address or one leaf node in the address tree. A Level-1 sub tree has a (l-1)-bit prefix and can contain up to two leaf nodes. In figure 1, [0xx] is a Level-2 sub tree containing addresses [000] through [011]. Note that every Level-k sub tree consists of exactly two Level-(k - 1) sub trees. We define the term Level-k sibling of a given address to be the sibling of the Level-k sub tree to which a given address belongs. By drawing entire sibling sub trees as triangles, we can create abstracted views of the address tree, as shown in figure 4.
Figure.2 Address tree of 3-bit binary address space.
For presentation purposes, nodes are sorted in increasing address order, from left to right. The actual physical links are represented by dotted lines connecting leaves in fig 3. Network Topology: This view represents the connectivity between nodes. In figure 3, the network from figure 2 is presented as a set of nodes and the physical connections between them. Each solid line is an actual physical connection, wired or wireless, and the sets of nodes from
Velammal College of Engineering and Technology, Madurai
Figure.4 Routing entries corresponding to figure 2. Node 100 has entries for sub trees 0xx, 11x (null entry) and 101.
Here, we show the siblings of all levels for the address [100] as triangles: the Level-0 sibling is [101], Level-1 is [11x], and the Level-2 sibling is [0xx]. Note that each address has exactly one Level-k sibling, and thus at most l siblings in total.
Page 8
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Finally, we define the identifier of a sub tree to be the min of the identifiers of all nodes that have addresses from that sub tree. In cases where the prefix sub graph constraint is temporarily violated, two disconnected instances of the address sub tree exist in the network. In this case, each instance is uniquely identified by the min of the subset of identifiers that belongs to its connected sub graph. Our addressing and routing schemes have several attractive properties. First, they can work with omni directional and directional antennas as well as wires. Second, we do not need to assume the existence of central servers or any other infrastructure, nor do we need to assume any geographical location information, such as GPS coordinates. However, if infrastructure and wires exist, they can, and will, be used to improve the performance. Third, we make no assumptions about mobility patterns, although high mobility will certainly lead to increased overhead and decreased throughput. Finally, since our approach was designed primarily for scalability, we do not need to limit the size of the network; most popular ad hoc routing protocols today implicitly impose network size restrictions. IV. Routing In this work, we use a form of proactive distance-vector routing, made scalable due to the hierarchical nature of the address space. Although we have chosen to use distance vector routing, we would like to point out that many of the advantages of dynamic addressing can be utilized by a linkstate protocol as well. Each node keeps some routing state, routing state about a node’s Level-i sibling is stored at position i in each of the respective arrays. The routing state for a sibling contains the information necessary to maintain a route toward a node (any node) in that sub tree. The address field contains the current address of the node, and bit i of the address is referred to as address[i], where i = 0 for the least significant bit of the address. Arrays next hop and cost are self-explanatory. The id array contains the identifier of the sub tree in question. As described earlier, the identifier of a sub tree is equal to the lowest out of all the identifiers of the nodes that constitute that sub tree. Finally, route log[i] contains the log of the current route to the sibling at level i, where bit b of log i is referenced by the syntax route log[i][b]. To identify the most significant bit that differs between the current node’s address and the destination address. In this case, the most significant differing bit is bit number 2. The node then looks up the entry with index two in the next hop table, and then sends the packet there. In our example, this is the neighbor with address [011]. The process is repeated until the packet has reached the given destination address. The hierarchical technique of only keeping track of sibling sub trees rather than complete addresses has three immediate benefits. One,
Velammal College of Engineering and Technology, Madurai
the amount of routing state kept at each node is drastically reduced. Two, the size of the routing updates is similarly reduced. Three, it provides an efficient routing abstraction such that routing entries for distant nodes can remain valid despite local topology changes in the vicinity of these nodes. A. Loop Avoidance DART uses a novel scheme for detecting and avoiding routing loops, which leverages the hierarchical nature of the address space to improve scalability. A simple way of implementing this is to concatenate a list of all visited nodes in the routing update, and to have nodes check this list before accepting an update. However, this approach has a scalability problem, in that routing updates will quickly grow to unwieldy sizes. Instead, DART makes use of the structured address space to create a new kind of loop avoidance scheme. In order to preserve scalability, we generalize the loop freedom rule above. For each sub tree, once a routing entry has left the sub tree, it is not allowed to re-enter. This effectively prevents loops, and can be implemented in a highly scalable manner. V. Node Lookup We propose to use a distributed node lookup table, which maps each identifier to an address, similar to what we proposed in [5]. Here, we assume that all nodes take part in the lookup table, each storing a few7 <identifier, address> entries. However, this node lookup scheme is only one possibility among many, and more work is needed to determine the best lookup scheme to deploy. For our proposed distributed lookup table, the question now arises: which node stores a given <identifier, address> entry? Let us call this node the anchor node of the identifier. We use a globally, and a priori, known hash function that takes an identifier as argument and returns an address where the entry can be found. If there exists a node that occupies this address, then that node is the anchor node. If there is no node with that address, then the node with the least edit distance between its own address and the destination address is the anchor node. To route packets to an anchor node, we use a slightly modified routing algorithm: If no route can be found to a sibling sub tree indicated by a bit in the address, that bit of the address is ignored, and the packet is routed to the sub tree indicated by the next (less significant) bit. When the last bit has been processed, the packet has reached its destination. This method effectively finds the node with the address minimum edit distance to the address returned by the hash function. For example, using figure 3 for reference, let’s assume a node with identifier ID1 has a current routing address of [010]. This node will periodically send an updated entry to the lookup table, namely <ID1, 010>. To figure out where to send the entry, the node uses the hash function to calculate an address, like so: hash (ID1). If the returned address is [100], the packet will simply be routed to the node with that address. However, if the returned address
Page 9
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
was instead [111], the packet could not be routed to the node with address [111] because there is no such node. In such a situation, the packet gets automatically routed to the node with the most similar address, which in this case would be [101] A. Improved Scalability We would like to stress that all node lookup operations use unicast only: no broadcasting or flooding is required. This maintains the advantage of proactive and distance vector based protocols over on-demand protocols: the routing overhead is independent of how many connections are active. When compared with other distance vector protocols, our scheme provides improved scalability by drastically reducing the size of the routing tables, as we described earlier. In addition, updates due to a topology change are in most cases contained within a lower level sub tree and do not affect distant nodes. This is efficient in terms of routing overhead. To further improve the performance of our node lookup operations, we envision using the locality optimization technique described in [5]. Here, each lookup entry is stored in several locations, at increasing distance from the node in question. By starting with a small, local lookup and gradually going to further away locations, we can avoid sending lookup requests across long distances to find a node that is nearby. B. Coping with Temporary Route Failures On occasion, due to link or node failure, a node will not have a completely accurate routing table. This could potentially lead to lookup packets, both updates and requests, terminating at the wrong node. The end result of this is that requests cannot be promptly served. In an effort to reduce the effect of such intermittent errors, a node can periodically check the lookup entries it stores, to see if a route to a more suitable host has been found. If this should be the case, the entry is forwarded in the direction of this more suitable host. Requests are handled in a similar manner: if the request could not be answered with an address, it is kept in a buffer awaiting either the arrival of the requested information, or the appearance of a route to a node which more closely matches the key requested. This way, even if a request packet arrives at the anchor node before the update has reached it; the request will be buffered and served as soon as the update information is available. C. Practical Considerations Due to the possibility of network partitioning and node failure, it is necessary to have some sort of redundancy mechanism built-in. We have opted for a method of periodic refresh, where every node periodically sends its information to its anchor node. By doing so, the node ensures that if its anchor node should become unavailable, the lookup information will be available once again within one refresh period. Similarly, without a mechanism of expiry, outdated
Velammal College of Engineering and Technology, Madurai
information may linger even after a node has left the network. Therefore, we set all lookup table entries to expire automatically after a period twice as long as the periodic refresh interval. VI. Dynamic Address Allocation To assess the feasibility of dynamic addressing, we develop a suite of protocols that implement such an approach. Our work effectively solves the main algorithmic problems, and forms a stable framework for further dynamic addressing research. When a node joins an existing network, it uses the periodic routing updates of its neighbors to identify and select an unoccupied and legitimate address. It starts out by selecting which neighbor to get an address from the neighbor with the highest level insertion point is selected as the best neighbor. The insertion point is defined as the highest level for which no routing entry exists in a given neighbor’s routing table. However, the fact that a routing entry happens to be unoccupied in one neighbor’s routing table does not guarantee that it represents a valid address choice. We discuss how the validity of an address is verified in the next subsection. The new node picks an address out of a possibly large set of available addresses. In our current implementation, we make nodes pick an address in the largest unoccupied address block. For example, in figure 4, a joining node connecting to the node with address [100] will pick an address in the [11x] sub tree. Figure 5 illustrates the address allocation procedure for a 3-bit address space.
Figure.5 Address allocation procedure for a 3-bit address space
There are several ways to choose among the available addresses, and we have presented only one such method. However, it has turned out that this method of address selection works well in simulation trials. Under steady-state, and discounting concurrency, the presented address selection technique leads to a legitimate address allocation: the joining node is by definition connected to neighbor it got its new address from, and the new address is taken from one of the neighbors’ empty sibling sub trees, so the prefix sub graph constraint is satisfied. Node A starts out alone with address [000]. When node B joins the network, it observes that A has a null routing entry corresponding to the sub tree [1xx], and
Page 10
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
picks the address [100]. Similarly when C joins the network by connecting to B, C picks the address [110]. Finally, when D joins via A, A’s [1xx] routing entry is now occupied. However, the entry corresponding to sibling [01x] is still empty, and so, D takes the address [010]. Merging Networks Efficiently - DART handles the merging of two initially separate networks as part of normal operations. In a nutshell, the nodes in the network with the higher identifier join the other network one by one8. The lower id network absorbs the other network slowly: the nodes at the border will first join the other network, and then their neighbors join them recursively. Dealing with Split Networks - Here, we describe how we deal with network partitioning. Intuitively, each partition can keep its addresses, but one of the partitions will need to change its network identifier. In this situation, there are generally no constraint violations. This reduces to the case where the node with the lowest identifier leaves the network. Since the previous lowest identifier node is no longer part of the network, the routing update from the new lowest identifier node can propagate through the network until all nodes are aware of the new network identifier. VII. Maintaining the Dynamic Routing Table While packet forwarding is a simple matter of looking up a next hop in a routing table, maintaining a consistent routing state does involve a moderate amount of sophistication. In addition to address allocation, loop detection and avoidance is crucial to correct protocol operation. DART nodes use periodic routing updates to notify their neighbors of the current state of their routing table. If, within a constant number of update periods, a node does not hear an update from a neighbor, it is removed from the list of neighbors, and its last update discarded. Every period, each node executes Refresh() function. Refresh() checks the validity of its current address, populates a routing table using the information received from its neighbors, and broadcasts a routing update. When populating the routing table, the entry for each level, i, in the received routing update is inspected is sequence, starting at the top level. For neighbors where the address prefix differs at bit i, we create a new routing entry, with a one-hop distance. It also has an empty route log, with the exception of bit i, which represents the level-i sub tree boundary that was just crossed. The sub tree identifier is computed using the id array. After this, the procedure returns, as the remaining routing information is internal to the neighbor’s sub tree, and irrelevant to the current node. For nodes with the same address prefix as the current node, we go on to inspect their routing entry for level i. First, we ensure that the entry is loop free. If so, then keep the routing entry as long as the identifier of the entry is the same or
Velammal College of Engineering and Technology, Madurai
smaller than what is already in the routing table, and as long as the distance is smaller. VIII. Simulation Results We conduct our experiments using two simulators. One is the well known ns-2 network simulator. In ns-2, we used the standard distribution; version 2.26 used the standard values for the Lucent Wave LAN physical layer, and the IEEE 802.11 MAC layer code, together with patch for a retry counter bug recently identified by Dan Berger at UC Riverside9. For all of the ns-2 simulations, we used the Random Waypoint mobility model with up to 800 nodes and a maximum speed of 5 m/s, a minimum speed of 0.5 m/ s, a maximum pause time of 100 seconds and a warm-up period of 3600 seconds10. The duration of all the ns-2 simulations was 300 seconds11, wherein the first 60 seconds are free of data traffic, allowing the initial address allocation to take place and for the network to thereby organize itself. The size of the simulation area was chosen to keep average node degree close to 9. For example, for a 400-node network, the size of the simulation area was 2800x2800 meters. This was done in order to maintain a mostly connected topology. Mobility parameters were chosen to simulate a moderately mobile network. DART is not suitable for networks with very high levels of mobility, as little route aggregation benefits are to be had when the current location of most nodes bear little relation to where these nodes were a few seconds ago. Our simulations focus on the address allocation and routing aspects of our protocol, not including the node lookup layer, which is replaced by a global lookup table accessible by all nodes in the simulation. The choice of lookup mechanism (for example distributed, hierarchical, replicated, centralized, or out-of-band) should be determined by network characteristics, and performance may vary depending on what mechanism is used. Here follows a summary of our findings. DSDV, due to its periodic updates and flat routing tables, experiences very high overhead growth as the network grows beyond 100 nodes, but nevertheless performs well in comparison with other protocols in the size ranges studied. AODV, due to its reactive nature, suffers from high overhead growth both as the size of the network, and the number of flows, grows. While AODV performs very well in small networks, the trend suggests that it is not recommendable for larger networks .DSR, in our simulations, performed well in small networks, and never experienced high overhead growth, likely due to its route caching functionality. However, due to excessive routing failures, DSR demonstrated unacceptable performance in larger networks. Finally, DART, demonstrated its scalability benefits in terms of no overhead growth with the number of flows, and logarithmic overhead growth with network size. The simulation results shown in Figure 6, 7, 8.
Page 11
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IX. Performance Evaluation We compare our protocol to reactive protocols (AODV, DSR) and a proactive protocol (DSDV). Our results shown in table.1 suggest that dynamic addressing and proactive routing together provide significant scalability advantages and high level addressing. DART reduces the overhead growth with the number of flows, and logarithmic overhead growth with network size shown in figure 1. No of Transmitted packets No of Lost packets Figure.6 Communication between source and destination
DSDV DSR AODV DART 10726 5138 23349 42661 82
234
1003
1654
Table.1 DART performance
Figure.7 Destination became member in group 3 Figure.9 Throughput vs. Network Size (Nodes)
Figure.8 Destination leaving group3 and move towards group4
Velammal College of Engineering and Technology, Madurai
X. Conclusion We proposed Dynamic address routing, an initial design toward scalable ad hoc routing. We outline the novel challenges involved in a dynamic addressing scheme, and proceeded to describe efficient algorithmic solutions. We show how our dynamic addressing can support scalable routing. We demonstrate, through simulation and analysis, that our approach has promising scalability properties and is a viable alternative to current ad hoc routing protocols. First, we qualitatively compare proactive and reactive overhead and determine the regime in which proactive routing exhibits less overhead that its reactive counterpart. Large scale simulations show that the average routing table size with DART grows logarithmically with the size of the network. Second, using the ns-2 simulator, we compare our routing scheme to AODV,
Page 12
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
DSR and DSDV, and observe that our approach achieves superior throughput, and with considerably smaller overhead, in networks larger than 400 nodes. The trend in simulated overhead, together with the analysis provided, strongly indicate that DART is the only feasible routing protocol for large networks. We believe that dynamic addressing can be the basis for ad hoc routing protocols that for massive adhoc and mesh networks. REFERENCES [1] Ram Ramanathan and Martha Steenstrup, “Hierarchically-organized, multihop mobile wireless networks for quality-of-service support,” Mobile Networks and Applications, vol. 3, no. 1, pp. 101–119, 1998. [2] Guangyu Pei, Mario Gerla, Xiaoyan Hong, and Ching-Chuan Chiang, “A wireless hierarchical routing protocol with group mobility,” in WCNC, 1999. [3] G. Pei, M. Gerla, and X. Hong, “Lanmar: Landmark routing for large scale wireless ad hoc networks with group mobility,” in ACM MobiHOC’00, 2000. [4] X. Hong, M. Gerla, G. Pei, and C. Chiang, “A group mobility model for ad hoc wireless networks,” 1999. [5] J. Eriksson, M. Faloutsos, and S. Krishnamurthy, “Peernet: Pushing peer-2-peer down the stack.,” in IPTPS, 2003. [6] C. Perkins, “Ad hoc on demand distance vector routing,” 1997. [7] Charles Perkins and Pravin Bhagwat, “Highly dynamic destinationsequenced distance-vector routing (DSDV) for mobile computers,” in ACM SIGCOMM’94, 1994. [8] David B Johnson and David A Maltz, “Dynamic source routing in ad hoc wireless networks,” in Mobile Computing, vol. 353. Kluwer Academic Publishers, 1996. [9] Xiaoyan Hong, Kaixin Xu, and Mario Gerla, “Scalable routing protocols for mobile ad hoc networks,” IEEE NETWORK, vol. 16, no. 4, 2002. [10] Z. Haas, “A new routing protocol for the reconfigurable wireless networks,” 1997. [11] Guangyu Pei, Mario Gerla, and Tsu-Wei Chen, “Fisheye state routing: A routing scheme for ad hoc wireless networks,” in ICC (1), 2000, pp. 70– 74. [12] Paul F. Tsuchiya, “The landmark hierarchy : A new hierarchy for routing in very large networks,” in SIGCOMM. 1988, ACM. [13] Benjie Chen and Robert Morris, “L+: Scalable landmark routing and address lookup for multi-hop wireless networks,” 2002. [14] Aline C. Viana, Marcelo D. de Amorim, Serge Fdida, and Jos F. de Rezende, “Indirect routing using distributed location information,” ACM Mobile Networks Applications, Special Issue on Mobile and Pervasive Computing, 2003. [15] Jakob Eriksson, Michalis Faloutsos, and Srikanth Krishnamurthy, “Scalable ad hoc routing: The case for dynamic addressing,” in IEEE InfoCom, 2004.
Velammal College of Engineering and Technology, Madurai
Page 13
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Framework for Comparison of Association Rule Mining using Genetic Algorithm K.Indira, Research Scholar, Department of CSE, Pondicherry Engineering College, Pondicherry, India
S.Kanmani Professor & Head Department of IT Pondicherry Engineering College, Pondicherry, India
[email protected]
[email protected]
Abstract – A new framework for comparing the literature on Genetic Algorithm for Association Rule Mining is proposed in this paper. Genetic Algorithms have emerged as practical, robust optimization and search methods to generate accurate and reliable Association Rules. The main motivation for using GAs in the discovery of high-level prediction rules is that they perform a global search and cope better with attribute interaction than the greedy rule induction algorithms often used in data mining. The objective of the paper is to compare the performance of different methods based on the methodology, datasets used and results achieved. It is shown that the modification introduced in GAs increases the prediction accuracy and also reduces the error rate in mining effective association rules. The time required for mining is also reduced. Keywords - Data Mining, Genetic Algorithm, Association Rule Mining,
I.
INTRODUCTION
• Data selection: at this step, the data relevant to the analysis is decided on and retrieved from the data collection. • Data transformation: also known as data consolidation, it is a phase in which the selected data is transformed into forms appropriate for the mining procedure. • Data mining: it is the crucial step in which clever techniques are applied to extract patterns potentially useful. • Pattern evaluation: in this step, strictly interesting patterns representing knowledge are identified based on given measures. • Knowledge representation: is the final phase in which the discovered knowledge is visually represented to the user. This essential step uses visualization techniques to help users understand and interpret the data mining results. The architecture of Data mining system is depicted in figure.1
In today’s jargon enormous amount of data are stored in files, databases, and other repositories. Hence it becomes necessary, to develop powerful means for analysis and interpretation of such data and for the extraction of interesting knowledge to help in decision-making. Thus, there is a clear need for (semi-) automatic methods for extracting knowledge from data. This need has led to the emergence of a field called data mining and knowledge discovery. Data Mining, also popularly known as Knowledge Discovery in Databases (KDD), refers to the nontrivial extraction of implicit, previously unknown and potentially useful information from data in databases. The Knowledge Discovery in Databases process comprises of a few steps starting from raw data collections to formation of new knowledge. The iterative process consists of the following steps: • Data cleaning: also known as data cleansing, is a phase in which noise data and irrelevant data are removed from the collection. • Data integration: at this stage, multiple data sources, often heterogeneous, may be combined in a common source.
Velammal College of Engineering and Technology, Madurai
Figure 1. Architecture of typical Data mining System.
Page 14
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
This paper reviews the works published in the literature, where basic Genetic Algorithm is modified in some form to address Association Rule Mining. The rest of the paper is organized as follows. Section II briefly explains Association Analysis. Section III gives a preliminary overview of Genetic Algorithm for Rule Mining. Section IV Reviews the different approaches reported in the literature based on Genetic Algorithm for Mining Association Rules. Section V lists the inferences attained from the comparison. Section VI presents the concluding remarks and suggestions for further research. I.
new population. The flowchart of the Basic Genetic Algorithm is given in figure 2.
ASSOCIATION ANALYSIS
Association analysis is the discovery of what are commonly called association rules. It studies the frequency of items occurring together in transactional databases, and based on a threshold called support, identifies the frequent item sets. Another threshold, confidence, which is the conditional probability that an item appears in a transaction when another item appears, is used to pinpoint association rules. The discovered association rules are of the form: P Q [s,c], where P and Q are conjunctions of attribute value-pairs, and s (for support) is the probability that P and Q appear together in a transaction and c (for confidence) is the conditional probability that Q appears in a transaction when P is present. III. GENETIC ALGORITHM A Genetic Algorithm (GA) is a procedure used to find approximate solutions to search problems through the application of the principles of evolutionary biology. Genetic algorithms use biologically inspired techniques such as genetic inheritance, natural selection, mutation, and sexual reproduction (recombination, or crossover). Genetic algorithms are typically implemented using computer simulations in which an optimization problem is specified. For this problem, members of a space of candidate solutions, called individuals, are represented using abstract representations called chromosomes. The GA consists of an iterative process that evolves a working set of individuals called a population toward an objective function, or fitness function. Traditionally, solutions are represented using fixed length strings, especially binary strings, but alternative encodings have been developed. The evolutionary process of a GA is a highly simplified and stylized simulation of the biological version. It starts from a population of individuals randomly generated according to some probability distribution, usually uniform and updates this population in steps called generations. Each generation, multiple individuals are randomly selected from the current population based upon some application of fitness, bred using crossover, and modified through mutation to form a
Velammal College of Engineering and Technology, Madurai
Figure 2. Basic Genetic Algorithm.
A. [Start] Generate random population of n chromosomes (suitable solutions for the problem) B. [Fitness] Evaluate the fitness f(x) of each chromosome x in the population C. [New population] Create a new population by repeating the following steps until the new population is complete i. [Selection] Select two parent chromosomes from a population according to their fitness (the better fitness, the bigger chance to be selected) ii. [Crossover] With a crossover probability cross over the parents to form a new offspring (children). If no crossover was performed, offspring is an exact copy of parents. iii. [Mutation] With a mutation probability mutate new offspring at each locus (position in chromosome). iv. [Accepting] Place new offspring in a new population D. [Replace] Use new generated population for a further run of algorithm E. [Test] If the end condition is satisfied, stop, and return the best solution in current population F. [Loop] Go to step 2
Page 15
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IV. ANALYSIS ON GENETIC ALGORITHM FOR MINING ASSOCIATION RULES Among the Genetic algorithms designed for the purpose of Association rule mining is discussed based on the following criteria 1. Genetic Operations 2. Encoding I. Initial Population II. Crossover III. Mutation IV. Fitness Threshold 3. Methodology. 4. Application areas. 5. Evaluation Parameters The various methodologies are listed in Table A1. given in Annexure. 1. Genetic Operations The basic steps in the traditional Genetic algorithm implementations are discussed in the previous section. Modifications are carried out in the traditional GA to increase the prediction accuracy thereby reducing error rate in mining association rules. The variations have been carried out in various steps of GA. 2. Encoding : Encoding is the process of representing the entities in datasets for mining. Rules or chromosomes can be represented either with fixed length data [2..18] or by varying length chromosomes[1], Fuzzy rules are implemented for encoding data in [10], In [4] and [6] coding is done using natural numbers, In [17] Gene expressions are used for representation of chromosomes and encoding is carried out using arrays in [7] . Initial Population: The initial population could be generated by random selection, seeded by users [1], single rule set generation [5] and Fuzzy Rules [10]. Crossover: The Crossover operator which produces new offspring and hence new population plays a vital role in enhancing the efficiency of the algorithms. The changes are carried out as discussed in Table A1. Saggar. M et. Al [2] describes whether crossover is to be performed or not and if required the locus point where the crossover begins is of prime importance. Crossover on same attributes of both offspring if present or random attributes in absence of similar attributes is carried out in [11]. In [12] setting the crossover rate dynamically for each generations are presented. The concept of Symbiotic combination, where instead of crossover the combination of chromosomes to generate a new chromosome based on Symbiogenesis in Ramin Halavathy et. Al [5] has proved to increase the speed of the rule generation system.
Velammal College of Engineering and Technology, Madurai
Mutation: Mutation is the process where attributes of selected entities are changed to from a new offspring. The Mutation Operator is altered based on the application domain into macro mutation in [1], changing locus points of mutation in [2]. The weight factor is taken into consideration for locus point of mutation in [5] so as to generate a better offspring, Dynamic mutation where the mutation point is decided on the particular entity and generation selected enhances the diversity of colony is introduced in [12], mutation 1 & mutation 2 where mutation is performed twice to generate offspring is performed in [16]. Adaptive mutation where the mutation rate differs for each generation is found to produce better offspring’s in [17]. Fitness Threshold: The passing of chromosomes from a population to new population for the succeeding generation depends on the fitness criteria. Changes to the fitness functions or threshold values alter the population and hence the effective fitness values lead to the efficiency of the rules generated. The negation of the attributes are taken into consideration while generating rules by including criteria like True Positive, True Negative, False Positive and False Negative[2]. By these criteria rules with negated conditions of attributes are also generated. By varying the fitness values dynamically in each generation, the speed of the system can be improved [4]. Factors like strength of implication of rules when considered while calculating fitness threshold proves to generate more interesting rules [6]. The Sustainability index, creditable index and inclusive index when considered for fitness threshold results in better predictive accuracy [7]. When the real values of confidence and support derived and applied in threshold rate found to generate faster than traditional methods [8]. The predictability and comprehensibility factors of rules tends to provide better classification performance[11]. 3. Methodology Rather than altering the operations in basic GA algorithm the changes made in the methodology has also proved to increase the performance. In [5] the crossover operation is replaced by symbiotic recombination techniques. Wenxiang Dou et. Al [8] describes the generation of rules and displayed to the user. The user decides the interesting rules thereby seeding the population by the user. Here instead of searching the whole database for support the system searches K- itemset alone and hence is faster than other methodologies. The real values of support and confidence are taken into account while generating threshold values. If user is not satisfied then more rules with next level of support could be generated. The response time of the system is found to be increased dramatically. Antonio Peregrin et. Al [9] states the distributed genetic algorithm where the Elit data pool is the central node. Each node connected are considered as Intercommunicating
Page 16
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Subpopulations. The Data Learning Flow (DLF) Copies the Training Example with Low Fitness to neighborhood and each node is assigned with different partitions of learning data. In [18] the concept of Dynamic Immune Evolution and Biometric Mechanism is introduced. The immune recognition, immune memory and immune regulation is applied to GA and thereby making the system faster and discovering interesting rules though support might be low.
• Predictive Accuracy of more than 95 percentage was seen in 55 % of datasets. • Predictive Accuracy between 80 - 94 percentage was seen in 15 % of datasets. • Predictive Accuracy between 75 – 79 percentage was seen in 10 % of datasets. • Predictive Accuracy of less than 75 percentage was seen in 20 % of datasets.
4. Application Areas
Predictive accuracy of 100 % is achieved with methodology [1] and [14]. The error rate is seen maximum in Pima dataset under UCI repository by SGERD method [10].
From the dataset in Table A1 it is derived that the Genetic Algorithm is flexible in nature and could be applied to varying data sets. Different areas of applications as Biology, biometrics, Education, Manufacturing Information System, Application Protocol interface records from Computers for Intrusion Detection, Software Engineering, Virus information from Computer data, Image data base, Finance information, Students Information etc. are taken up for study. It is seen that by altering representations and operators the Genetic algorithm could be applied for any fields without compromising the efficiency. 5. Evaluation Parameters To achieve the best performance from any system setting of the evaluation parameters are vital. For mining efficient association rules using the evolutionary Genetic Algorithm the parameters that affects the functioning of the system are representation of chromosomes and their length, population Size, Mutation probability, Crossover probability and the Fitness threshold values. The use of support and confidence factor in deciding the threshold values also increases the performance of the system. It is seen form the Table A1 that parameters don’t have predefined boundaries. The values of the parameters are based upon the methodology applied and the dataset. It is found that the mutation rate differs from 0.005 to 0.6. Similarly the Crossover probability ranges from 0.1 to 0.9. The fitness function is found to be the most crucial parameter. The optimum method for the fitness threshold evaluation has to be arrived in order to achieve the highest performance. The carrying over of offspring from one generation to other is based on fitness function the population. Different factors of the system are takem into consideration while generating fitness function. Efficient the fitness function, efficient and accurate are the association rule that are generated by the system V.
INFERENCES
The study of the different system using Genetic Algorithm for Association rule mining was carried out. The predictive accuracy based on the methods under comparison is noted that
Velammal College of Engineering and Technology, Madurai
Based on the time for mining association rules as low as 5 and 9.81 seconds were achieved in SGERD [10] and Quick response data mining model [8] respectively and maximum of 7012 seconds is taken up for mining KDDCUP99 dataset using Genetic algorithm with symbiogenesis [5]. The tools used for mining are GA Toolbox from Mat lab, Java, C++ and Weka for training the datasets. From the Study it could be inferred that • The system is flexible and could be applied to varying datasets and applications. • The performance of the system is improved when compared to other existing methods like Apriori, C4.5, REGAL, Data-miner etc. • The Genetic operators are the key in achieving increased performance. • The speed of the system increases rapidly when compared with other methods. • The optimum values for the mutation, crossover and threshold decide on the performance of the system. • Better and precise rules are generated. • Classification error rates are reduced in comparison.. • Negated attributes could be derived in rules. • Critical rules could be generated even if support is low. VI. CONCLUSION The framework for comparing the effectiveness of Genetic Algorithm in mining Association rules is taken up for study. Various systems applying GA for mining Association rules are considered and the tabulation was done with the data chosen. From the analysis it could be derived that the classification performance of the system was found to be robust. Genetic Algorithm based implementation Outperforms Conventional Methods. The speedup achieved is remarkable. For future work the combination of one or two methodologies could be carried out for basic GA operators. The given system for a particular domain can be further modified for other domains.
Page 17
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Table A1 : Genetic Algorithms For Association Rule Mining: A Comparative Study Data Set Ref .No
Methodology Steps In Genetic Algorithm Represen -tation
[1]
Varying Length
Selectio n
Seeded Populati on
Mutation
Macro Mutations
Cross Over
As GA
As GA
[2]
[3]
[4]
[5]
[6]
Binary Coding
As GA
Using Natural Numbers
As GA
Using Natural Numbers
Roulette Wheel Selection
As GA
Roulette Wheel Selection
As GA
Optimum Value
Whether Required Or Not And If Point Of Crossove r
As GA
Optimum Value
Changes In Weight Symbiotic Generate Of Combinati Single Membersh on Rule Sets ip Operator Function
As GA
As GA
As GA
TP,TN,FP, FN Used
As GA
As GA
Grilled Mushrooms in Agaricus and Lepiota Family.
Synthetic Database for the Selection of Electives for a Course
Data Set from MIS Synthetic Image Database KDDCUP99 Classes 2 CRX Classes 2
As GA
Individual Evaluation Using Strength Of Implication
[7]
Array Represent a -Tion
As GA
As GA
As GA
Based On Sustaining, Creditable And Inclusive Index
[8]
As GA
Done By
As GA
As GA
Based On Real
Evaluated By / Parameters Size / Pc/ Suppor Confi t dence
Fitness
Vehicle Silhouette Dataset Whether Required Or Not And If Point Of Mutation
Sample Size
Fitness
Gen er atio ns
8128 Of 23 Species 22 Attributes 846 Records 18 Attributes
*
100
*
*
0.1
0.005
5%
User define d
*
5% *
*
.25% To 2%
*
4 To 14
*
Completen ess consider -red
User defined
*
*
-
*
*
Features 15 Size 690 Features 4 Size 150
Vote Classes 2
Features 16 Size 435
Wine Classes 3
Features 13 Size 178
10000 0.4
10
0.85
0.4
0.2
0.8
0.6
6
0.6
8
50
*
Six Datasets from Irvine Repository
*
S.I 1.0 C I 1.0 I.I 1.0
10
* Single Table
Velammal College of Engineering and Technology, Madurai
Accuracy On Training Dataset Between 95 To 100. On Test Data Between 62 And 71 Rules With Negation Of Attributes As Well As General Rules Generated Based On Time And Fetching Knowledge GA Is Faster Than Apriori. Runs 2 To 5 Times Faster Than Apriori
Features 41 Size 494021
IRIS Classes 3
Car Test Results Dataset
Pm
Results
40
10%
50%
50
SEA Has Better Or Similar Results When Compared With GA SEA Much Faster Than GA
During Generation Between 1 To 200 Interesting Rules For Whatever Be The Threshold. Predictive Accuracy Better Than CN2 And AntMiner Methods Response Got
Page 18
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Users From Rules Generate d •
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Support And Confidence
Produced Randomly with 100 Transactions
Multiple Intercommunicating Subpopulations Distributed Data And DLF • Central Elite Pool UCI : Nursery • The Data Learning Flow (DLF) Copies The Training Example With Low Fitness To Neighborhood • Each Node Is Assigned With Different Partitions Of Learning Data Fuzzy Rules
As GA
As GA
As GA
As GA
Gene String Struct ure •
0 *
*
0.01
0.9
•
As GA
Elitist As GA Recombinat ion Method
As GA
Attributes
As GA
As GA
As GA
As GA
As GA
As GA
As GA
As GA
Dynamic
Adaptive
As GA
As GA
Mutation1 & Mutation 2
Adaptive
As GA
Done On Same Attributes If Present Or Random
Dynamic
As GA
As GA
As GA
As GA
Predictive Accuracy
11 Data Sets from Irvine Machine Learning Repository Adult
Nursery Datasets from UCI
Based On Last Generation
Finance Service Data of Certain City
Individual Based
Database Of Student Achievement in Schools in Recent Years
Modified To Decide KDD CUP99 Whether A Dataset Chromosome Is Right Or Not
Feature Selection Is Applied
CM1, KC1, KC2, PC1 From UCI Repository Vehicle Dataset And
As GA
Adaptive
Based On Distance Between Rules Lympography Dataset From UCI ML Measure Of Overall Performance
Dynamic Immune Evolution And Biometric
Real Case Data
Computers
Velammal College of Engineering and Technology, Madurai
12960 Instances
*
*
optimu m 48842 Instances 15attributes
40
Optimu m
0.3
200
Optimu m
0.6
0.9
*
*
*
0.01
0.7
*
*
148 Records 18 Attributes 4 Classes
Varie s
240
Length Of Chromosome 41 • Generations 100 • Generation Gap 0.9
Varie s
Varie s
0.1
0.01 0.01
846 Records, 18 Attributes 4 Classes
*
*
50
*
•
22
*
0.05 *
12960 With 9 Attributes
2050 Groups
optimu m
50 0
600
0.6
0.8
0.005
*
Varies *
*
*
In Ten Seconds Whereas For Apriori It Is More Than 3000 Seconds Faster And Better Behavior Number Of Rules Generated Is Between 60%80% Smaller Classification Error Rates Are Low Outperforms C4.5 Better Classification Performance Produces Partial Association Rules After 252 Generations Whereas It Is 850 In traditional GA The Algorithm Based On 0.1 Support And 0.7 Confidence Is Close To Actual Situation Rules Generated Are Useful In Detecting Intrusion Generated Rules That Provide Better Estimation And Explanation Of Defective Modules
10 0
GRA Outperforms Conventional Methods
50
Performance And Effectiveness Of Proposed Model Is Close With Real World Analysis Faster &
Page 19
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
•
Mechanism Is Introduced Immune Recognition, Immune Memory And Immune Regulation Is Applied To GA
Daily Records Of API
0.26 *
*
*
*
Discovers New Critical Rules Though Support Not High
Note : *- not defined in literature REFERENCES [1]. Cattral, R., Oppacher, F., Deugo, D.,”Rule Acquisition with a Genetic Algorithm”, Proceedings of the 1999 Congress on Evolutionary Computation,. CEC 99, 1999. [2]. Saggar, M., Agrawal, A.K., Lad, A., “Optimization of Association Rule Mining”, IEEE International Conference on Systems, Man and Cybernetics, Vol. 4, Page(s): 3725 – 3729, 2004 [3]. Cunrong li, Mingzhong Yang, “Association Rules Data mining in Manufacturing”, 3rd International Conference on Computational Electromagnetics and Its Applications, Page(s): 153 – 156, 2004. [4]. Shangping Dai, Li Gao, Qiang Zhu, Changwu Zhu, “A Novel Genetic Algorithm Based on Image Databases for Mining Association Rules”, 6th IEEE/ACIS International Conference on Computer and Information Science, Page(s): 977 – 980, 2007 [5]. Halavati, R., Shouraki, S.B., Esfandiar, P., Lotfi, S., “Rule Based Classifier Generation Using Symbiotic Evolutionary Algorithm” , 19th IEEE International Conference on Tools with Artificial Intelligence, Volume: 1, Page(s): 458 – 464, 2007. [6]. Zhou Jun, Li Shu-you, Mei Hong-yan, Liu Haixia, “A Method for Finding Implicating Rules Based on the Genetic Algorithm”, Third International Conference on Natural Computation, Volume: 3, Page(s): 400 – 405, 2007. [7]. Hua Tang, Jun Lu, “A Hybrid Algorithm Combined Genetic Algorithm with Information Entropy for Data Mining”, 2nd IEEE Conference on Industrial Electronics and Applications, Page(s): 753 – 757, 2007. [8]. Wenxiang Dou, Jinglu Hu, Hirasawa, K., Gengfeng Wu, “Quick Response Data Mining Model using Genetic Algorithm”, SICE Annual Conference, Page(s): 1214 – 1219, 2008 [9]. Peregrin, A., Rodriguez, M.A., “Efficient Distributed Genetic Algorithm for Rule Extraction”,. Eighth International Conference on Hybrid Intelligent Systems, HIS '08. Page(s): 531 – 536, 2008 [10]. Mansoori, E.G., Zolghadri, M.J., Katebi, S.D., “SGERD: A Steady-State Genetic Algorithm for Extracting Fuzzy Classification Rules From Data”, IEEE Transactions on Fuzzy Systems, Volume: 16 , Issue: 4 , Page(s): 1061 – 1071, 2008. [11]. Xian-Jun Shi, Hong Lei, “A Genetic AlgorithmBased Approach for Classification Rule Discovery”,
Velammal College of Engineering and Technology, Madurai
International Conference on Information Management, Innovation Management and Industrial Engineering, ICIII '08, Volume: 1 , Page(s): 175 – 178, 2008. [12]. Xiaoyuan Zhu, Yongquan Yu, Xueyan Guo, “Genetic Algorithm Based on Evolution Strategy and the Application in Data Mining”, First International Workshop on Education Technology and Computer Science, ETCS '09, Volume: 1 , Page(s): 848 – 852, 2009 [13]. Hong Guo, Ya Zhou, “An Algorithm for Mining Association Rules Based on Improved Genetic Algorithm and its Application”, 3rd International Conference on Genetic and Evolutionary Computing, WGEC '09, Page(s): 117 – 120, 2009 [14]. Yong Wang, Dawu Gu, Xiuxia Tian, Jing Li, “Genetic Algorithm Rule Definition for Denial of Services Network Intrusion Detection”, International Conference on Computational Intelligence and Natural Computing, CINC '09, Volume: 1 , Page(s): 99 – 102, 2009 [15]. Rodriguez, D., Riquelme, J.C., Ruiz, R., AguilarRuiz, J.S., “Searching for Rules to find Defective Modules in Unbalanced Data Sets”, 1st International Symposium on Search Based Software Engineering, Page(s): 89 – 92, 2009 [16]. Gonzales, E., Mabu, S., Taboada, K., Shimada, K., Hirasawa, K., “Mining Multi-class Datasets using Genetic Relation Algorithm for Rule Reduction”, IEEE Congress on Evolutionary Computation, CEC '09, Page(s): 3249 – 3255, 2009 [17]. Haiying Ma, Xin Li, “Application of Data Mining in Preventing Credit Card Fraud”, International Conference on Management and Service Science, MASS '09, Page(s): 1 – 6, 2009 [18]. Genxiang Zhang, Haishan Chen, “Immune Optimization Based Genetic Algorithm for Incremental Association Rules Mining”, International Conference on Artificial Intelligence and Computational Intelligence, AICI '09, Volume: 4, Page(s): 341 – 345, 2009
Page 20
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Content Management through Electronic Document Management System T.Vengattaraman#1, A.Ramalingam*2, P.Dhavachelvan#3 #
Department of Computer Science, Pondicherry University
Puducherry, India. 1
[email protected] 3 [email protected]
*
Department of Master in Computer Applications, Sri Manakula Vinayagar Engineering College
Puducherry, India. 2
[email protected]
Abstract— This paper presents “Electronic Document Management System” which is to develop a web based application which can be used to manage the contents hosted on a web site. Electronic document management system provides services in the areas of providing information to the different areas such as Development, Business, Management, Technology etc. The main functionality of this system is Content Creation/Hosting, which is the administration section of the system in which the content is created by a group of professional possessing different specialization. This section is divided into four stages such as authoring, editing, approving and deploying. After going through these four phases of content management the content is available to the users. Each content is placed in one of the three different categories. And content View is the one where the content to be viewed is divided into three halves as developer Content, manager Content and protected Content. The Protected content is login protected where as the other two are open to all. The developer content is used by the all for content/information modification that helps in the dynamic display of different information for providing services to the Business, Management, etc whereas the manager content is also used by all but the contents are used for customization purpose. Keywords— Electronic document management system, Content management, Web publishing.
I. INTRODUCTION To develop a web based interface for creating new content for the site or manage existing content. The system should provide help in managing different personnel capable of working in different areas of content creation. To make the best possible content available as an end result. Divide the complex task of content creating into no of specialists. To allow the floating of content with the system before being hosted on site. Different levels of approval to make the content very precise as per the requirements. To speed up the processing with in the
Velammal College of Engineering and Technology, Madurai
Content Creation. The current system although semi atomized required manual processing of approval and editing before being approved for the deployment. Transfer of information content between different sections of the current system is in the form of documents. Selection of a person for a task is done by manually approaching the person and confirming the availability of the person. Due to mismanagement the work is delayed to later date than the due date. Although the current system is confined to only one type of a website management. It can be generalized a General Electronic Document Management system with which any type of website can be managed. The current system is interactive with the database provides efforts can be made so that the system can adopt the available database features of a new site to make is as a part of content management. II. LITERATURE SURVEY As organizations increasingly move towards a paperless environment due to an increasing need in storing more and more documents, an Electronic Document Management System (EDMS) must be in place to cope not only with the increasing volume of these documents but also to make sure that these documents are searchable for use in the future. EDMS is proven to be an appropriate tool for handling necessary documents of organizations around the world. To aid the researchers pursue this study, they have gathered some studies related to their proposed system. Electronic Document Management System (EDMS) seemed to be the answer in organizing business information for easy access and quick retrieval and archiving important electronic records in a secure and stable environment. EDMS might just provide the tool to effectively manage electronic documents and records throughout its life cycle. After implementing the EDMS with Accounts Payable application and Electronic Reports Management technology to AEC, access and retrieval of documents is more efficient and there have been no reported “lost” or “misfiled” documents. Employees in
Page 21
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
remote locations can quickly retrieve and view records not requiring much time and effort, and more paper. Since EDMS provides a secure and stable environment, records cannot be altered, and any document access or action can be tracked through audit history. Vital records stored in EDMS are secured and backed up daily[1-3]. To implement an automated document analyzer that will provide the file organization, searching, retrieval and segmentation of files is the main objective of the thesis by Revano [7]. To achieve his goal, he used three data mining algorithms such as the Aho-Corasick algorithm for exact set matching, the Segmentation algorithm which generates a series of cluster based on the identified relationships in a dataset, and the Text mining algorithm or automatic text categorization. Based from findings of the study, the researchers had proven that there was a significant difference between the proposed system and the existing system being implemented in CITE department. Based from the gathered related studies, the researchers have come up with the idea of developing an electronic document management system. Organizing documents, paper and electronic, is now a concern in any organization. In relation with the studies mentioned, the researchers agreed that exhibiting the basic components of electronic document management systems such as capturing, storage and archiving, indexing and retrieval, distribution, and security will produce a good EDMS. III. SYSTEM OVERVIEW The complete system can be divided into six halves on basis of access levels. The system architecture is given in the figure 1. A. Account management: Using this part of an application the administrator can view the list of users and their area of specialization. The administrator can create a new users, modify existing user. An administrator provides permission to the newly created user by placing the new user into set of roles such as an author, approver, and editor or deploy. This part of the application is only accessible to the administrator. Login Area
Processing Area Database Approving
User Name
Account Login
Deploy
Database
Editing
Password
Authorizing
B. Utilities: Utilities section of the application is used to shut down the site for the normal person to browse as well as to up the site back for its use. C. Authoring: An administrator or a person with the author privileges can access this part of the application. This part of the application includes creating new content in the form of stories which is normally done by the developers or content writers. The newly created content may include no of notes which will guide the editor at the time of editing the content. The newly created content then can be posted to editor for editing. D. Editor: An editor receives the content posted by the author. An editor can view the content and later post the content to a new revision or to an existing revision. If content is found unsuitable to the cause the content is returned back to the author. This part of the application can be explored only by an administrator or the users who possess an editor privilege. The editor can withdraw the content from being hosted if found unfit for hosting. E. Approver: An approver is a person who will approve the contents to be hosted on the site. An approver can approve the content to the deploy section or Discontinue the content usage or return the content back to the editor for revision. The returned content should accompany with a message to the editor regarding the revision in the content. This part of the application can be accessed by the administrator of the person who possesses an Approver privilege. F. Deploy: This area of the application includes the deployment part of an application. A deploy person can view the content before deploying it. The person can also return the content if found unfit to be hosted on the site. The returned content is sent back to the approver. The deployment of the content includes the content to be placed in specific area of the hosting environment. The hosting environment is divided into three categories. The Deploy content, the manager content, the protected content. These categories are subdivided into no of sections. G. Administrator: An administrator has all the privileges that of the guest as well as the normal registered user. Along with these common features an administrator has the administrator related features such as creating new users and granting roles to those newly created users. The roles granted by the administrator cannot be changes by the user. An administrator can create new user as a guest or as a user
Fig 1 General Architecture Diagram
Velammal College of Engineering and Technology, Madurai
Page 22
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
or an administrator. The access levels are as per the grants done by the administrator. An administrator can also be part of a team and could lead a project team this is possible only if administrator when building a team includes himself in the team section. If included as a manager he is not a part of the team but supervisor of the team. The register option on the homepage of the application is provided only to register a new user as a guest. The information of the entire system will be maintained at a centralized data base. Notifications between sections is provides in terms of content list notification in the users area. Provide Interactive interface through which a user can interact with different areas of application easily. Disapproved content is returned back to the lower for redesign. Approved content is removed from the user list and made as part of the user’s list to which the content is being notified. Deploy the application on a single system and make is available on all the systems within the network, thereby reducing the maintenance cost of software. The system screen shots are shown in the figure 2 and figure 3.
Fig. 2 The Contents are hosted in this area
IV. CONCLUSION This system has been appreciated by all the users and is easy to use, since it uses the GUI provided in the user dialog. User friendly screens are provided. The usage of software increases the efficiency, decreases the effort. It has been efficiently employed as a Content management mechanism. It has been thoroughly tested and implemented. The application is capable of managing only the current site. It can be converted into general Electronic Document Management system by providing a tight integration of the existing database with the one the site holds. The current database should be redesigned so that it can adapt to the changes with respect to the site. The current Electronic Document Management system manages only the articles and there exists no provision for the new categories to be added and placed into the content section. This can be brought into the current on user request. The application can be converted into a Mobile based using ASP.net with which the deployment of application will be done only on enterprises server and is accessible to all other departments of the organization. The current application is confined to only one enterprise. REFERENCES [1] Adam, “Fundamentals of EDRMS. Implementing Electronic Document and Record Management Systems”, CRC Press. Madison Ave. New York, 2007 [2] J. Feldman and E. Freuder, “Integrating Business Rules and Constraint Programming Technologies for EDM”, Business Rules Forum 2008. [3] Georgia Archives, “Electronic Document Management System Technologies” ,November 2008. [4] P. Immanuel, “Basic Components of Electronic Document Management System – Essential Characteristics of an Effective Document Management System”, November 2008 [5] Laserfiche, “A guide to the benefits, technology, and implementation of electronic document management solutions”, Archival and scanning services, Nov 2008. [6] D.P. Quiambao, “Document Management System at the Property Management Division of Subic Bay Metropolitan Authority”, 2004 [7] T.J. Revano, “Development of Automated Document Analyzer using Text Mining and Aho-Corasick Algorithm” . TIP Manila: Unpublished Thesis, 2008 [8] Sire Technologies, “A Guide to Understanding Electronic Document Management Systems: The Future File Cabinets”, 2006.
Fig. 3 To deploy the content.
Velammal College of Engineering and Technology, Madurai
Page 23
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Intelligent Agent based Data Cleaning to improve the Accuracy of WiFi Positioning System Using Geographical Information System (GIS) T.JOSHVA DEVADAS Lecturer, Department of MCA, The American College, Madurai [email protected] Abstract — Wifi positioning system uses GIS to achieve higher accuracy by means of comparing the error distance. The objective of this paper is to minimize the distance error generated during the process of positioning. Wifi positioning system needs to have a proper placement of those access points which are free from error distance. Data cleaning is introduced in this concept to clean the errors that are generated from the experimental results. An agent is a Intelligent Processing system that is situated in the environment and perform autonomous action in order to meet the design objectives. Agents are introduced in this system uses its intelligence during the data cleaning task by using its dynamic and flexible characteristics. This paper aims at describing the functionalities of such a data cleaning agent so as to improve the performance of Wifi positioning system. Keywords— Wireless positioning, WiFi, GIS, Constraint optimization, Data cleaning, Agents
1.
Introduction
WiFi stands for wireless fidelity which is one of the Institute of Electrical and Electronics Engineers or IEEE standards for WLAN. Wifi is truly a freedom to make a connection to Internet without the old-fashioned network cable. WiFi or 802.11b is one of 802.11 specifications family and it is also the first dominating standard in the market. Nowadays new wireless networking devices are adopting the new standard of IEEE 802.11g, which offers higher data speed, namely driving from 11 Mbps to 54 Mbps, while sharing common operating frequency of 2.4 GHz.
WiFi technology is evolving technically and practically in past four years leading WLAN to be a common sight at universities, airports, coffee shops, offices and organizations. That service point or access point is often referred to as “wireless hotspot” or “hotspot” in short. Incentives for developing and standardizing WLAN are definitely mobility and flexibility. This phenomenon stimulates the creation of supportive features for new-age laptops. Data preprocessing is a procedural activity which is to be done at the first step of data mining process. Data preprocessing is a process of requesting the data to be
Velammal College of Engineering and Technology, Madurai
cleaned. Data mining process need to analyze the possible occurrence of the noisy data[14]. The noise will occur due to faults generated during the data transmission. To resolve this we need to develop a system which can produce an error free data by identifying or removing the outliers identified in the system. In this paper we consider only those error positions which are to be eliminated during the wifi positioning process. Agents are intelligent autonomous computer systems, which are capable of functioning autonomously and react to the environment[1]. The intelligent agent doesn’t just react to the changes to its environment but itself takes the initiative under specific circumstance. This characteristics request the agent should have a well defined goal. Agents incorporated in the data cleaning task have a specific goal of removing the errors by identifying the outliers. In this paper section 2 deals about related work and background details of wifi positioning system. Section 3 describes the positioning computation by describing radio propagation model, free-space model and wall-attenuation method. Section 4 explains data cleaning process along with agent participation, section 5 describes performance evaluation.
2. Related works and Background Now days all the WLAN is the emerging research area that varies from real-world application to pure theoretical aspect. Major issue in wireless security is the wireless positioning application. WEP and WPA are encryption algorithms that are to be realized by the realworld globally. Hence the development for wireless security turns to be the positioning technique. Many researchers use the positioning technique to improve the accuracy of their need with various platform and architecture. The objective of this paper is to improve the accuracy by the means of positioning the access points in two areas with the aid of GIS. The experiment considers the indoor environment with two access points.
Page 24
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
There are two types of computation to perform this positioning process. One is location fingerprinting model and the other one is the propagation based computation model[8]. Location fingerprinting requires site survey and to collect signal strength samples at every single equally divided area of that building to build up a signal strength database. Propagation based computation is used to get the strength of signal but it don’t have the facility to store the database. These are the some of the computation to do the mapping process. The following is the test site for the system where we need to position the access points
the corridor. Those fifteen markers are points of interest in our experiment to be used for learning and testing purposes. They are 1.25 meters apart from one another.
FIG. 2.PROCESS OF NETSTUMBLER Multipath Fading Minimization:
Fig .1. Q5 floor plan with markers and access points locations In the test site we need to place the access points in two different areas and the corridor is selected for our test process and get the signal from two access points namely NAL-Q51 and DSP-Q52. Triangulation is a basic geometric concept for finding crossing point on the basis of distance or angle [6][7]. Therefore it is radiating circularly in horizontal direction and so the access point is in the pattern of circle. The existing system needs to be improved in order to decrease the error distance between the two access points and therefore it needs to perform the operation to obtain the calculated signal strength and error distance[2]. The next step in this process is to place the access point in an error free area. The existing system makes use of the Netstumbler software to detect the signal strength[10]. Signals are received from the two access points with the help of PDA or laptop and that are detected using the Netstumbler software. Netstumbler imports the signal strengths and are mapped into excel sheet to calculate the signal average. MATLAB is used to find the erroneous data and later the positioning is done by using the floor position and calculated positions. Predetermined location of the access points named DSPQ52 and NAL-Q51 is included in Fig 1 with markers along
Velammal College of Engineering and Technology, Madurai
From all the four directions, the signal strength is collected to minimize the multipath fading. The Netstumbler software is used to detect the signal strength. Signal strength samples directly reflect the multipath effect in terms of direction sensitivity and angle of reception. In principle, we should achieve higher resolution of sampling and higher accuracy. Experiment Methodology: Equipments required for capturing signal strength, raw data processing and position determination are listed in categories of Hardware and Software accordingly. i) Hardware Access Point or Hotspot is functioning as a signal transmitter. Wireless Network Interface Card (Wireless NIC) or WiFi card is considered as wireless LAN signal receiver in this application. Wireless NIC is available in various interface types and it can operate on both laptop or desktop computer and Personal Digital Assistant (PDA) correspondingly. Latest laptop computer and PDA are equipped with internal Wireless NIC. Desktop computer is not suitable due to obvious reason of mobility from point to point around the floor. ii) Software NetStumbler is a freeware program and it is one of the most widely used wireless network auditing tools allowing the detection of 802.11b, 802.11a, and 802.11g networks. This software offers two versions for both Windows and WindowsCE platform, which are Netstumbler 0.4.0 and
Page 25
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Ministumbler 0.4.0 correspondingly. Signal strength samples are saved into NS1 file and it can be further be exported into Microsoft EXCEL for calculation, but Ministumbler does not have that feature. EXCEL is used to calculate the signal strength average for each particular point on the floor plan. MATLAB Characterizes wireless propagation model and minimize the error distance.
transmission model is described as PR = PTGTGR { λ / 4пd }2 where PR and PT are receiving and transmitting power in unit of watts. GT and GR are receiving and transmitting antenna gain. λ is a wavelength and d is a distance between transmitter and receiver.
3. Position Computation
Our experiment considers this model because it is considered to be more suitable for indoor environment. This model represents a common and reliable in building characteristics. Attenuation due to an intervening wall or partition is described as PR[dBm] = Po[dBm]- 10 log{d/d0}n –WAF where n is a path loss exponent and WAF is a Wall Attenuation Factor. Building environment may have the different electrical and physical characteristics which have obstruction that varies from the place to place. These values can be obtained by experiments.
Conceptual Overview Our experiment considers first floor plan found in Fig 1. To locate the access point’s position x-y coordinates is chosen. Distance error is calculated with respect to the exact location on the floor plan in hand. Signal strength should be collected from the NAL-Q51 and DSP-Q52 will be interpreted to any of the two models namely free-space model and wall attenuation model.
Wall Attenuation Factor Model:
Learning Process Position determination is based on free-space model and it does not require learning process as universal relationship is implied every environment considered. WAF model is used to achieve more accurate result and to represent more realistic indoor environment[9]. Before positioning service WAF and n have been computed from experiment. Once the measurement of signal strength at marking points is done, linear regression is applied to those data sets resulting in WAF and n parameters. Fig.3. Position determination overview Radio Propagation models Basically the radio channel associates with reflection, refraction, diffraction and scattering of radio waves that influence propagating paths. Transmitted signal from direct and indirect propagation path are combined either constructively or destructively causing variation of received signal strength at the receiving station. The situation is even more severe for indoors communication. The building may have different architectures, construction material which results in receiving challenging and unpredictable signal. Free-space Model This model is used for the worst case consideration. This model can be implemented with GIS but it unveils a trend for further improvement. This model is not suitable to implement because this may disturb the other signals. Also, this model is not appropriate for the indoor environment to do the positioning system. Free-space model or Friis
Velammal College of Engineering and Technology, Madurai
Fig. 4. Line of sight for NAL access point Line joining between the two access points to every single point of interest on the floor plan can at least differentiate amount of signal attenuation due to variety of obstructing objects. Longer the range from a transmitter, the lower signal strength from a receiver can perceive. Shorter range from transmitter can yields lower signal strength only in the case of presence of a large barrier between transmitter and receiver. Regional division is essential for the case of
Page 26
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
NAL-Q51 access point as received signal strength does not undergo common obstruction as the case of DSP-Q52 access point. Essentially, Geographical information is now used to separate region under consideration. System Simulation Signal strength collected is translated into transmitterreceiver distance with the aid of two radio propagation models. Transmitter-receiver distance is the separation distance of access point from user and the distance are visualized as a radius of antenna radiation circle with center at that access point of signal reception. User can be at any position on that antenna radiation circle. Signal strength received from DSP-Q52 is translated into radius of antenna radiation circle with the center at the access point itself. 4. Data cleaning process Data cleaning is a process of identifying and removing the erroneous data[4]. Since the data is received with poor quality, it is necessary to clean the data before using it for processing. Poor quality of data is obtained due to some noise. The noisy data may have erroneous data, missing values, duplicate/ redundant data, and useless information. The error occur in the collected data are more common. One way of addressing this problem is to painstakingly check through the data with the help of data cleaning methods. Data Cleaning Methods Data cleaning is done by using either decision trees, or filling the missing values, or removal of noisy data and eliminate the redundancies. Decisions trees induced from training data can often be simplified, without loss of accuracy, by discarding misclassified instances from the training set. Filling of the missing values is done by replacing with a global constant, such as attribute mean value, string constant, and numeric constants. A noise is nothing but a random error which can be removed either by binning method or clustering method or with the help of human inspection [9]. We consider the clustering method is more suitable for cleaning the data of a wifi positioning system, because the formation of clusters clearly identifies the outliers. Identification of the outlier will eliminate irrelevant information by forming various clusters. Need and strength of cleaning
Erroneous data is assumed to be the access points coordinates which causes intersection with the signal strength. Since the data collected may have many such erroneous data items and the errors are removed by a data cleaning process or method. Agents Agents are intelligent processing elements which can work on behalf of others[13]. An intelligent agent may be categorized as a human agent, or a hardware agent or a software agent. A software agent may in turn be categorized into information agent, cooperation agent and transaction agent. The agents are categorized into two major categories as internal and external properties. Internal properties, describes the actions within the agent, which includes the ability to learn, reactivity, autonomy and goal-orientedness[11]. The external properties include all those characteristics which are associated with other agents for communication I n t e r a c t i o n
I n t e r a c t i o n
Action Information fusion
Information processing
Fig. 5.Workflow of the agent Work process of intelligent agent The functionality of the agent can be described as a black box. Input is received through perceptions by the agent. After receiving input, agent use the information received for its intelligent processing and send the output as action performed over the given input[5]. Intelligent processing of the agent is done as a series of actions performed over the input. Interaction component found in this process is responsible of sending and receiving data from or to the external components. Information fusion is responsible for collecting the inputs from the interaction component. The information processing component is responsible of handling the input by carefully analyzing the data to perform action on it. The action component is capable of performing action on the data received and sends the output to the interaction component.
In wifi positioning system access points PDA or laptop detect the signal strength and export into excel sheet is consider as primary data for data cleaning. At this stage the data collected may have erroneous information.
Velammal College of Engineering and Technology, Madurai
Page 27
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Need and the Role of Agents in Data Cleaning Agents are introduced in the data cleaning process are capable of reacting with the environment with their intelligent behavior and is used to improve performance of wifi positioning system. The Intelligent behavior is achieved by systematically analyzing the data and store the useful information alone in the knowledgebase. Intelligent agents incorporated in the preprocessing phase of the data cleaning task uses its flexible and dynamic characteristics to clean the data. Agent uses its knowledge base in the cleaning task reacts with the system by verifying the existence of the required information. Use of agents in wifi positioning system will improve the performance by not repeating the same procedure for the access positions which has intersection in signal strength. Intersection of such signal strengths causes an error and the errors positions are stored in the knowledge base of an agent. When the agents receive the data, it checks its knowledgebase to verify the existence of erroneous data. If an erroneous position was found then the agent filters this information otherwise it sends the details to perform action. Outliers are identified by using the domain of the erroneous data. Also, Agent based cleaning process distinguishes the access points which do not have the crossing point in their signal strength. Functionality of Intelligent Agents in WiFi positioning system DSP NAL
Excel sheet
Data cleaning statergies
Cleaned data
Fig.6.Data cleaning process Agents collect data from excel data sheet through the interaction component and send it to the information fusion component which is responsible of receiving data. Data collected through excel data sheet may have some erroneous signal coordinates where we cannot locate either the DSP or NAL. These signal coordinates collected through information fusion component are transferred to information processing component. Information processing component is responsible of computing the error distance between access points. If there is an error between the access points then those access points are
Velammal College of Engineering and Technology, Madurai
transferred to the action component of the agent, which in turn stores the details of the access point coordinates in its knowledge base as action. The error free access point coordinates alone be transferred as output of the action component through the interaction component. At this stage the data processing handles only those access points which are error free causes improvement in average signal strength computation. Ultimately the computation time required to accomplish the MATLAB simulation will also be greatly be reduced. Hence, the use of agents in the wifi positioning system improves the performance in one or more aspects. Data cleaning Agent Procedure PDA receives signals from the access points. The netstumbler software detects the signal strength from the PDA and explores the signals into the excel data sheet. This data sheet is given as an input to the Data Preprocessing component. The preprocessing component is to clean the given data sheet with the help of intelligent agents. Agents are incorporated in the intelligence component to improve the performance of the system. Agents analyze the data access points carefully and compute the error distance to each of the access points. If agents encounters an erroneous data then agents learns to update its knowledge base. Clusters are formed using the erroneous data that are stored in the knowledge base. Before sending the details of the access points to the action component, agent stores all such erroneous coordinates of the access points in its knowledge base. Data cleaning is a treatment for decreasing possible negative influence[3]. An outlier is an observation which deviates so much from other observations as to arouse suspicions that was generated by a different mechanism. The goal of outlier detection is to uncover the “different mechanism”. If samples of the different mechanism exist in knowledge base, then build a classifier to learn from samples. For fine classification, a function that designates the distribution of samples is necessary. Agents are intelligent systems that verify the input data (access points) from its knowledge base to identify the outlier. If the access point coordinates are error free then the agent uses the access point coordinates to position the laptop. Otherwise the agent filters the details of erroneous access points by not sending them to the next layer. Use of agents in this process eliminates the computation time required to find the erroneous data using MATLAB. Agent based data cleaning eliminates the computation time associated with the detection of erroneous data which causes performance improvement in the system.
Page 28
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
What makes the difference in the performance? In the existing system the work process is done manually and the errors are detected individually until the positioning is done. Though the manual work progress successfully, the work process is repeated for all the access point signal strength. To overcome this repetition problem and not go beyond the cross point’s intelligent agents are introduced in the cleaning system. Agent makes use of the error distances stored in its knowledge base to clean the erroneous signals.
Fig.7. Agents (Intelligent Process) in positioning process 5. Results and Performance evaluation What the existing system will do? In the existing system, access points are positioned with the help of GIS and the errors positions are to be ignored are calculated with the help of MATLAB. All the operations taken individually and the signals are transformed into the excel sheet with the help of the software Netstumbler. Error distance is calculated for each of the floor positions using MATLAB and the error signal strengths are stored in the form of matrix. The next step is to position the access point according to the erroneous distance. This process to done manually and every signal strength are stored separately requires more time and makes the work process as a long one.
The process is done for all the areas and the agent stores only the erroneous coordinates in the knowledgebase. Time is saved by the use of agent while experimenting this process. The accuracy of the positioning is improved when agents are incorporated in the system with is the aid of geographical information system. In the existing system performance is measured by means of error distance which requires the help of GIS but for the agent based system error distance is calculated and stored in the knowledgebase.
Fig. 8. Floor plan with corridor dimension What Agent based system is doing? The main objective to introduce the agents in the positioning system improves the performance by cleaning the erroneous data and to store it in the knowledge base for further processing. The accuracy and the time required to position the lap top is reduced. The signals are broadcasted from the access point and the signals make use of the Netstumbler to export the signal into excel sheet. Agent reads the excel sheet and store only the error details in its knowledge base. Agent makes use of its knowledge base to verify the error details and quit the work process for erroneous data. Reduction of processing time by the use of agents not only improves the performance but also not repeating the operations which are already done during the cleaning task will also be considered as one of the performance improvement factor. Thus Agents present the data cleaning system enable the system with intelligent behavior and improves positioning strategies.
Velammal College of Engineering and Technology, Madurai
The above Fig 7 displays a floor plan of our test site along with specific x and y dimensions of corridor section. Now the X and Y ranges from 0 and 19.125 and from 9.25 to 11.125 meters respectively. Use of agent in positioning system improves the accuracy and reduces the working time of the process. Also it scopes down and eliminates unlikely intersection area from positioning algorithm in the optimization phase.
6. Conclusion In this paper, Intelligent Agents are introduced to clean the erroneous access point coordinates for the wifi positioning system that uses the knowledge of geographical Information system. The performance variation of the wifi positioning system is described in terms of agent based and normal positioning. Agents are introduced between the excel data sheet and the MATLAB simulation data
Page 29
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
processing component. Future enhancement of this paper may introduce Multi-Agents to perform the cleaning task with distinct functionality.
7.References
[13] Walter Brener,Rudiger Zarnekow, “Software Agents Foundation & Application”, Morgan publishers,Elsevier,2002 [14]William E Winkler Conference SIGKDD’03
“Data
Cleaning
Methods:,
[1] Ayse Yasemin SEYDIM “Intelligent Agents: A Data Mining Perpespective”,CiteSeer .IST, Literature Digital Library. 1999. [2] Bahi,P , Padmanabhan V RADAR ; An In-building RF based User Location and Tracking System, IEEE Infocom ,March 2000 [3] S.D.Bay and M.Schwabacher. Mining distance-based Outliers in near linear time with randomization and a simple pruning rule. In Proceedings of Ninth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Washington, D.C.USA, pages29–38, 2003. [4]Erhard Rahm, Hong Hai Do, “ Data Cleaning : Problems and Current approaches” [5] Gerhard Weiss , “ Multiagent Systems, “A Modern Approach to Distributed Artificial Intelligence”, The MIT Press, , 1999 [6]Hightower,J and Borriello G “Location Systems for Ubiquitous Computing”, IEEE Communic. Mag. August 2001, 57-66 [7] Hightower J and Borriello G, “Location Sensing Techniques”, IEEE Communic. Mag , August 2001 Technical Report [8] Jan R H, Lee Y.R “ An Indoor Geolocation System for wireless Lans” Proceedings of the International Conference Parallel Process Workshop (ICPPW’03) 2003. [9] Jiawei Han , Micheline Kamber “ Data mining: Concepts and Techiniques “ , Morgan Kaufmann Publishers, Elsevier, 2001 [10] http://www.netstumbler.com [11] S.Russell, P. Norvig, Artificial Intelligence, “ A Modern Approach, Printice-Hall,1995. [12]Tussanai Parthornratt , Kittiphan Techakittiroj, “ Improving Accuracy of WiFi positioning system by using Geographical Information System “, AU J.T. 10(1) : 38-44, Jul 2006
Velammal College of Engineering and Technology, Madurai
Page 30
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Framework for Multiple Classifier Systems Comparison (MCSCF) P.Shunmugapriya ,
S.Kanmani
Research Scholar, Department of CSE, Pondicherry Engineering College, Pondicherry, India
Professor & Head Department of IT Pondicherry Engineering College, Pondicherry, India
[email protected] Abstract In this Paper, we propose a new framework for comparing Multiple Classifier Systems in the literature. A Classifier Ensemble or Multiple Classifier System combines a finite number of classifiers of same kind or different types, which are trained simultaneously for a common classification task. The Ensemble is able to efficiently improve the generalization ability of the classifier when compared with a single classifier. The objective of this paper is to introduce a framework MCSCF that analyses the existing research work on Classifier Ensembles. Our framework compares the classifier ensembles on the basis of Constituent classifiers of an ensemble and Combination methods, Classifier selection basis, Standard datasets, Evaluation criteria and Behavior of classifier ensemble outcomes. It is observed that, different types of classifiers are combined using a number of different fusion methods and classification accuracy is highly improved in the ensembles irrespective of the application domain. Keywords - Data Mining, Pattern Classification, Classifier Ensemble, Multiple classifier Systems
I. INTRODUCTION Combining classifiers to get higher classification accuracy is rapidly growing and enjoying a lot of attention from pattern recognition and machine learning communities. For ensembling, the classification capabilities of a single classifier are not required to be very strong. What is important is to use suitable combinative strategies to improve the generalization of classifier ensembles. In order to speed up the convergences and simplify structures, the combinative components are often “weak” or simple [1]. Classifiers like Region Based Classifier ,Contour Based Classifier, Enhanced Loci classifier, Histogram-based classifier , Crossing Based Classifier, Neural Networks classifier, K-nearest neighbor classifier, SVM, Anorm, KDE, KMP(Kernel Matching Pursuit), Minimum Distance Classifier, Maximum Likelihood Classifier, Mahalanobis Classifier, Naïve Bayesian, Decision tree, Fisher classifier, Nearest Means classifier are often combined to form classifier fusion or classifier ensemble as shown in [Table I].
Velammal College of Engineering and Technology, Madurai
[email protected]
Also there are many classifier combination methods available. By using different combination methods to a number of various classifiers, classifier ensembles result with variations in their performance and classification accuracy. So far, different methods of classifier ensembles have been experimented on different application domain. The important objective of this paper is to provide a comparison framework of works on classifier combination .i.e. what classifiers are used in ensembles, what are the methods of combination that have been applied, what application domain and datasets are considered. [2] and [9] are classifier ensemble works carried out in context aware framework using evolutionary algorithm and [2] on line proposal. [3] uses voting method to form an ensemble that classifies Hyper spectral data. [5] considers distribution characteristic and diversity as the base features for classifier selection and uses k-means algorithm, a new kernel based probability c-means (KPCM) algorithm is designed for classifier ensemble. [6] is double bagging, a variable method of bagging .This uses SVM with 4 kernels and has shown enhanced classification accuracy than bagging and boosting. A-priori knowledge about classifiers is considered as an important criterion in selecting the classifiers [7] and uses a number of classifiers and ensemble methods to form an ensemble that has always 90% accuracy of classification. Boosting method is employed to obtain diversity among the classifiers and uses very effective method of ensembling - stacked generalization [8]. In [10] feature selection and GA are used and the ensemble performs well in accurate prediction of cancer data. For the area of Remote sensing, [11] considers diversity as an important feature and works on satellite image. [12], [14], [18] and [19] are also works carried out giving importance to diversity of classifiers. [16] is based on feature selection that uses product rule for the ensemble. Certain works are carried out in interest on looking at the performance of classifier ensembles, by using several combinations of classifiers and then look for the best combination [17], [24] and [25].
Page 31
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The text is organized as follows: Section II is about data mining, pattern classifier and classifier ensemble. Section III presents the comparative framework of works on classifier ensemble. In section IV, we discuss the comparison of ensemble results and our inference from the framework. Section V is our conclusion. II. DATA MINING Data mining or Knowledge Discovery in Databases (KDD) is a method for data analysis developed in recent years. It is an interdisciplinary research focusing upon methodologies for discovering and extracting implicit, previously unknown and potentially useful knowledge (rules, patterns, regularities as well as constraints) from data. The mining task is mainly divided into 4 types: Class/Concept description, Association analysis, Classification or Prediction and Clustering analysis. Among them, the classification task has received more attention. In the task of classification, the mining algorithm will generate the exact class description for the classification of unknown data by analyzing the existing data. A. Pattern Classification Classification is the action of assigning an object to a category according to the characteristics of the object. In Data Mining, Classification refers to a form of Data analysis which can be used to extract models describing important data classes or to predict future data trends. i.e. To classify objects(or patterns) into categories(or classes). Classification has 2 stages: • Model is determined from a set of data called training data. i.e. classes have been established before hand. The model is known as rules, decision trees or mathematical formula. • Correctness of the evolved model is estimated. This is done by studying the results of the evolved model’s function on a set of data called test data. Classes of the test data are determined before hand. Classification techniques that are available are Decision Tree Learning, Decision Rule Learning, Naive Bayesian Classifier, Bayesian Belief Networks, Neural Networks, kNearest Neighbor Classifier, Support Vector classifier. These techniques differ in the learning mechanism and in the representation of the learned model. Classification algorithms are Decision tree, Correlation analysis, Bayesian classification, Neural networks, Nearest neighbor, Genetic Algorithms, Rough sets and Fuzzy
Velammal College of Engineering and Technology, Madurai
technology. Also methods like Support Vector Machine (SVM), Neural Networks, K Nearest Neighbor, K Means, and Boosting and Quadratic classifiers exist.. SVM is the most preferred method in many classification and machine learning tasks. Also many standard tools like WEKA, KNIME, Rapid Miner, R Project, MATLAB tools are available for classification. Several frameworks like LIBLINEAR and LIBSVM are useful resources available on the web for classification. Many people and industries are interested in the decision support system and prediction systems for the better choice and risk. Classification forms the basis of these systems. B. Multiple Classifier Combination Though many classifiers exist and are widely used for classification in different application areas, the classifiers are not able to give a good accuracy of classification (Performance accuracy). To enhance the performance of classification algorithms, many works were carried out and there comes the idea of combining different classifiers to perform the classification task. Such a combination is called by names like Combination of Multiple Classifiers, Committee Machine, Classifier Ensemble and Classifier Fusion. The most popular usage is Classifier Ensemble. Combining classifiers to get a higher accuracy is an important research topic in the recent years. Classifier Ensemble is a learning paradigm where several classifiers are combined in order to generate a single classification result with improved performance accuracy. There are many methods available for combining the classifiers. Some of the popular ensembling methods are Majority voting, Decision templates, naïve Bayesian, Boosting, Bagging etc. There are many proposals for classifier ensembles and also works are still under pursue. To have an idea about what are the types of ensembles that have been proposed so far, we have given a comparative framework of existing ensemble methods in section III. [1].3
COMPARATIVE FRAMEWORK OF WORKS ON CLASSIFIER ENSEMBLE
The framework for comparing the existing classifier ensembles is given in Table I. Features for classifier selection, types of classifiers, combination methods, datasets, evaluation criteria are the important features discussed in the framework. Ensemble works are carried out in 2 perspectives: a. To look for the performance of ensemble, by combining a number of different classifiers using various combination methods. In certain works, the classifiers that are to be put into combination are selected on the basis of several features like,
Page 32
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
diversity, accuracy, feature selection, feature subspaces, random sampling etc.. b. Application of different ensemble works to a number of application domain and several standard datasets and conclude the performance of the ensemble. A. Classifier Selection Basis Classifiers are selected for a particular ensemble, based on several features as specified in the 3rd column of Table I. The features are Prior knowledge about classifiers, Diversity of classifiers, Feature selection, Reject option and Random subspaces. L.R.B.Schomaker et al. [7] considers Prior knowledge about the behavior of classifiers as an important factor. A prior knowledge about classifiers performing classification in single or when put into combination the way they behave, plays an important role in his work. Another criterion for selection of classifiers is Diversity of classifiers. The diversity of the classifier outputs is therefore a vital requirement for the success of the ensemble. [1]. Diversity indirectly talks about the correctness of the individual classifiers that are to be combined. [5], [8], [11], [12], [14], [19] are some samples that say about ensemble works in which, classifiers are selected on the basis of diversity of classifiers. The diversity among the combination of classifiers is defined as: if one classifier has some errors, then for combination, we look for classifiers which have errors on different objects [1]. There are two kinds of diversity measures: pairwise and nonpairwise. Pairwise measures are taken ,considering two classifiers at a time and Non pairwise measures are taken considering all the classifiers at one time [23]. Both these measures have different evaluation criteria. Diverse classifiers for any work are obtained by bootstrapping of the original dataset, resampling of training set and adoption of different feature set with Bagging [20], Boosting [21] and Random Subspace method [22]. When classifiers are more diverse, it is likely to expect very good classification accuracy from them. This is because the more, the diversity of the classifiers, they are very good finding different error conditions. Feature Selection is another criterion for selecting the classifiers. Data objects contain a number of features. There may be some irrelevant features that are not helpful for classification. In that case, features that are important in identifying the uniqueness of an object are selected by using several feature selection algorithms. After deciding the features, classifiers that can work on this particular feature subset are selected and combined [10], [15] discusses the ensemble proposals on the basis of feature selection
Velammal College of Engineering and Technology, Madurai
Classifiers of an ensemble are also selected based on the assumption that they are independent of each other and also has some diversity [5]. Distribution characteristic is difference in performance of classifiers in different region of characteristic space. Apart from performing the task of classification correctly, classifiers should also be capable of identifying data objects that do not belong to a particular class and capable of rejecting them. When such classifiers are selected for the ensemble, it is called as classifiers with Reject option. Apart from the above features, classifiers are also selected related to the type of application domain and the method of ensemble to be applied. Certain ensembles are fit into context aware framework and in most of such cases evolutionary algorithm is used for classification leading to a very good classification performance. [2] and [9] discusses such type of methodology. GA plays a very important role in combining classifiers [11]. When GA is the method of ensemble to be applied, classifiers that can work well in GA framework are selected for the ensemble. B. Classifiers for Ensemble Different kinds of standard classifiers exist and most of the classifiers perform well when put into an ensemble. Region Based Classifier, Contour Based Classifier, Enhanced Loci classifier, Histogram-based classifier, Crossing based classifier, Nearest Neighbor, K- Nearest Neighbor ( K taking values 1,2,3 etc), Anorm, KDE, Minimum Distance classifier, Maximum Likelihood Classifier, Mahalanobis Classifier , SVM, KMP(Kernel Matching Pursuit), Decision Tree, Binary Decision Tree, Nearest Mean Classifier, Maximum Entropy Model, Heterogeneous Base Learner, Naïve Bayesian classifier, Linear Discriminant Analysis, Non-Linear Discriminant Analysis and Rough Set Theory are the classifiers used in most of the ensemble applications. When diversity is the base feature considered, then by using methods like Bagging, Boosting, Random Sampling, the classifiers are trained and thereby the classifiers are diversified. The diverse classifiers are then put into ensemble. In case of Context Aware Framework, the classifiers are trained with different context conditions before put into ensemble [2], [9]. Cost sensitivity is considered as the important factor for a classifier and in such cases the classifier is trained to possess cost sensitivity [18]. SVM is one such classifier trained to have cost sensitivity and it is CS-SVC.
Page 33
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
D. Datasets for Classifier Ensemble SVM is used as such in many ensembles. There are 4 kernel functions with which SVM can combine and yield very good classification accuracy: Linear Kernel, Polynomial Kernel, Radial Basis Kernel and Sigmoid kernel. SVM with Radial Basis Kernel yields a very good performance.
C. Classifier Combination Methods Majority Voting, Weighted Majority Voting, Naïve Bayesian Combination, Multinomial Methods, Probabilistic Approximation, Singular Value Decomposition, Trainable Combiners, Non Trainable Combiners, Decision Templates, Dempster-Shafer method, Bagging, Boosting and Stacking(Stacked Generalization) are the methods available for combining the classifiers according to Ludmila I Kuncheva [1].
Datasets that are used for an individual classifier can serve as the datasets for ensemble too. Same as classifiers, ensembles have their usage in all most all the areas like Machine Learning, Image processing, Bioinformatics, Geology, Chemistry, Medicine, Robotics, Expert and Decision making systems, Diagnosis systems, Remote sensing etc. According to the type of application domain, datasets are taken from various available standard Databases. UCI datasets [12],[14],[16] and [17], ELENA database[5] and [8], EFERET, EYALE, EINHA [2] and [9] are some of the benchmark datasets that are used for experiments in most of the ensemble works. For remote sensing and land cover classification, satellite images are used. Some Researchers collect datasets on their own for their works. E. Evaluation Methods
Out of the available methods, Majority Voting, Dempster-Shafer method, Boosting, Stacking and Bagging are the popular methods which are used in most of the ensemble works [6],[7],[8],[12],[14],[17],[18],[19]. Bayesian decision rules are another popular method of ensembling. The rules are Product Rule, Sum rule, max rule, min rule and median rule.[11]. Product Rule is the rule which is most commonly used among the 5 rules. [16]. A multinomial method has two types: Behavior Knowledge Space (BKS) Method and Wernecke’s Method. Out of these 2 methods BKS is widely used [7]. In Boosting, Adaboost algorithm is the most commonly used ensemble algorithm [12]. Apart from this, GA is used for forming the classifier ensemble [2], [5], [9] and [10]. Classifiers are also fit into context aware framework for the task of ensembling [2] and [9]. In several cases, a method which is capable of integrating the advantages of existing classifiers is used for combining the classifiers. Another way is initially after classification is done by some classifiers, an additional classifier either of the same type or of a different kind is used finally for classification [3].
The performance of a classifier is a compound characteristic, whose most important component is the classification accuracy [1]. Some of the standard measures that are very often used for evaluating the performance of classifiers and ensembles are Cross validation(confusion matrix as a result of Cross validation), Kappa statistics, McNemar’s test, Cochran’s Q test, F-test, OAO(One-Against-One) approach, OAA(One-Against-All) approach, ROC(Receiver Operating Characteristic), AUC(Area Under the ROC Curve), FAR, FRR, Multinomial Selection Procedure for Comparing Classifiers [9], [10], [11]. [14], [16], [17], [19],[24]. Cross-Validation is the most popular method for estimating the classification accuracy. Some of the variants of Cross-Validation obtained by varying the training and testing data are: K-Hold-Out Paired t-Test, K-Fold Cross-Validation Paired t-Test, Diettterich’s 5 x 2-Fold Cross-Validation Paired t-Test (5x2cv). IV. RESULTS AND DISCUSSION Most of the works discussed here are carried out in order to enhance the performance of multiple classifiers ensemble and simplify the system design. Different classifiers are combined using a number of different classifier combination methods and has been experimented with standard datasets.
A new ensemble method is designed, a new ensemble algorithm is proposed or modifications are made to existing ensemble methods and algorithms in certain cases without using the existing methods [12], [16] and [24].
It is seen from [2], [3],[5],[6],[7],[8],[10],[12],[14],[16],[17][19] and [24] the new methods designed are highly robust and show improved classification accuracy than the existing ensemble methods and the individual constituent classifier. Most of the new methods show a better performance compared to individual classifiers and
Velammal College of Engineering and Technology, Madurai
Page 34
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the ensembles Bagging, Boosting, Rotation Forest and Random Forest. Not only that, the running time is improved in [14] and the training time for the ensemble is reduced by 82% without any performance degradation [18]. Also the new methods designed using Evolutionary algorithm in Context aware framework has resulted in ensembles with highest recognition rate and reduced error rates [2] and [9]. It could be inferred from the framework that: • Feature Selection and Diversity of classifiers play an important role in selecting the classifiers for an ensemble. • 75% of the ensemble works uses Cross-Validation for evaluating the performance of classifier ensemble. • Bayesian Decision Rules, Bagging, Boosting, Stacking and voting are the methods most commonly used for combination of classifiers. • UCI datasets are taken for most of the applications. • 80% of the works are on two-class classification problem. • WEKA and MATLAB are the tools used for the implementation of most of the classifier works. • 100% Classification accuracy has not been achieved in any ensemble work.
V. CONCLUSION AND FUTURE WORK A classifier ensemble is a very successful technique where the outputs of a set of separately trained classifiers are combined to form one unified prediction. First it improves the generalization performance of a classification system greatly. Second, it can be viewed as an effective approach for classification as a result of its variety of potential applications and validity. Although classifier ensembles have been used widely, the key problem for researchers is to effectively design the individual classifiers that are highly correct, with diversities between them and thereby the ensemble is also highly accurate. We have given a framework that consolidates the types of existing ensemble methods. It uses classifier selection features, types of classifiers and combination methods for comparison. Any new work can easily fit into our framework and will be useful to the researchers for making an analysis of existing work. As a Future work, Ensembles can be tried for 100% classification accuracy and for many unexplored application domain. Most of the works on binary classification can be extended for multi class classification.
Table I: Comparative Framework of Works on Classifier Ensemble Ref. No
[7]
On the Basis of
A-Priori Knowledge
Classifiers Used
Method of Ensembling
1.Region Based Classifier 2.Contour Based Classifier 3.Enhanced Loci 4.Histogrambased 5.Crossing based
1.Majority Voting(MV) 2.Dempster Shafer (DS) 3.Behavioral Knowledge Space(BKS)
Dataset
Hand Written Numerals
4 UCI datasets : 1.Pima [12]
Diversity of Classifiers
1.Knn (k=1) 2.Knn (k=3) 3.Svm (r=1) 4.Anorm 5.Kde.
Sample Size
*
764 patterns 8 features 2 classes
AdaBoosting 2.Spam
4016 patterns 54 features 2 classes
3.Haberman
2 classes
Velammal College of Engineering and Technology, Madurai
Evaluated By Comparing the classification results with that of constituent classifiers
Comparing the classification results with that of constituent classifiers
Results 1.Classificati on Accuracy ≈ 90% 2.Better Accuracy than the constituent classifiers Better Performance of the new method than the individual classifiers of the ensemble
Page 35
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
4.Horse-colic
[8]
[11]
[14]
[19]
Diversity of Classifiers
3 different classifiers diversified by Boosting
1.Boosting 2.Stacked Generalizati on
SATIMAGE Dataset from ELENA Database
Diversity of Classifiers
1.Minimum Distance Classifier 2.Maximum Likelihood Classifier 3.Mahalanobis Classifier 4.K nearest Neighbor Classifier
Five Bayesian Decision rules: 1.Product rule 2. sum rule 3. max rule 4. min rule, 5. median rule
Remote Sensing Image –SPOT IV Satellite Image
Diversity of Classifiers
1.Diversity of Classifiers
SVM KMP(Kernel Matching Pursuit)
1.Maximum Entropy Model 2.Heterogeneous
Bagging and Majority voting
L1 Regularized maximum Entropy Model
I.UCI Datasets 1.Waveform 2.Shuttle 3.Sat II. Image Recognition6 Plane Class Images
Customer Relationship Management ORANGE – A French Telecom
Velammal College of Engineering and Technology, Madurai
6435 pixels 36 attributes 6 classes
Comparing the classification results with that of constituent classifiers
6 Land Cover classes
1.Overall Accuracy 2.Kappa statistics 3.McNemar’s Test 4.Cochran’s Q Test 5.F-Test
614 Sheets 128 X 128 pixels
1.OAO Approach 2. Comparing the classification results with that of constituent classifiers
2 Versions 1.Larger Version15,000 Feature Variables 50,000
1.By Cross Validation 2. Overall Accuracy
1.Boosting generates more diverse classifiers than Cross validation. 2.Highly Robust compared to original Boosting and Stacking 1. Diversity is not always beneficial. 2.Increasing the number of Base classifiers in the ensemble will not increase the Classificati on accuracy. 1. A New Ensemble of SVM and KMP is designed. 2. High Classificati on Accuracy of SVM. 3. Quick Running Time of KMP. 1. Good Classificati on Accuracy. 2.Has won 3rd place in
Page 36
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Base Learner 2.Best Ensemble Proposal KDD Cup 2009
3.Naïve Bayesian
Boosting
MODL CriterionSelective Naïve Bayesian
Company’s Dataset 3 tasks (Churn, Appetency, Upselling)
examples
Ensemble proposal of KDD Cup 2009
2.Smaller Version: 230 features 50,000 examples
Post processing the results of 3 methods with SVM
[5]
[18]
[10]
[16]
1.Distrib ution Character istics 2.Diversi ty
1.Feature Subspace s 2.Diversi ty
Feature Selection
Feature Selection
Classifiers with Diversity
Cost Sensitive SVC s
1.Fisher classifier 2.Binary Decision Tree 3. Nearest mean Classifier 4.SVM 5.Nearest Neighbor(1-nn)
‘n’ number of SVM s
1. Kernel Clustering 2. New KPCM Algorithm
Phenome Dataset from ELENA Database
2 classes 5000 samples 5 features
Tr.Set 7931 positive 7869negative
In comparison with Bagging applied to the same dataset.
Comparison with Conventional SVCs in terms of Detection Rate and Cost Expectations
1.Random Subspace method 2.Bagging
Hidden Signal DetectionHidden Signal Dataset
Ensemble Feature Selection based on GA
1.Colon cancer data
Tr.Set – 40 Tst.Set– 22
1.10- Fold Cross Validation and
2.Hepato Cellular Carcinoma Data
Tr.Set – 33 Tst.Set– 27
2. Assigning Weights
Tst.Set 9426 positive 179,528 negative
Tr.Set – 21 Tst.Set– 29
Better Performance than bagging and any other constituent ensemble classifier 1. SVC parameter optimizations reduced by 89%. 2. Reduction in overall Training time by 82% without performance degradation.
Better prediction Accuracy
3.High grade glioma dataset
Product Rule
12 UCI benchmark Datasets
Velammal College of Engineering and Technology, Madurai
*
4- Fold CrossValidation
1.Simplifie d Datasets 2.Reduced time complexity, 3.A new FSCE
Page 37
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[9]
[2]
[17]
1.Evoluti onary Fusion 2.Context awarenes s 1.Evoluti onary Fusion 2.Context awarenes s
*
Best classifiers in each ensemble obtained by GA
1.K-Means Algorithm 2.GA
‘n’ number of different classifiers trained with different context conditions
Embedding Classifier Ensemble into Context Aware Framework
1.Naïve Bayesian 2.K-Nearest Neighbor
1.Static Majority Voting(SMV ) 2.Wighted Majority Voting(WM V) 3.Dynamic WMV(DW MV)
1.Decision Tree
[6]
*
2. SVM With 4 kernels: 1. linear, 2.polynomial 3.radial basis 4.sigmoid
1.Bagging 2. Double Bagging
Face Recognition FERET Dataset 4 Face Recognition Systems 1.E- FERET 2.E-YALE 3.E-INHA 4.IT (All 4 datasets are further divided into 3 separate datasets I, II,III )
6 contexts
Nine data sets Each containing 10000 face images under 100 kinds of illumination
1.ROC 2.FAR 3.FRR
By creating similar offline system(without Result Evaluator)trained and tested on the same dataset
UCI DataSets 1.TicTacToe EndGame 2.Chess EndGame
1.Condition Diagnosis of Electric Power ApparatusGIS dataset 2. 15 UCI Benchmark Datasets
Velammal College of Engineering and Technology, Madurai
958 Instances 9 features 2 classes 3196 Instances 36 features 2 classes
Each dataset has different number of objects, classes and features
Cross Validation
In comparison with Other ensembles’ performance on the same data.
algorithm is proposed 4.Higher Classificati on Accuracy 1.Reduced Error Rates 2.Good Recognition Rate 1.Highest Recognition Rate than the individual classifiers 2.Most Stable Performanc e 1.Better Classificati on Accuracy than the individual Classifier. 2. DWMV has Higher classificatio n Accuracy. Better performanc e than popular ensemble methods like Bagging, boosting, Random forest and Rotation
Page 38
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Forest.
[3]
[24]
[25]
*
‘n’ number of SVMs
*
*
Any classifier capable of classifying CSP
1.Rough Set Theory 2.Decision Tree 3.SVM
Additional SVM is used to fusion the outputs of all SVMs
HyperSpectral AVIRIS Data
1.Majority Voting 2.Rejecting the Outliers
BCIEEG signals
Integrating the advantages of RST, DT and SVM
UCITeaching Assistant Evaluation (TAE) Dataset
220 data channels, 145 x 145 pixels, 6 land cover classes
*
151 Instances 6 Features 2 classes
1.Overall classification Results compared to an individual classifier. 2.By simple voting.
20 – Cross Validations
1. 102(68%) – Tr.Data 49(32%) -Tst. Data 2. 6-Folds Cross Validation
Good Accuracy than a single constituent classifier in the Ensemble 1.A new ensemble CSPE is designed 2.Better performanc e than LDA , RLDA and SVM 3.Average accuracy of 83.02% in BCI (Comparitiv ely a good Accuracy) 1.Improved Class Prediction with Acceptable Accuracy. 2. Enhanced Rule Generation
Note: * - Undefined in the literature REFERENCES:
[2]. Zhan Yu, Mi Young Nam and Phill Kyu Rhee, “Online Evolutionary Context-Aware Classifier Ensemble Framework For Object Recognition”, Proceedings of the 2009 IEEE International Conference on Systems, Man, and
Cybernetics,San Antonio, TX, USA - October 2009, Page(s): 3428 – 3433, 2009 [3]. Jón Atli Benediktsson, Xavier Ceamanos Garcia, Björn Waske, Jocelyn Chanussot, Johannes R. Sveinsson, and Mathieu Fauvel, “Ensemble Methods For Classification Of Hyperspectral Data”, IEEE International Geoscience and Remote Sensing Symposium, IGARSS 2008, Volume: 1, Page(s): I-62 - I-65, 2008 [4]. Gao Daqi, Zhu Shangming, Chen Wei and Li Yongli “A Classifier Ensemble Model and Its Applns” IEEE
Velammal College of Engineering and Technology, Madurai
Page 39
[1]. Ludmila I. Kuncheva,” Combining Pattern Classifiers Methods and Algorithms” a john wiley & sons, inc., publication, tk7882.p3k83 2004
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
International Joint Conference Proceedings on Neural Networks, Volume: 2, Page(s): 1172 – 1177, 2005 [5]. Cheng Xian-yi and Guo Hong-ling, “The Technology of Selective Multiple Classifiers Ensemble Based on Kernel Clustering”, Second International Symposium on Intelligent Information Technology Application, Second International Symposium on Intelligent Information Technology Application, Volume: 2, Page(s): 146 – 150, 2008 [6]. Faisal M. Zaman Hideo Hirose, “Double SVMBagging: A New Double Bagging with Support Vector Machine”, Engineering Letter, 17:2, EL_17_2_09, (Advance online publication: 22 May 2009) [7]. In: L.R.B. Schomaker and L.G. Vuurpijl (Eds.), “Classifier combination: the role of a-priori knowledge”, Proceedings of the Seventh International Workshop on Frontiers in Handwriting Recognition, September 11-13, 2000, Amsterdam, ISBN 90-76942-01-3, Nijmegen: International Unipen Foundation, pp 143-152. [8]. Nima Hatami and Reza Ebrahimpour , “Combining Multiple Classifiers: Diversify with Boosting and Combining by Stacking”, IJCSNS International Journal of Computer Science and Network Security, VOL.7, No.1, January 2007 [9]. Suman Sedai and Phill Kyu Rhee “Evolutionary Classifier Fusion for Optimizing Face Recognition”, Frontiers in the Convergence of Bioscience and Information Technologies, Page(s): 728 – 733, 2007 [10]. Kun-Hong Liu, De Shuang Huang “Microarray data prediction by evolutionary classifier ensemble system” Kun-Hong Liu; De-Shuang Huang; Jun Zhang; IEEE Congress on Evolutionary Computation, Page(s): 634 – 637, 2007 [11]. Man Sing WONG, and Wai Yeung YAN “ Investigation of Diversity and Accuracy in Ensemble of Classifiers using Bayesian Decision Rules” , International Workshop on Earth Observation and Remote Sensing Applications, Page(s): 1 – 6, 2008 [12]. Abbas Golestani , Kushan Ahmadian Ali Amiri,and MohammadReza JahedMotlagh “ A Novel Adaptive-Boost-Based Strategy for Combining Classifiers Using Diversity Concept” 6th IEEE/ACIS International Conference on Computer and Information Science (ICIS 2007), Page(s): 128 – 134, 2007 [13]. Gou Shuiping ,Mao Shasha and Licheng Jiao,” Isomerous Multiple Classifier Ensemble Method with SVM and KMP “ International Conference on Audio, Language and Image Processing, Page(s): 1234 – 1239, 2008 [14]. D.M.J. Tax and R.P.W. Duin “Growing a Multi Class Classifier with a Reject Option” Pattern Recognition Letters 29 (2008 ) 1565-1570 [15]. Bing Chen, Hua-Xiang Zhang “An Approach of Multiple Classifiers Ensemble Based on Feature Selection”
Fifth International Conference on Fuzzy Systems and Knowledge Discovery, Volume: 2, Page(s): 390 – 394, 2008 [16]. Dr.S.Kanmani, Ch P C Rao, M.Venkateswaralu and M.Mirshad “Ensemble Methods for Data Mining” IE (I) Journal-CP, Vol 90, May 2009, 15-19 [17]. Barry Y. Chen, Tracy D. Lemmond, and William G. Hanley, “Building Ultra-Low False Alarm Rate Support Vector Classifier Ensembles Using Random Subspaces” IEEE Symposium on Computational Intelligence and Data Mining, Page(s): 1 – 8,2009. [18]. Hung-Yi Lo, Kai-Wei Chang, Shang-Tse Chen, Tsung-Hsien Chiang, Chun-Sung Ferng, Cho-Jui Hsieh, YiKuang Ko, Tsung-Ting Kuo, Hung-Che Lai, Ken-Yi Lin, Chia-Hsuan Wang, Hsiang-Fu Yu, Chih-Jen Lin, Hsuan-Tien Lin, Shou-de Lin” An Ensemble of Three Classifiers for KDD Cup 2009: Expanded Linear Model, Heterogeneous Boosting, and Selective Naıve Bayes” JMLR: Workshop and Conference Proceedings 1: 1-16 KDD cup, 57-64, 2009 [19]. L. Breiman, “Bagging predictors,” Machine Learning, vol. 24, no. 2, pp.123-140, August 1996. [20]. Y. Freund and R. Schapire, “Experiments with a new boosting algorithm,” Proceedings of 13th International Conference on Machine Learning, Bari, Italy, pp. 148-156, July 1996. [21]. T. K. Ho, “The random subspace method for constructing decision forests,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 20, no. 8, pp. 832844, August 1998.
Velammal College of Engineering and Technology, Madurai
Page 40
[22]. L. I. Kuncheva, and C. J. Whitaker, “Ten measures of diversity in classifier ensembles: limit for two classifiers,” DERA/IEE Workshop on Intelligent Sensor Processing, Birmingham, U.K., pp.10/1-10/6, February 2001. [23]. Xu Lei, Ping Yang, Peng Xu, Tie-Jun Liu, and DeZhong Yao, “Common Spatial Pattern Ensemble Classifier and Its Application in Brain-Computer Interface”, Journal of Electronic And Technology Since Of China, Vol. 7, No. 1, March 2009 [24]. Li-Fei Chen, “Improve class prediction performance using a hybrid data mining approach “, International Conference on Machine Learning and Cybernetics, Vol 1, pp 210-214, 2009
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Efficient Apriori Hybrid Algorithm For Pattern Extraction Process 1,2
J.Kavitha1,D.Magdalene Delighta Angeline2,P.Ramasubramanian3 Lecturers,Department of Computer Science and Engineering, Dr.G.U.Pope College of Engineering, Sawyerpuram-628251, Tamilnadu, India.,3 Professor, Vellammal College of Engineering and Technology, Madurai 1
[email protected] [email protected] 3 [email protected] 2
Abstract— Student’s placement in industry for the practicum training is difficult due to the large number of students and organizations involved. Further the matching process is complex due to the various criteria set by the organization and students. This paper will discuss the results of a pattern extraction process using association rules of data mining technique where Apriori Hybrid algorithm was chosen. The data use consists of Bachelor of engineering students in the teaching organization from the year 2009 till 2010. Keywords— Data mining, KDD, association rules, apriori Algorithm, Knowledge Discovery
student’s majoring. Usually, organization will request student with a specific majoring details. Other criterion is student’s Percentage. Also, due to certain work prospect, some organization request student based on the gender and race. These criteria have been considered by the program coordinator in the placement process to ensure the right student being sent to the right organization. This study aim to identify the patterns in matching organization and student and to extract hidden information from previously matched practicum placement datasets. This paper discusses the application of data mining technique particularly association rules to extract the historical placement pattern. This pattern will be a useful guideline for future organization and student matching. The data consist of all engineering undergraduate students from the year 2009 till 2010. We then discuss how the best features of Apriori and AprioriTid can be combined into a hybrid algorithm, called AprioriHybrid. The problem of finding association rules falls within the purview of database mining, also called knowledge discovery in databases.
VI. INTRODUCTION The teaching organization is responsible with the placement of students in the industry for the internship program. It is experiencing difficulty in matching organization‘s requirement with students profile for several reasons. This situation could lead to a mismatched between organization’s requirement and students’ background. Hence, students will face problems in giving good service to the company. On the other hand, companies too could be facing difficulties in training the students and assigning them with a project. The placement must be based on certain criteria in order to best serve the organization and student. For example, student who lives in Chennai should not be sent to an organization located in Bangalore. This is to avoid problems in terms of accommodation, financial, and social. It has been decided that practicum students’ should match the organization’s requirement. However, due to the large number of students registered every semester, matching the organization with the students is a very tedious process. The current procedures in matching organization and students involve several steps. First, the registered city1 (is the first choice for students) and city2 (is the second choice for students) will be examined. A match between organizations location and student’s hometown will be determined. The next criterion is the
VII. LITERATURE REVIEW Data mining have been applied in various research works. One of the popular techniques used for mining data in KDD for pattern discovery is the association rule [1]. According to [2] an association rule implies certain association relationships among a set of objects. It attracted a lot of attention in current data mining research due to its capability of discovering useful patterns for decision support, selective marketing, financial forecast, medical diagnosis and many other applications. The association rules technique works by finding all rules in a database that satisfies the determined minimum support and minimum confidence [3]. An algorithm for association rule induction is the Apriori algorithm, proven to be one of the popular data mining
Velammal College of Engineering and Technology, Madurai
Page 41
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
techniques used to extract association rules [4], implemented the Apriori algorithm to mine single-dimensional Boolean association rules from transactional databases. The rules produced by Apriori algorithm makes it easier for the user to understand and further apply the result. [5] Employed the association rule method specifically Apriori algorithm for automatically identifying new, unexpected, and potentially interesting patterns in hospital infection control. Another study by employed Apriori algorithm to generate the frequent item sets and designed the model for economic forecasting, presented their methods on modelling and inferring user’s intention via data. VIII. PATTERN EXTRACTION All paragraphs must be indented. All paragraphs must be justified, i.e. both left-justified and right-justified. A. Selection The data have been generated by different reports among others Registered Students Report, Students’ Mark Report, Students’ List Based on City Report. This data include all 2009 and 2010 Bachelor in Engineering students. The initial data contains the performance profile gathered from a number of 125 students with 20 listed attributes which include Register Number, Programme, Duration, Program Code, City1, City2, Address, Address State, Percentage, Gender, Race Code, Race, Organization, Address1, Address2, Postcode, City3 and State. The data contains various types of values either string or numeric value. The target is represented as Organization’s name. The Organization’s name was grouped according to two categories (Government and Private). Based on the discussion with the program coordinator, all 125 data are used in this study. The selected attributes are Majoring, Percentage, Gender, City1, Race, Organization and City3 chosen based on the suitability of the condition of the problems being discussed. The data were then processed for generating rules. B. Pre-processing No Upon initial examination on the data, missing values of the attributes City1, Percentage, Race, Gender, Organization and City3 were found and removed according to the numbers of missing values in one instance as part of the data cleansing process. C. Transformation According to [9], after the cleansing process, data is converted into a common format to make sure that the data mining process can be easily performed besides ensuring a
Velammal College of Engineering and Technology, Madurai
meaningful result produced. The following rules are used to transform the Percentage to string data. 1. If the Percentage = 81 Till 90 THEN Replace Percentage by S1 2. If the Percentage = 75 Till 80 THEN Replace Percentage by S2 3. If the Percentage = 70 Till 74 THEN Replace Percentage by S3 4. If the Percentage = 65 Till 69 THEN Replace Percentage by S4 Transformation has also been applied to attributes city1 and city3 by grouping several cities together according to their location or region, decoded into new region using code of each state. For example, KODAMPAKAM and GUINDY have the same code 02 then they were converted into one Region (N_Region). Organization’s name was also transformed by into two categories (Government and Private). After all pre-processing and transformation have been implemented, the data was than ready to be mined using association rules.
D. Pattern Extraction using Apriori Algorithm In this study, the association rules using Apriori and Apriori TID Algorithm was applied to the data for generating rules. E. Apriori Algorithm Fig.1 gives the Apriori algorithm. The first pass of the algorithm simply counts item occurrences to determine the large 1-itemsets. A subsequent pass, say pass k, consists of two phases. First, the large itemsets Lk-1 found in the (k-1)th pass are used to generate the candidate itemsets Ck, using the apriori-gen function. Next, the database is scanned and the support of candidates in Ck is counted. For fast counting, we need to efficiently determine the candidates in Ck that are contained in a given transaction t. 1) L1 = {large 1-itemsets}; 2) for ( k = 2; Lk-1≠0; k++ ) do begin 3) Ck = apriori-gen(Lk-1 ); // New candidates 4) for all transactions t ε D do begin 5) Ct = subset(Ck , t); // Candidates contained in t 6) for all candidates c ε Ct do 7) c:count++; 8) end 9) Lk = {c ε Ck | c:count _ minsup} 10) end 11) Answer = n Lk; Fig. 1 Algorithm Apriori
Page 42
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
F. Algorithm AprioriTid The AprioriTid algorithm, shown in Fig. 2, also uses the apriori-gen function to determine the candidate item sets before the pass begins. The interesting feature of this algorithm is that the database D is not used for counting support after the first pass. Rather, the set Ck is used for this purpose. Each member of the set Ck is of the form < TID; fXkg >, where each Xk is a potentially large k-item set present in the transaction with identifier TID. For k = 1, C1 corresponds to the database D, although conceptually each item i is replaced by the item set fig. For k > 1, Ck is generated by the algorithm (step 10). The member of Ck corresponding to transaction t is <t:T ID, fc 2 Ckjc contained in tg>. If a transaction does not contain any candidate k-item set, then Ck will not have an entry for this transaction. Thus, the number of entries in Ck may be smaller than the number of transactions in the database, especially for large values of k. In addition, for large values of k, each entry may be smaller than the corresponding transaction because very few candidates may be contained in the transaction. However, for small values for k, each entry may be larger than the corresponding transaction because an entry in Ck includes all candidate k-item sets contained in the transaction.
earlier in the study while the degree of rules coverage was shown through the value of support parameter. IV EXPERIMENTS In this experiment, the data has been grouped into three groups based on the Organization category. Again, the experiment was conducted using Apriori algorithm with the same specifications. Table 1 shows the results generated by Apriori for all two categories of organizations.
TABLE 1 EXTRACTED PATTERN BASED ON ORGANIZATION CATEGORY Organization
Region
Criteria (Apriori) Major=Computer Science and Engineering Percentage=75-80
1) L1 = flarge 1-itemsetsg; 2) C1 = database D; 3) for ( k = 2; Lk-1 ≠0; k++ ) do begin 4) Ck = apriori-gen(Lk-1 ); // New candidates 5) Ck =0; 6) for all entries t ε Ck-1 do begin 7) // determine candidate itemsets in Ck contained // in the transaction with identi_er t.TID Ct = {c ε Ck | (c - c[k]) ε t:set-of-itemsets ^ (c - c[k-1]) ε t.set-of-itemsets}; 8) for all candidates c ε Ct do 9) c:count++; 10) if (Ct ≠0 ;) then Ck += < t:TID;Ct >; 11) end 12) Lk = {c ε Ck | c:count _ minsup} 13) end 14) Answer = Ụk Lk;
Gender=Male Race = Guindy
N_Region1
Major=Electronics Communication and Engineering Percentage=75-80
Government
Gender=Male Race = Guindy
Major=Electrical and Electronics W_Region2
Engineering Percentage=75-80
Fig. 2 Algorithm AprioriTid
Gender=Male Race = Guindy
G Interpretation/ Evaluation
During the process of pattern extraction, the acceptance of the output produced was evaluated in terms of accuracy and converge. This is to make sure that the generated rules are reliable and accurate. The accuracy of rules was obtained according to the value of confidence parameter determined
Velammal College of Engineering and Technology, Madurai
N_Region1
Major=Computer Science and Engineering Percentage=70-74 or 75-80 Gender=Female or Male
Page 43
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Race = Guindy or Kodambakam W_Region2
Major=Electronics Communication and Engineering
Private
Percentage=70-74 Gender=Male Race = Guindy
Major=Electrical and Electronics Engineering Percentage=70-74 Gender=Male Race = Guindy
A Discussion on the Apriori result From the pattern extracted, it was found that Apriori algorithm could generate patterns that are believed to be the factors that affect the matching process. From the experiment, extraction of the hidden information reveals that organization requirement can be fulfilled based on only three or four criteria. The best rules were selected where the Organization was set as the target of the students. The rules were evaluated based on the confidence and support. Upon examining Table 1, example of pattern extracted is IF students are from the Computer Science and Engineering AND Their Percentage is between 75-80 AND They are Guindy THEN The students were placed in the Northern Region and In an Government Organization IF students are from the Electrical and Electronics Engineering and Electronics Communication and Engineering Majoring AND Their Percentage is between 70-74 AND They are Guindy THEN The students were placed in the Western Region and In a Private Organization
Velammal College of Engineering and Technology, Madurai
V CONCLUSIONS This study has been implemented and conducted on existing data from the teaching organization. In this study data mining techniques namely association rule was used to achieve the goal and extract the patterns from the large set of data. Using organization category as the target, the patterns extracted can provide information of the practicum placement and how the matching of the organization’s requirement and student’s criteria was done previously. Further analysis can be done by changing the target attributes. ACKNOWLEDGMENT
Magdalene Delighta Angeline is a Lecturer in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. She obtained her Bachelor degree in Information Technology from Anna University, Chennai in the year 2007 and she is doing Master degree in Computer and Information Technology in Manonmaniam
Page 44
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Sundaranar University, Tirunelveli. She has over 3 years of Teaching Experience and published two paper in national conferences. Her current area of research includes Image Processing, Neural Networks, and Data Mining. Email: [email protected]. J.Kavitha is a Lecturer in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. She obtained her MCA from IGNOU and Master degree in Engineering from Anna University, Chennai in the year 2004 and 2007 respectively. She has over 3 years of Teaching Experience and published 2 papers in National conferences. Her current area of research includes Image Processing, Image Restoration and Data Mining. Email: [email protected]. P.Ramasubramanian, is Professor and Head in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. He obtained his Bachelor and Master degree in Computer Science and Engineering from M.K.University, Madurai in the year 1989 and 1996 respectively. He has submitted his Ph.D thesis to Madurai Kamaraj University, Madurai. He has over 22 years of Teaching Experience and authored 15 books and 22 research papers in International, National Journals and Conferences. His current area of research includes Data Mining, Data Ware housing, Neural Networks and Fuzzy logic. He is a member of various societies like ISTE, International Association of Engineers, Computer Science Teachers Association, International association of Computer Science and Information Technology and Fellow in Institution of Engineers (India). Email: [email protected], [email protected].
[4] Agrawal, R., C. Faloutsos, and A. N. Swami (1994). Efficient similarity search in sequence databases. In D. Lomet (Ed.), Proceedings of the 4th International Conference of Foundations of Data Organization and Algorithms (FODO), Chicago, Illinois, pp. 69-84. Springer Verlag [5] Ma, Y., Liu, B., Wong, C. K., Yu, .S., & Lee, S. M. (2000). Targeting the Right Student Using Data Mining , ACM, PP. 457-463. [6] R. Agrawal, T. Imielinski, and A.Swami.Database mining: A performance perspective.IEEE Transactions on Knowledge and Data Engineering, 5(6):914{925, December 1993. Special Issue on Learning and Discovery in Knowledge-Based Databases. [7] Almahdi Mohammed Ahmed , Norita Md Norwawi , Wan Hussain Wan Ishak(2009), Identifying Student and Organization Matching Pattern Using Apriori Algorithm for Practicum Placement, International Conference on Electrical Engineering and Informatics ,Selangor, Malaysia. [8] Jiawei Han, Micheline Kamber. "Data Mining : Concepts and Techniques " book: Data mining (2001). [9] Zhigang Li, Margaret H. Dunham, Yongqiao Xiao: STIFF: A Forecasting Framework for SpatioTemporal Data. Revised Papers from MDM/KDD and PAKDD/KDMCD 2002: 183-198. [10] Rakesh Agrawal Ramakrishnan Srikant, Fast Algorithms for Mining Association Rules, Proceedings of the 20th VLDB Conference Santiago, Chile (1994).
REFERENCES [1] Hipp, J., Guntzer, U., Gholamreza, N. (2000). Algorithm for Association Rule Mining: A General Survey and Comparison, ACM SIGKDD, volume 2 (Issue 1), p. 58. [2] Fayyad, U. M., Shapiro, G. P., Smyth, P., and Uthurusamy, R. (1996). Advances in Knowledge Discovery and Data Mining, Cambridge, AAAI/MIT press [3] Liu, B., Hsu, W., Ma, Y. (1998). Integrating Classification and Association Rule Mining, American Association for Artificial Intelligence
Velammal College of Engineering and Technology, Madurai
Page 45
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
CLD for Improving Overall Throughput in Wireless Networks Dr. P. Seethalakshmi1, Ms. A. Subasri2
1
Professor,Department of Computer Science, Anna University, Trichy, TamilNadu, India. 2 Student, Anna University, Trichy, TamilNadu, India 2
[email protected]
Abstract— In this project a cross-layer design approach to enhance throughput in multi-hop wireless network is proposed. The lossy nature of wireless link decreases the data rate of a given flow and the data flow becomes smaller and smaller along its routing path. As a consequence, the data rate received successfully at the destination node is typically lower than the transmission rate at the source node. The losses are due to interference and fading. These are to be minimized by adjusting the rate of flow and maximum throughput is to be achieved. The different types of interference at each layer are minimized applying various techniques like modulation and coding techniques at each layer. Maximum throughput is obtained by devising throughput maximization algorithms, and Interference aware algorithm.. The Received signal strength, quality of the link, signal to interference noise ratio play an important role in maximize overall throughput in the network. Keywords— Cross-layer design, Throughput , Interference, multi-hop
IX. INTRODUCTION In the past couple of decades, wireless communications have gained dramatic development and have been recently considered as an alternative to wire line networks in providing the last-mile broadband services. Such development further stimulates the emergence of multimedia applications, which require wireless networks to support broader bandwidth, higher transmission rate, and lower endto-end delay. For wireless communications, the challenge to provide multimedia services stems from the hostile wireless channel conditions. Besides channel noise, the time-variant channel fluctuation(i.e., channel fading) severely affects the transmission accuracy and the of quality of service(QoS).In order to combat interference and channel fading various diversity techniques, modulation and coding techniques are used. In the paper[1] the lossy feature of wireless links is studied and a leaky pipe flow model is designed where the flow rate changes per hop, which naturally points to hop-byhop rate control. The effective network utility is determined by considering two constraints, With link outage constraints and with path outage constraints. Hop-by-hop rate
Velammal College of Engineering and Technology, Madurai
control algorithm that are jointly optimized are used. CSIchannel state information is used. In the above work the interference considerations are not analysed. For achieving throughput maximization only estimations are done and analysed. Rayleigh fading model is used. II. RELATED WORKS This scheme [8] is based on belief propagation, which is capable of fully exploiting the statistics of interference. Consider, the detection of a sequence of symbols of the desired user with one strong interferer of the same signalling format, where the fading processes of both the desired user and the interference are Gauss-Markov in nature. Belief propagation is an iterative message passing algorithm for performing statistical inference on graphical models by propagating locally computed beliefs. Belief propagation algorithm has significant performance gain over the traditional interference suppression schemes. A. Local Balancing And Interference Aware Channel Allocation: This algorithm is used for reducing the overall interference in the network. Several approaches have been proposed for minimizing the Adjacent channel effects[5] ranging from coordinating the multiple radios in the wireless node and adjusting antenna parameters and the filter characteristics to using the channel overlaps for routing data across devices operating on the non-overlapping channels. One of the popular approaches for mitigating the interference effects[4] is to choose the transmission channels carefully by making sure that nearby links are on channels that do not interfere sufficiently. However, due to the dynamic nature of the links in a wireless network, the interference characteristics may vary, and therefore the channel allocation should be adaptable to these variations.
Page 46
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
B. Analog Network Coding Wireless networks strive to avoid scheduling multiple transmissions at the same time in order to prevent interference. This paper[6] encourages strategically picked senders to interfere instead of forwarding packets, routers forward the interfering signals. The destination leverages network level information to cancel the interference and recover the signal destined to it, The result is analog network coding because , signals are sent and not bits. C. Interference Aware Load Balancing. [4]Path weight function and routing schemes to provide interference aware and multichannel aware load balancing for mesh networks is done. The objective of load balancing is essentially to distribute traffic among different paths to avoid creating congested areas and improve network performance. D. Co- Operative Diversity To determine the optimal trade-off between the amount of throughput gain obtained via co-operation and the amount of interference introduced to the network. [2]Co-operative regionfor each active node is maintained, The nodes lying in such a region are allowed to co-operate with the source. They adopt decode and forward scheme at the relays and use the physical interference model to determine the probability that a relay node correctly decodes its corresponding source. Limitations Of Existing System All the techniques have their own limitations according to the design they have arrived. Tradeoffs between certain parameters is made. Link quality varies drastically. III. DESIGN AND IMPLEMENTATION In this project overall network throughput is to be enhanced by considering all the performance metrics and analyzing the parameters at each layer. The cross layer design is proposed for interactions between different layers. Using cross layer design the layers to be considered are Physical layer, MAC/ Data link layer.
S1
D1
Fig. 1 One source and one destination node
Velammal College of Engineering and Technology, Madurai
. In the above scenario there is no interference as there is only one source and one destination node.
S
S
D
D
S
D
Fig. 2 Nodes with two interfering sources In this fig. 2 there are two interfering sources which causes an adverse effect by introducing noise . As distance between source and destination node increases the received signal strength at the destination node decreases and interference increases which leads to the degradation in the quality of the signal. Normally by setting a threshold value for the received signal strength the packets are dropped or received. Here in this project an additional information is included in terms of interference which is to decide which packets are to be dropped and which are to be passed based on the signal to interference noise ratio calculation. A. Architectural Design: For each packet, a packet error rate(PER) can be computed as a function of received power, interference from concurrent transmissions. Here interference is of high importance. The underlying modulation is BPSK(Binary Phase Shift Keying) followed by a variable rate channel code. For each packet received a signal to interference and noise ratio is calculated. The bit error rate corresponding to this SINR is obtained by lookup tables. Gaussian approximation is used for the multiuser interference and for computing the SINR. From the computed BER the PER is obtained by standard approximation. The PER is used as the parameter of a binary random variable used to decide whether the packet is properly received or not.
Page 47
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The various interferences[3] are to be analysed at each layer and their effects in increasing or reducing the throughput is taken into account and the various qos metrics like throughput, delay and jitter have to be improved by changing the parameters that are in each layer which account for various degradation. The layers to be analysed are Physical layer – Received signal strength, bit error rate, Interference or signal to Interference noise ratio, modulation techniques, Data link layer- addressing or frame checking, reliability of link, Network layer- path reliability, routing, hop count, Transport –packet delivery ratio. A rate adaptation scheme is to be proposed. B. Implementation The impact of interference on the link from traffic on neighbouring links is determined. Quality of link is measured in terms of packet delivery ratio/loss rate. In the absence of interference the link capacity would be the product of the maximum sending rate of the sender and delivery ratio of the link. QoS - metrics in networks to be dealt with includes Available bandwidth, Packet loss rate/bit error rate, delay. Current routing protocols are AODV and DSR. They select the shortest path between source and Destination. The issue in shortest path selection is that intermediate nodes having less battery power, more delay, less bandwidth, high congestion, more noise ratio in such cases the path is not optimal for long and bandwidth oriented transmissions such as multimedia or real time applications, as such applications need more bandwidth, less delay and more link life. An adaptive and optimal route can only be designed using cross layer approach in which the source and destination select route on the basis of many parameters from different layers.
Fig. 3 End-to-end- Latency
The fig. 4 shows the computed packet delivery ratio with respect to the number of nodes active in the network . As packet delivery ratio increases the overall network throughput increases.
Fig. 4 Packet Delivery Ratio Vs Time
IV RESULTS As distance between the source and destination increases the delay increases due to interference. As delay increases throughput decreases. As packet delivery ratio increases delay decreases or end to end latency decreases. The fig. 3 shows the end to end latency with respect to time .
The fig. 5 represents the reporting node which are active and listening to the transmission and the number of packetsReceived reducing the interference and latency.
Velammal College of Engineering and Technology, Madurai
Page 48
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[8] Yan Zhu, Dongning Guo and Michael L. Honig , “Joint Channel Estimation and Co-Channel Interference Mitigation in Wireless Networks Using Belief Propagation”, Department of Electrical Engineering and Computer Science Northwestern University, Evanston, Illinois 60208. [9] http://www.wirelesscommunication.nl/reference/chaptr04 /outage/compouta.htm [10] http://www.wirelesscommunication.nl/reference/chaptr05/ spreadsp /ber.htm [11] htttp://www.isi.edu/nsnam/ns/tutorial/index.html
Fig. 5 No. Of Packets Received Vs Reporting Node
CONCLUSION This project aims at reducing the different types of interferences at each layer and improving the overall throughput by introducing throughput maximization algorithm and interference aware algorithm along with rate adaptation. NS2 simulation environment is implemented to test the results. By introducing the interference aware algorithm the throughput has been improved to an increase of 3% when compared to the standard existing system.
REFERENCES [1] Qinghai Gao, Junshan Zhang and Stephen V. Hanly, “Cross Layer rate control in wireless networks with lossy links: leaky pipeflow, effective network utility maximization and hop by hop algorithms”, IEEE Transactions of wireless communications vol8,no. 6,june 2000. [2] Kaveh Pahlavan,Prashant Krishnamurthy, “Principles of Wireless Networks”, Prentice Hall of India Private Limited, 2006. [3] Jochen Schiller, “Mobile Communication” 2nd Edition, Pearson Education 2003. [4] Yaling Yang, Jun Wang and Robin Kravets, “ Interference-aware Load Balancing for Multi-hopWireless Networks”, University of Illinois at Urbana- Champaign. [5] Nitin H. Vaidya , Vijay Raman , “Adjacent Channel Interference Reduction in Multichannel Wireless NetworksUsing Intelligent Channel Allocation”, Technical Report (August 2009),University of Illinois at Urbana Champaign. [6] Sachin Katti ,Shyamnath Gollakota, Dina Katabi, ”EmbracingWireless Interference: Analog Network Coding”, mit.edu. [7] David Tse, Pramod Viswanath , “Fundamentals of Wireless Communications “, University of California, Berkeley, August13, 2004.
Velammal College of Engineering and Technology, Madurai
Page 49
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Particle Swarm Optimization Algorithm In Grid Computing Mrs.R.Aghila1, Assistant Professor, [email protected]
Mr.R.Arun Kumar2 Lecturer [email protected]
M.Harine3, G.Priyadharshini3 Final year student, [email protected]
Department of Computer Science and Engineering, K.L.N College of Information Technology, Pottapalayam, Sivagangai district Abstract—Computational grid is a hardware and software infrastructure that provides dependable, consistent, pervasive and inexpensive access to high end computational capabilities. Grid computing environment has coordinated resources, Open Standard Protocols & frameworks, QOS (Quality of Service). Grid is mainly to extract enormous power of network, heterogeneous system. Grid can be applied in biomedicine, earth science, a high energy physics, astronomy, weather forecasting etc... The main need of grid computing application is discovering and scheduling of tasks and workflows. Scheduling plays an important role in load balancing and thus avoids processing delays and over commitment of resources. An effective task scheduling algorithm is necessary to satisfy the above need. In this paper, we’ve implemented Particle Swarm Optimization (PSO) algorithm for effective task scheduling using gridsim tool in Java. We’ve also compared its performance with Genetic Algorithm (GA). We aim to generate minimum completion time while executing the tasks using PSO than that obtained using GA. Keywords— Genetic algorithm, Simulated annealing, Particle Optimization algorithm
I. A.
INTRODUCTION
GRID COMPUTING:
Grid [1] is defined as an infrastructure that couples computers, software, databases, special instruments and people across the internet and presents them as an unified integrated resource. It is mainly used to extract enormous power of network, heterogeneous system. Grids have emerged as a global cyber-infrastructure for the nextgeneration of e-Science and e-business applications, by integrating large-scale, distributed and heterogeneous resources. Scientific communities in areas such as highenergy physics, gravitational-wave physics, geophysics, astronomy and bioinformatics, are utilizing Grids to share, manage and process large data sets.
Velammal College of Engineering and Technology, Madurai
B.
TASK SCHEDULING:
The need for task scheduling [1-2] arises from the requirement for most modern systems to perform multitasking (executing more than one process at a time) and multiplexing (transmit multiple flows simultaneously). Task scheduling is one of the NP-complete problems [5]. Heuristic optimization algorithm is widely used to solve a variety of NP-Complete problems. With the development of the network technology, grid computing used to solve larger scale complex problems becomes a focus technology. Task scheduling is a challenging problem in grid computing environment. If large numbers of tasks are computed on the geographically distributed resources, a reasonable scheduling algorithm must be adopted in order to get the minimum completion time. So task scheduling which is one of NPComplete problems become a focus by many of scholars in grid computing area. Abraham et al and Braun et al [6] presented three basic heuristics implied by Nature for Grid scheduling, namely Genetic Algorithm (GA) [3-5], Simulated Annealing (SA) [5] and Tabu Search (TS) [11], and heuristics derived by a combination of these three algorithms. GA and SA are powerful stochastic optimization methods, which are inspired form the nature.PSO [7-10] shares many similarities with evolutionary computational techniques such as Genetic Algorithms. The system is initialized with a population of random solutions and searches for optima by updating generations. However unlike GA, PSO has no evolution operators such as cross over and mutation. In PSO, the potential solutions, called particles, fly through the problem space by following the current optimum particles. This paper is organized as follows. In section 2, the mostly used task scheduling algorithms in grid computing. In section 3, our proposed system Particle Swarm Optimization algorithm is explained. Its implementation is also discussed in this section. The section 4 and 5 deals with concluding discussions and references respectively.
Page 50
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
II.
MOSTLY USED TASK SCHEDULING ALGORITHMS IN GRID COMPUTING:
In this paper, we will discuss the problem that ‘numt’ tasks work on ‘res’ computing resources with an objective of minimizing the completion time and utilizing the resources effectively. If the number of tasks is less than the number of resources in grid environment, the tasks can be allocated on the resources according to the first- come-first-serve rule. If the number of task is more than the number of resources, the allocation of tasks is to be made by some scheduling schemes. A.
SIMULATED ANNEALING (SA):
Simulated Annealing (SA) derives from the Monte Carlo method for statistically searching the global. The concept is originally from the way in which crystalline structures can be formed into a more ordered state by use of the annealing process, which repeats the heating and slowly cooling a structure [5]. SA has been used to select a suitable size of a set of machines for scheduling in a Grid environment.
Genetic algorithms (GAs) provide robust search techniques that allow a high-quality solution to be derived from a large search space in polynomial time by applying the principle of evolution [5]. A genetic algorithm combines exploitation of best solutions from past searches with the exploration of new regions of the solution space. Any solution in the search space of the problem is represented by an individual. A genetic algorithm maintains a population of individuals that evolves over generations. The quality of an individual in the population is determined by a fitness function. The fitness value indicates how good the individual is compared to others in the population. It first creates an initial population consisting of randomly generated solutions. After applying genetic operators, namely selection, crossover and mutation, one after the other, new offspring are generated. Then the evaluation of the fitness of each individual in the population is conducted. The fittest individuals are selected to be carried over next generation. The above steps are repeated until the termination condition is satisfied. Typically, a GA is terminated after a certain number of iterations, or if a certain level of fitness value has been reached.
Figure 2. Flowchart – Genetic Algorithm
III. A. PARTICLE ALGORITHM: Figure 1. Flowchart – Simulated Annealing algorithm
B.
GENETIC ALGORITHMS (GA):
Velammal College of Engineering and Technology, Madurai
OUR PROPOSED WORK: SWARM
OPTIMIZATION
PSO [7] is a robust stochastic optimization technique based on the movement and intelligence of swarms. PSO applies
Page 51
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the concept of social interaction to problem solving. It was developed in 1995 by James Kennedy (social-psychologist) and Russell Eberhart (electrical engineer). In PSO each single solution is a ‘bird’ in the search space. We call it “particle”. All of particles have fitness values which are evaluated by the fitness function to be optimized and have velocities which direct the flying of particles. Each particle keeps track of its coordinates in the solution space which are associated with the best solution (fitness) that has achieved so far by that particle. This value is called personal best, pbest. Another best value that is tracked by the PSO is the best value obtained so far by any particle in the neighborhood of that particle. This value is called gbest. The basic concept of PSO lies in accelerating each particle toward its pbest and the gbest locations, with a random weighted acceleration at each time
w: weighting function, cj : weighting factor, rand : uniformly distributed random number between 0 and 1, sik : current position of agent i at iteration k, pbesti : pbest of agent i, gbest: gbest of the group. The following weighting function is usually utilized in (1) w=wMax-[(wMax-wMin)xiter]/maxIter (2) Where wMax= initial weight, wMin = final weight, maxIter = maximum iteration number, iter = current iteration number. sik+1 = sik + Vik+1
(3)
3.2 FLOW CHART DEPICTING THE GENERAL PSO ALGORITHM:
Figure 3.Basic concept of PSO
sk : current searching point. sk+1: modified searching point. vk: current velocity. vk+1: modified velocity. vpbest : velocity based on pbest, vgbest : velocity based on gbest Each particle tries to modify its position using the following information: ) the current positions, ) the current velocities, ) the distance between the current position and pbest, ) The distance between the current position and the gbest.
Figure 4. PSO-Flowchart
The modification of the particle’s position can be mathematically modeled according the following equation:
B.
Vik+1 = wVik +c1 rand1 (…) x (pbesti-sik) + c2 rand2 (…) x (gbest-sik) ….. (1)
Unlike in genetic algorithms, evolutionary programming and evolutionary strategies, in PSO, there is no selection operation. All particles in PSO are kept as members of the population through the course of the run PSO [8] are the only
where, vik : velocity of agent i at iteration k,
Velammal College of Engineering and Technology, Madurai
ADVANTAGES OF PSO OVER GA AND SA:
Page 52
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
algorithm that does not implement the survival of the fittest. No crossover operation in PSO. C.
IMPLEMENTATION:
Module 1: PSO Implementation The number of processors i.e., the number of resources and the number of jobs are obtained as input. From the obtained input, gridlets are generated. The PSO algorithm is used to allocate the jobs to correct resources so that the completion time of the process will get minimized. Module 2: Comparison with GA The implementation is made to run for certain trials and the output is obtained. It is then compared with that of the results obtained using Genetic Algorithm and the graph is drawn.
Figure 6. Example implementation – Generation of completion time
Consider the number of resources is 3 and the number of tasks is 10. The speed of three resources are 4,3 and 2 respectively. The results of GA algorithm running 10 times were {26, 25.4, 25.8, 25.8, 25, 25, 25.8, 26, 25.4, 25}, with an average value of 25.52.the best result based on PSO algorithm for (3, 10) is shown in table 1. Table 1. Example of the best result
Figure 5.Example implementation-Getting the user inputs
Task(T)/Resources(R) T1 T2 T3 T4 T5 T6 T7 T8 T9 T10
R1 0 1 1 0 0 1 0 1 0 0
R2 1 0 0 0 1 0 0 0 1 0
R3 1 0 0 1 0 0 1 0 0 1
The results of PSO algorithm running 10 times were {25, 25, 25.4, 25.4, 25, 25, 25, 25, 25, 25.2}, with an average value of 25.1. PSO algorithm provided the best result 7 times, while GA algorithm provided the best result 3 times.
Velammal College of Engineering and Technology, Madurai
Page 53
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
proposed approach and compared it with genetic algorithm under the same condition. From the simulated experiment, the result of PSO algorithm is better than GA. Simulation results demonstrate that PSO algorithm can get better effect for a large scale optimization problem. Task scheduling algorithm based on PSO algorithm can be applied in the computational grid environment. REFERENCES:
Figure 7: Performance of PSO and GA scheduling algorithm about 5 processors and 100 tasks
Figure 8: The performance of curves of different numbers of processors running different Number of tasks
It shows PSO usually spent the shorter time to complete the scheduling than GA algorithm. It is to be noted that PSO usually spent the shorter time to accomplish the various task scheduling tasks and had the better result compared with GA algorithm. IV.
[9] Foster and C. Kesselman (editors), The Grid: Blueprint for a Future Computing Infrastructure,Morgan Kaufman Publishers, USA, 1999. [10] Y. Gao, H.Q Rong and J.Z. Huang, Adaptive grid job scheduling with genetic algorithms, Future Generation Computer Systems, pp.1510-161 Elsevier,21(2005). [11] M. Aggarwal, R.D. Kent and A. Ngom, Genetic Algorithm Based Scheduler for Computational Grids, in Proc. of the 19th Annual International Symposium on High Performance Computing Systems and Application (HPCS’05), ,pp.209-215 Guelph, Ontario Canada, May 2005. [12] S. Song, Y. Kwok, and K. Hwang, Security-Driven Heuristics and A Fast Genetic Algorithm for Trusted Grid Job Scheduling, in Proc. of 19th IEEE International Parallel and Distributed Processing Symposium (IPDPS’05), pp.65-74, Denver, Colorado USA, April 2005. [13] Workflow scheduling algorithm for grid computing Jia Yu and Rajkumar Buyya Grid computing and distributed system(GRIDS) laboratory Department of Computer Science and Software Engineering The University of Melbourne,Australia. [14] Abraham, R. Buyya and B. Nath, Nature's Heuristics for Scheduling Jobs on Computational Grids, The 8th IEEE International Conference on Advanced Computing and Communications (ADCOM 2000), pp. 4552,Cochin, India, December 2000,. [15] Kennedy J. and Eberhart R. Swarm interllignece,Morgan Kaufmann, 2001. [16] J. Kennedy and R. C. Eberhard, “Particle swarm optimization”, Proc. of IEEE Int’l Conf. on Neural Networks, pp.1942-1948, Piscataway, NJ, USA, ,1995. [17] J.F. Schute and A.A. Groenwold, A study of global optimization using particle swarms, Journal of Global Optimization, pp.93-108, Kluwer Academic Publisher,31(2005). [18] M. Fatih Tasgetiren, Yun-Chia Liang, MehmetSevkli, and Gunes Gencyilmaz, “Particle Swarm Optimization and Differential Evolution for Single Machine Total Weighted Tardiness Problem,” International Journal of Production Research, pp. 4737-4754 , vol. 44, no. 22, 2006. [19] R. Braun, H. Siegel, N. Beck, L. Boloni, M. Maheswaran, A. Reuther, J. Robertson, M. Theys, B. Yao, D. Hensgen and R. Freund, A Comparison of Eleven Static Heuristics for Mapping a Class of Independent Tasks onto Heterogeneous Distributed Computing Systems, pp. 810-837, J. of Parallel andDistributed Computing, vol.61, No. 6, 2001.
CONCLUSION:
In this paper, scheduling algorithm based on PSO is proposed for task scheduling problem on computational grids. Each particle represents a feasible solution. The position vector is transformed from the continuous values to the discrete values based on SPV rules, accordingly, a permutation formed. Our approach is to generate an optimal schedule so as to complete the tasks in a minimum time as well as utilizing the resources in an efficient way. We evaluate the performance of our
Velammal College of Engineering and Technology, Madurai
Page 54
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
NTRU - Public Key Cryptosystem For Constrained Memory Devices V.Pushparani#1 and Kannan Balasubramaniam*2
#1
Faculty of Kamaraj College of Engg and Technology, Virudhunagar, Tamilnadu. 1
[email protected]
*2
Professor of Mepco Schlenk Engg College, Sivakasi. Tamilnadu. 2
[email protected]
ABSTRACT -- In many business sectors secure and efficient data transfer is essential. To ensure the security to the applications of business, the business sectors use Public Key cryptographic Systems (PKCS).. Secure public key authentication and digital signatures are increasingly important for electronic communications, commerce and security. They are required not only on high powered desktop computers, but also on smart cards and wireless devices with severely constrained memory and processing capabilities. An RSA and NTRU system generally belongs to the category of PKCS. The efficiency of a public key cryptographic system is mainly measured in computational overheads, key size. Although the security of RSA is beyond doubt, the evolution in computing power has caused a growth in the necessary key length. The fact that most chips on smart cards cannot process keys extending 1024 bit shows that there is a need for alternative. NTRU is such an alternative and it is a collection of mathematical algorithms based on manipulating lists of very small integers and polynomials. NTRU is the secure public key cryptosystem not based on factorization or discrete logarithm problems. The security of the NTRU cryptosystem came from the interaction of the polynomial mixing system with the independence of reduction modulo two relatively prime integers’ p and q. KeyWords -- NTRU, patterns, encryption, decryption, digital signature
I. INTRODUCTION A. NTRU NTRU (Number Theory Research Unit) is relatively new and was conceived by Jeffrey Hoff stein, Jill Pipher and Joseph. H. Silverman. NTRU uses polynomial algebra combined with clustering principle based on elementary mathematical theory. The security of NTRU comes from the interaction of polynomial mixing system with the independence of reduction modulo two relatively prime numbers. The basic collection of objects used by the NTRU Public Key Cryptosystem in the ring R that consists of all truncated polynomials of degree N-1 having integer coefficients a =a 2 + a 3X 3 +… + a N-2X N-2 + a N-1X N-1. 0+ a 1X + a 2X Polynomials are added in the usual way. They are also
Velammal College of Engineering and Technology, Madurai
multiplied more or less as usual, except that XN is replaced by 1, XN+1 is replaced by X, X N+2 is replaced by X2 and so on. 1)
Truncated Polynomial Rings:
The NTRU public-key algorithm, explained in chapter II, uses random polynomials which are generated from a [X]/(XN-1). The polynomial ring of the form R[X] = polynomials that form the ring R[X] have a degree smaller than N. The polynomials in the truncated ring R[X] are added in a regular way by adding their coefficients. The polynomial XN ≡ 1. Said differently, the maximum degree of the resultant polynomial of a multiplication between two polynomials of the ring cannot be greater than N - 1. The product operation of two polynomials in R[X], is defined as c(X) = a(X) * b(X) where ck is the kth coefficient of c(X) and is computed as ck=a0bk + aibk-1 + akb0 +.... + aN-1bk+1 The product of polynomials in R[X] is also called the star multiplication. II. RELATED WORKS A. NTRU: A Ring Based Public Key Cryptosystem - J. Hoffstein, J. Pipher, and J. H. Silverman N −1
The polynomial is represented as a(X)=
∑
aiXi € R. When
i =0
NTRU was formally introduced in 1998, Silverman presented the polynomial multiplication as the cyclic convolution of two polynomials as ck =
k
∑
i= 0
a i .b k − i +
N −1
∑
a i .b N
i= k +1
+ k −i
=
∑
i + j = k (mod
a i .b
j
N )
Ultimately, this straightforward method requires N2 multiplications to perform a polynomial multiplication for NTRU. B. NTRU in Constrained Devices – D. V. Bailey, D. Coffin, A. Elbirt, J. H. Silverman, and A. D. Woodbury
Page 55
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
In 2001, Bailey et al. introduced a fast convolution algorithm, to perform a polynomial multiplication for NTRU. This algorithm makes the realization that almost every polynomial multiplication involved with NTRU has one polynomial that is random. The random polynomials are assumed to have binary coefficients. In addition, the random polynomial is assumed to consist of three smaller polynomials of low Hamming weight(d), The product can be computed using only additions over and rotations of the coefficient vector of a. For each non-zero coefficient bi of b, polynomials of the form Xia are added in order to compute the product ab. The multiplication of a polynomial a with a monomial Xi in R corresponds to i right rotations of the coefficient vector of a, where the right rotation is defined as the mapping
( a 0 ,..., a N −1 ) a ( a N −1 , a 0 ,..., a N − 2 )
Fig.1 Multiplications of a, b using additions and rotations
The fast convolution algorithm reduces the complexity of NTRU's polynomial multiplication to dN additions and no multiplications. C. Sliding Window Method for NTRU - M.-K. Lee1, J. W. Kim, J. E. Song, and K. Park. In 2007, Lee et al observed that it is possible to reduce the number of additions needed to compute the product of polynomial a, b by using bit patterns of the binary polynomial b. By a bit pattern, we understand two 1s separated by a (possibly empty) sequence of 0s. We say that such a bit pattern has length l if the two 1s are separated by l - 1 0s. More generally, it is possible to reduce the number of additions needed to compute the product ab whenever a bit (a0 ,..., aN−1more ) a (than aN−1 ,once a0 ,..., pattern occurs inab. N −2It) is thus desirable to choose bit patterns in a way that maximizes the number of pattern occurrences and to efficiently identify the patterns in b. The algorithm of Lee et al. only considers bit patterns of length less than or equal to a parameter w. For each pattern l length l = 1…..w, the polynomial a + X a is precomputed and stored in a lookup table. The non-zero coefficients not belonging to any such bit pattern are treated as in the algorithm of Bailey et al. Binary polynomials are represented as bit strings.
Velammal College of Engineering and Technology, Madurai
Fig.2 Multiplications of a, b using bit patterns
Considering bit strings containing more than two 1s does not achieve any notable speedup because the probability that these strings occur more than once in b is very low. D. An Algorithm of Dynamic Patterns for NTRU - Bu Shan Yue Zhang Han Yan Wang RuChuan Bu Shan Yue et al proposed algorithm uses bit patterns, but the patterns can be of arbitrary length, and only the patterns actually occurring in b are considered. Thus, all non-zero coefficients of b belong to a pattern, except for a single coefficient in case that the hamming weight of b is odd. They omit the precomputation step of the algorithm of Lee et al. and instead compute the polynomials a+Xla when needed. They also represent binary polynomials as the sequence of the degrees of their monomials. It shows that pattern finding can be performed much easier and faster in this representation. E. NSS: The NTRU Signature Scheme - Jeffrey Hoffstein, Jill Pipher, Joseph H. Silverman Secure public key authentication and digital signatures are increasingly important for electronic communications and commerce, and they are required not only on high powered desktop computers, but also on Smart Cards and wireless devices with severely constrained memory and processing capabilities. Silverman et al introduced a NTRU signature scheme in which the public and private keys are formed as follows. Choosing two polynomials f and g having the form f = f0 + pf1 and g = g0 + pg1. Here f0 and g0 are fixed universal polynomials (e.g., f0 = 1 and g0 = 1-X) and f1 and g1 are polynomials with small coefficients chosen from the sets Ff and Fg, respectively. Next computes the inverse f-1 of f modulo q, that is, f-1 satisfies f-1* f ≡ 1 (mod q). Public verification key is the polynomial h * f-1 ≡g (mod q). Private signing key is the pair (f; g). III. NTRU PUBLIC KEY CRYPTOSYSTEM A. Parameters The basic collection of objects used by the NTRU Public Key Cryptosystem in the ring R that consists of all truncated polynomials of degree N-1 having integer coefficients
Page 56
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
a =a 0+ a 1X + a 2X 2 + a 3X 3 +…...+a N-2X N-2 +a N-1X N-1. All computations are performed in the ring of convolution modular polynomials R = Z [ X ] / ( X N − 1 ) where polynomials of degree less than N are used as representatives for the residue classes. Polynomials are added in the usual way. They are also multiplied more or less as usual, except that XN is replaced by 1, XN+1 is replaced by X, X N+2 is replaced by X2 and so on. The NTRU public-key algorithm is defined by the following parameters: • N: The degree parameter. Defines the degree N - 1 of the polynomials in R. • q: Large modulo. Polynomial coefficients are reduced modulo q. • p: Small modulo. The coefficients of the message are reduced modulo p in decryption. • dF: Private key space. Fixes the polynomial form defining the number of positive ones for the private key f, the negative ones are fixed by df - 1. • dg: Public key space. Fixes the polynomial form defining the number of positive and negative ones for the random polynomial g used to calculate the public key. • dr: Blinding value space. Fixes the polynomial form defining the number of positive and negative ones of the random polynomial r used in the encryption process. • D(d): The set of binary polynomials of degree less than N with hamming weight d. B. Key generation 1) Pattern Finding: A binary polynomial b of hamming weight d is represented by the sequence D0....Dd-1 of the degrees of its monomials in ascending order. The polynomial is traversed once in reverse order, starting at Dd-1. For each possible pattern length l € 1...N - d+1 a list Ll of pattern locations is created. Every pair of degrees (Di....Di-1) represents a bit pattern of length Di – Di1. The degree Di is stored in the list LDi....Di-1 and i is decreased by 2. In case that d is odd, the remaining single degree D0 is stored separately in a list L0. 2) Pattern multiplication: Each non-empty list Ll, , l > 0 represents a bit pattern of b with length l. For each such Ll, the corresponding pattern polynomial p = a + X l a is computed. For each element D of the list Ll, this pattern polynomial is right rotated D times and added to the resulting polynomial. A possibly remaining single degree stored in L0 is treated separately without computing a pattern polynomial.
During Key Generation the process Random Polynomial is invoked to generate the polynomial f and g in order to calculate the public key h. The polynomials are generated with random coefficients from a truncated ring of polynomials R. Random Polynomial receives the number of positive and negative ones and generates the random polynomial of N coefficients. 4) Inversion modulo q: During Key Generation the polynomial fq, the inverse of f modulo q, is also computed. The polynomial fq is necessary, together with g, to calculate the public key h. Inversion modulo q computes the inverse of a certain polynomial f modulo q in fq, computing fq = f-1 (mod q) which satisfies f* fq ≡ 1 (mod q). 5) Inversion modulo p: Inversion modulo p computes the inverse of a polynomial f in modulo p such that f * fp ≡1(mod p). 6) Key generation: Choose uniformly
at
random
and polynomials f . f −1 ≡ 1(mod q ) pF. If the congruence
the
binary
Compute f = 1 + has a solution,
calculate such a solution Otherwise, start over. = f − 1 pg mod q Compute the hpolynomial . The notation a = b mod q stands for reducing the coefficients of b modulo q and assigning the result to a. The private key is f, the public key is h.
Fig. 3 Key Generation
3) Random Polynomial generation:
Velammal College of Engineering and Technology, Madurai
Page 57
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
C. Encryption Encryption is the simplest part in the NTRU PKC. Encryption only requires generating a random polynomial r from the ring R that obscures the message. Then the polynomial r is multiplied by the public key h. And finally the product of r and h is added to the desired message to encrypt. This means Encryption just needs to receive a message in the polynomial form m and the public key h.
Dev(a; b) = #{i: Ai ≠ Bi}: Intuitively, Dev(a; b) is the number of coefficients of a mod q and b mod q that differ modulo p. 1) Key Generation: Chooses two polynomials f and g. Compute the inverse f-1 of f modulo q. Public verification key is the polynomial h * f-1 ≡g (mod q) and his private signing key is the pair (f; g). 2) Signing: User’s document is a polynomial m modulo p. Chooses a polynomial w € Fw of the form w = m + w1 + pw2; where w1 and w2 are small polynomials. Then computes s≡ f* w (mod q): Signed message is the pair (m; s).
Fig.4 Encryption specification D. Decryption
The Decryption process requires the encrypted message e and the private key set (f.. fp) to decrypt the encrypted message e into the clear message c.
3) Verification: In order to verify signature s on the message m, First checks that s≠0 and then verifies the following two conditions: 1. Compares s to f0*m by checking if their deviation satisfies Dmin≤ Dev(s. f0 * m)≤Dmax 2. Use public verification key h to compute the polynomial t≡h*s (mod q), putting the coefficients of t into the range [-q=2; q=2] as usual. Then checks if the deviation of t from g0*m satisfies Dmin≤ Dev(t. g0 * m)≤Dmax If signature passes tests (A) and (B), then accepts it as valid. IV. PERFORMANCE ANALYSIS In order to grasp how well NTRU performs for different applications, a timing analysis was conducted for the Key generation, Encryption, and Decryption functions. The test values for the parameters of NTRU used for this performance analysis are listed in Table ITABLE I TEST VALUES USED for PERFORMANCE ANALYSIS
Parameters
Fig.5 Decryption specification
E. The NTRU Signature Scheme The key computation involves the deviation between two polynomials. Let a(X) and b(X) be two polynomials in R. We first reduce their coefficients modulo q to lie in the range between -q=2 to q=2, then we reduce their coefficients modulo p to lie in the range between -p=2 and p=2. Let A(X) = A0+A1X+.....+AN-1XN-1 and B(X) = B0+.......+BN-1X N-1 be these reduced polynomials. Then the deviation of a and b is
Velammal College of Engineering and Technology, Madurai
N q p NumOnes f NumNegOnes f NumOnes g NumNegOnes g NumOnes r NumNegOnes r NumOnes m NumNegOnes m
107 NTRU 107 64 3 15 14 12 12 5 5 25 25
503 NTRU 503 256 3 216 215 72 72 55 55 165 165
Page 58
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The results of the timing analysis are shown in Table II. TABLE II TIMINGS for NTRU
Functions in NTRU Key Generation (ms) Encryption (ms) Decryption (ms)
107 NTRU
503 NTRU
16.2
699.5
0.6
15.0
1.4
29.4
Systems (CHES ’01), volume 2162 of Lecture Notes in Computer Science, pages 262–272. Springer Verlag. [3]. J. Hoffstein, J. Pipher, and J. H. Silverman. NTRU: A Ring-Based Public Key Cryptosystem. 1998, In Proceedings of the Third International Symposium on Algorithmic Number Theory, volume 1423 of Lecture Notes in Computer Science, pages 267–288.Springer Verlag. [4]. Jeffrey Hoffstein, Jill Pipher, Joseph H. Silverman,2005,NSS: The NTRU Signature Scheme. [5]. R. Lindner, J. Buchmann, M. Doering, 2008, Efficiency Improvements for NTRU, Sicherheit 2008, LNI vol. 128. [6]. M.-K. Lee1, J. W. Kim, J. E. Song, and K. Park. 2007, Sliding Window Method for NTRU. In Proceedings of ACNS 2007, volume 4521 of Lecture Notes in Computer Science, pages 432–442. Springer Verlag. [7]. Narasimham Challa and Jayaram Pradhan, 2008, Performance Analysis and Public key cryptographic systems RSA and NTRU, IJCSNS International Journal of computer science and Network security.
The Key Generation function takes the longest time because it requires two polynomial inversions and a polynomial multiplication. In addition, since the Decryption function requires two polynomial multiplications, it takes over two times as long as the Encryption functions. Since Encryption requires only one polynomial multiplication, it is fair to use Encryption's timing to estimate the time to perform a single polynomial multiplication. Altogether, the timing analysis in Table II shows that NTRU has potential in offering high performance. V. CONCLUSION The NTRU is suitable for applications where it requires security based on the environment. Because of the RSA’s time complexity O (n 3) and that of NTRU’s O (n log (n)), NTRU confirms its cryptography and delivers encryption, decryption and authentication at speeds of multiple times faster than RSA. NTRU is ideally suited for applications where high performance, high security and low power consumption are required. ACKNOWLEDGMENT The authors are grateful to the management of Kamaraj College of Engineering and Technology, Virudhunagar, India and Mepco Schlenk Engineering College, Sivakasi, India for granting permission to undertake this work. Our thanks are due to the Head of the Department of Computer Science and Engineering of Kamaraj College of Engineering and Technology for allowing us the use of the laboratories and computing facilities. REFERENCES [1]. Bu Shan Yue Zhang Han Yan Wang RuChuan Dept. of Comput. Eng., Huaiyin Inst. of Technol., Huaiyin; 2008, An Algorithm of Dynamic Patterns for NTRU Wireless communications, Networking and Mobile computing,. WiCOM ’08. [2]. D. V. Bailey, D. Coffin, A. Elbirt, J. H. Silverman, and A. D. Woodbury. 2001, NTRU in Constrained Devices. In Proceedings of the Third International Workshop on Cryptographic Hardware and Embedded
Velammal College of Engineering and Technology, Madurai
Page 59
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Novel Randomized Key Multimedia Encryption Algorithm Security Against Several Attacks S. Arul Jothi Department of Computer Science, NMS Sermathai Vasan College for Women Madurai, Tamilnadu, India [email protected] Abstract— Encryption is the science of changing data so that it is unrecognisable and useless to an unauthorised person. Decryption is changing it back to its original form. The wide availability of digital multimedia contents as well as the accelerated growth of wired and wireless communication technologies have brought the multimedia content security issue to the forefront. In particular, the problem of efficient multimedia data encryption has recently gained more attention in both academia and industry. Unlike ordinary computer applications, multimedia applications generate large amounts of data that has to be processed in real time. So, a number of encryption schemes for multimedia applications have been proposed in recent years. The most secure techniques use a mathematical algorithm and a variable value known as a 'key'. As the entire operation is dependent upon the security of the keys, it is sometimes appropriate to devise a fairly complex mechanism to manage them. With the rapid progress of information technology, security becomes one of the key factors in information storage, communication and processing. For the reason of speed and security I develop a novel encryption algorithm secure to several attacks. As opposed to traditional key algorithms here the key is randomized. Keywords— Cryptanalysis, Algorithm, Cryptography, Block cipher, Security, Encryption and Decryption.
I. INTRODUCTION Information security deals with several different "trust" aspects of information. Another common term is information assurance. Information security is not confined to computer systems, nor to information in an electronic or machine-readable form. It applies to all aspects of safeguarding or protecting information or data, in whatever form. Cryptography [3] [10]can also be defined as the science and art of manipulating message to make them secure. In this the original message to be transformed is called the plaintext and resulting message after transformation is called the cipher text. There are several ways of classifying cryptographic algorithms. They will be categorized based on the number of keys that are employed for encryption and decryption, and further defined by their application and use. The three types of algorithms that will be discussed are
Velammal College of Engineering and Technology, Madurai
secret key cryptography, public key cryptography and hash functions. With secret key cryptography[1] [7], a single key is used for both encryption and decryption. As shown in figure below, the sender uses the key to encrypt the plaintext and sends the cipher text to the receiver. The receiver applies the same key to decrypt the message and recover the plaintext. Because a single key is used for both functions, secret key cryptography is also called symmetric encryption Public-key cryptography is based on the notion that encryption keys are related pairs, private and public. The private key remains concealed by the key owner; the public key is freely disseminated to various partners. Data encrypted using the public key can be decrypted only by using the associated private key and vice versa. Because the key used to encrypt plaintext is different from the key used to decrypt the corresponding cipher text, public-key cryptography is also known as asymmetric cryptography. Cryptanalysis [21] is the study of methods for obtaining the meaning of encrypted information, without access to the secret information which is normally required to do so. Typically, this involves finding the secret key. In non-technical language, this is the practice of code breaking or cracking the code, although these phrases also have a specialised technical meaning. Block ciphers [13] [14] encrypt blocks of data (typically 64 or 128 bits) in a fixed key-dependent way. The design of block ciphers is a well-studied area of research. The best-known block ciphers are the Data Encryption Standard (DES) and the Advanced Encryption Standard (AES). In the past decade, many new attacks on block ciphers have emerged, the most important ones being differential and linear cryptanalysis. Differential cryptanalysis is an example of a chosen-plaintext attack, while linear cryptanalysis is a known-plaintext attack. A good design should at least be resistant to these attacks. This algorithm can overcome this two attacks. A randomized algorithm or probabilistic algorithm is an algorithm which employs a degree of randomness as part of its logic. The algorithm typically uses uniformly random bits as an auxiliary input to guide its behaviour, in the hope of
Page 60
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
achieving good performance in the "average case" over all possible choices of random bits [4] [5]. Formally, the algorithm's performance will be a random variable determined by the random bits; thus either the running time, or the output (or both) are random variables. Here in this algorithm the key is nothing but the random positions of the given n × n matrix.. II. BRIEF OVERVIEW OF CRYPTOGRAPHY The history of cryptography [3]begins thousands of years ago. Until recent decades, it has been the story of what might be called classic cryptography that is, of methods of encryption that use pen and paper, or perhaps simple mechanical aids. In the early 20th century, the invention of complex mechanical and electromechanical machines, such as the Enigma rotor machine, provided more sophisticated and efficient means of encryption; and the subsequent introduction of electronics and computing has allowed elaborate schemes of still greater complexity, most of which are entirely unsuited to pen and paper. The development of cryptography has been paralleled by the development of cryptanalysis of the "breaking" of codes and ciphers. The discovery and application, early on, of frequency analysis to the reading of encrypted communications has on occasion altered the course of history. Thus the Zimmermann Telegram triggered the United States' entry into World War I; and Allied reading of Nazi Germany's ciphers shortened World War II, in some evaluations by as much as two years. The Germans made heavy use (in several variants) of an electromechanical rotor based cipher system known as Enigma. The German military also deployed several mechanical attempts at a one time pad. Bletchley Park called them the Fish cipher's, and Max Newman and colleagues designed and deployed the world's first programmable digital electronic computer, the Colossus, to help with their cryptanalysis. The German Foreign Office began to use the one-time pad in 1919; some of this traffic was read in WWII partly as the result of recovery of some key material in South America that was insufficiently carefully discarded by a German courier. The era of modern cryptography [7] really begins with Claude Shannon, arguably the father of mathematical cryptography. In 1949 he published the paper Communication Theory of Secrecy Systems in the Bell System Technical Journal, and a little later the book Mathematical Theory of Communication with Warren Weaver. These, in addition to his other works on information and communication theory established a solid theoretical basis for cryptography and for cryptanalysis. And with that, cryptography more or less disappeared into secret government communications organizations such as
Velammal College of Engineering and Technology, Madurai
the NSA. Very little work was again made public until the mid '70s, when everything changed. In 1969[6] two major public (ie, non-secret) advances. First was the DES (Data Encryption Standard) submitted Cryptography/History by IBM , at the invitation of the National Bureau of Standards (now NIST), in an effort to develop secure electronic communication facilities for businesses such as banks and other large financial organizations. After 'advice' and modification by the NSA, it was adopted and published as a FIPS Publication (Federal Information Processing Standard) in 1977. It has been made effectively obsolete by the adoption in 2001 of the Advanced Encryption Standard, also a NIST competition, as FIPS 197. DES was the first publicly accessible cipher algorithm to be 'blessed' by a national crypto agency such as NSA. The release of its design details by NBS stimulated an explosion of public and academic interest in cryptography. DES [19], and more secure variants of it, are still used today, although DES was officially supplanted by AES (Advanced Encryption Standard)[18] in 2001 when NIST announced the selection of Rijndael, by two Belgian cryptographers. DES remains in wide use nonetheless, having been incorporated into many national and organizational standards. However, its 56-bit key-size has been shown to be insufficient to guard against brute-force attacks (one such attack, undertaken by cyber civil-rights group The Electronic Frontier Foundation, succeeded in 56 hours the story is in Cracking DES, published by O'Reilly and Associates). As a result, use of straight DES encryption is now without doubt insecure for use in new crypto system designs, and messages protected by older crypto systems using DES[19] should also be regarded as insecure. The DES key size (56-bits) was thought to be too small by some even in 1976, perhaps most publicly Whitfield Diffie. There was suspicion that government organizations even then had sufficient computing power to break DES messages and that there may be a back door due to the lack of randomness in the 'S' boxes. Second was the publication of the paper New Directions in Cryptography by Whitfield Diffie and Martin Hellman. This paper introduced a radically new method of distributing cryptographic keys, which went far toward solving one of the fundamental problems of cryptography [8], key distribution. It has become known as Diffie-Hellman key exchange. The article also stimulated the almost immediate public development of a new class of enciphering algorithms, the asymmetric key algorithms. In contrast, with asymmetric key encryption, there is a pair of mathematically related keys for the algorithm, one of which is used for encryption and the other for decryption. Some, but not all, of these algorithms have the additional property that one of the keys may be made public since the other cannot be (by any currently known method) deduced from the 'public'
Page 61
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
key. The other key in these systems is kept secret and is usually called, somewhat confusingly, the 'private' key[12]. An algorithm of this kind is known as a public key / private key algorithm, although the term asymmetric key cryptography is preferred by those who wish to avoid the ambiguity of using that term for all such algorithms, and to stress that there are two distinct keys with different secrecy requirements. III. SECURITY ANALYSIS In this section we discuss the security strength of the randomized encryption paradigm under five most common cryptographic attack types: Brute force attack, cipher textonly attack, known plaintext attack, chosen plaintext attack and chosen cipher text attack. G. Brute Force attack In cryptography, a brute force attack[15] is a strategy used to break the encryption of data. It involves traversing the search space of possible keys until the correct key is found. The selection of an appropriate key length depends on the practical feasibility of performing a brute force attack. By obfuscating the data to be encoded, brute force attacks are made less effective as it is more difficult to determine when one has succeeded in breaking the code. In my algorithm, If n is large, then the key space is of n2 dimension and it is impossible to search the key space. H. Cipher text only A known cipher text attack[16] is an attack where the cryptanalyst only has access to encrypted cipher text. A known cipher text attack is the easiest of the common cryptanalysis attacks to mount, because is requires the least amount of control over the encryption device. Conversely, the known cipher text is the most difficult of the common methods of cryptanalysis to execute successfully, because so little knowledge is known to begin with. Let the cipher text c be available. Since the key stores the bit positions for both encryption and decryption, it is not possible to obtain the key from the cipher text and the encryption /decryption process. I. Known plain text attack The known-plaintext attack (KPA)[20] is an attack model for cryptanalysis where the attacker has samples of both the plaintext and its encrypted version (cipher text) and is at liberty to make use of them to reveal further secret information such as secret keys and code books. Suppose a plain text_ cipher text pair (p, c) is available. The ijth element of c, cij is obtained by the form
Velammal College of Engineering and Technology, Madurai
⎛n l ⎞ cij= ⎜ ∑ cil ∑cmj-pij⎟ %2 --- > 1 ⎝i=1 j=1 ⎠ where cil, cmj values are defined based on the of (i, l) and (m, j) in the key. Since cil and cmj can take either 0 or 1, eqn1 involves 22n-2 possibilities. Hence, it is impossible to solve. Therefore we cannot find the key k. J. Chosen-plaintext attack A chosen-plaintext attack (CPA) [11] is an attack model for cryptanalysis which presumes that the attacker has the capability to choose arbitrary plaintexts to be encrypted and obtain the corresponding cipher texts. The goal of the attack is to gain some further information which reduces the security of the encryption scheme. In the worst case, a chosen-plaintext attack could reveal the scheme's secret key. Suppose we take a plaintext p as a n × n block of zeros then the encryption process yields the cipher text c as (n × n) block of zeros. If we take a plain text block p with 0 entry except one, say(i, j) only equal to 1. Then the encryption process spreads the value randomly (as per the key order) in the cipher block c. from the pair (p-c) it is impossible to obtain the key k. K. Chosen-cipher text attack A chosen-cipher text attack (CCA)[2] is an attack model for cryptanalysis in which the cryptanalyst gathers information, at least in part, by choosing a cipher text and obtaining its decryption under an unknown key. Suppose a cipher text c is choosen. The decryption process uses the key in the reverse order and yields the plain text. The key spreads zeros & ones randomly in the plain text according to the position order. Infact, there would be unbalanced number of zeros and ones, as compared to the cipher text. Therefore impossible to find the key. IV ALGORITHM Security is a major concern in an increasingly multimediadefined universe where the internet serves as an indispensable resource for information and entertainment. This algorithm protect and provide access to critical and time-sensitive copyrighted material or personal information. L. Step 1: Randomized key generation Generate a pair of values (i,j) randomly, 1<=i<=m, 1<=j<=m, such that no pair (i,j) is repeated more than once.
Let K denote the collection of pairs (i,j) stored as per the order of generation. M. Step 2: Encryption Let p be a plaintext of size n. Let a = n(mod m²).
Page 62
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
If a = 0 set k = n/m² else set k = n/m² + 1. The plaintext p is divided in to k blocks, say p1,p2,---,pk, such that the first m entries of p form the first row of p1, second m elements the second row of p1 and so on. Similarly we define p2,p3,---,pk. If a ≠ 0 then we add the require number of zeros at the end of the plaintext, so that the last block pk is of order m × m. For l = 1,…,k. Let aij be the (i,j)th element of pl. Change aij as per the generation order of the pair(i,j) as follows, n ⎛ n ⎞ a^ij = ⎜ ∑arj+ ∑ais -aij ⎟ ⎝ r=1 s=1 ⎠ If a^ij (mod 2) = 0, set aij= 1 else set aij = 0. Let cl be the m × m matrix with the modified aij. c1 is the ciphertext corresponding to p1. N. Step 3: Decryption For l = 1,---,k. Let aij be the (ij)th element of cl. Change aij as per the reverse order of generation of the pair (i,j) as follows, n ⎛ n ⎞ a^ij = ⎜ ∑arj + ∑ais -aij ⎟ ⎝ r=1 s=1 ⎠ If a^ij (mod 2) = 0, set aij = 1 else set aij = 0. The (aij) = pl, the original plaintext block. The plaintext p = p1,p2,---,pk.
IV. EXPERIMENTS AND PERFORMANCE EVALUATION With the rapid development of communication techniques, it becomes more and more practical to talk with anyone anywhere. As an important aspect, multimedia (image, video, audio, etc.) enriches humans’ daily life. Nowadays, multimedia communication is in close relation with the activities in entertainment, politics, economics, militaries, industries, etc., which makes it urgent for protecting multimedia security, e.g., confidentiality, integrity, ownership or identity. Generally, multimedia security is different from text/binary data security since multimedia content is often of large volumes, with interactive operations, and requires real-time responses. As a result, the last decade has witnessed an explosive advance in multimedia applications that are dramatically changing every aspect of our modern life, from business, manufacturing, transportation, medicine and finance to education, research, entertainment and government. Although encrypting the entire multimedia content by a
Velammal College of Engineering and Technology, Madurai
traditional cryptographic cipher (e.g., the block or stream cipher) yields a satisfactory level of security, such an approach does have several shortcomings. First, the computational cost associated with encrypting the entire multimedia content is often high due to the large data size. Second, the encryption and decryption operations add another level of complexity to the system. In most cases, additional hardware or software functions are needed in order to implement it. Hence, I have developed an efficient yet secure multimedia encryption technique. The algorithm uses the experimental environment, CPU: Intel(R) core(TM)2 DVOE7200 @ 2.53GHz, 0.99 GB of RAM; Operating System: Windows XP Professional. O. Results: Block size : 8 bits COMPARISON OF DIFFERENT FILE FORMATS S. NO.
FILE TYPE & SIZE
TIME FOR ENCRYPTION
TIME FOR DECRYPTION
1.
Txt,1kb
78ms
63ms
2.
Doc,551kb
1secs 360 ms
1secs 47 ms
3.
Bmp,14.3kb
141 ms
143 ms
4.
Jpeg,1.69mb
3 secs 469 ms
3 secs 266ms
5.
Pdf,245kb
562 ms
531 ms
6.
Xls,29kb
141 ms
125 ms
7.
Mp3,50.3kb
171 ms
128 ms
8.
Wav,45.3mb
36 secs
35 secs
9.
Vob,62.8mb
47 secs
46 secs
Recent advances in the field of video communications have emphasized the importance of delivering video contents to a wide variety of devices, through different networks, and using a growing number of codecs and standards. In addition to the performance challenges behind video communications, the security issues have received increasing attention from the research community. Among the most relevant security aspects of contemporary video communication systems are: encryption, watermarking and authentication. File format: video file File size:62.8mb
Page 63
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
COMPARISON OF TIME FOR DIFFERENT BLOCK SIZES OF A VIDEO FILE BLOCK SIZE
TIME FOR
TIME FOR
(BITS)
ENCRYPTION
DECRYPTION
8
47 secs
46 secs
16
1 min 10 secs
1 min 9 secs
32
1 min 31 secs
1 min 30 secs
64
2 min 35 secs
2 min 34 secs
V. CONCLUSION Since the development of cryptology in the industrial and academic worlds in the seventies, public knowledge and expertise have grown in a tremendous way, notably because of the increasing presence of electronic communication means in our lives. Block ciphers are inevitable building blocks of the security of various electronic systems. Recently, many advances have been published in the field of public-key cryptography, being in the understanding of involved security models or in the mathematical security proofs applied to precise cryptosystems. Unfortunately, this is still not the case in the world of symmetric-key cryptography and the current state of knowledge is far from reaching such a goal. In this paper we developed a novel dynamic symmetric key generation scheme for multimedia data encryption. VI. REFERENCES [1] M. Bellare and P. Rogaway, “Robust computational secret sharing and a unified account of classical secret-sharing goals”, Proceedings of the 14th ACM Conference on Computer and Communications Security (CCS), ACM, 2007. [2] J. Zhou, Z. Liang, Y. Chen, and O. C. Au, “Security analysis of multimedia encryption schemes based on multiple Huffman table”, IEEE Signal Processing Letters, vol. 14, no. 3, pp. 201– 204, 2007. [3] M. Bellare, A. Boldyreva and A. O'Neill, “Deterministic and efficiently searchable encryption”, Advances in Cryptology - Crypto 2007 Proceedings, Lecture Notes in Computer Science Vol. 4622, A. Menezes ed, Springer-Verlag, 2007. [4] Elaine Barker and John Kelsey, “Recommendation for Random Number Generation Using Deterministic Random Bit Generators”, NIST Special Publication 800-90. Revised March 2007. [5] M. Grangetto, E. Magli and G. Olmo, “Multimedia selective encryption by means of randomized arithmetic
Velammal College of Engineering and Technology, Madurai
coding”, IEEE Transactions on Multimedia, vol. 8, no. 5, 2006. [6] M. Abdalla, M. Bellare, D. Catalano, E. Kiltz, T. Kohno, T. Lange, J. Malone-Lee, G. Neven, P. Paillier and H. Shi, “Searchable Encryption Revisited: Consistency Properties”, Relation to Anonymous IBE, and Extensions. Advances in Cryptology - Crypto 2005 Proceedings, Lecture Notes in Computer Science Vol. 3621, V. Shoup ed, SpringerVerlag, 2005. [7] K. Pietrzak, M. Bellare and P. Rogaway, “Improved Security Analyses for CBC MACs”, Advances in Cryptology - Crypto 2005 Proceedings, Lecture Notes in Computer Science Vol. 3621, Springer-Verlag, 2005. [8] Y. Lu and S. Vaudenay, “Faster correlation attack on Bluetooth keystream generator E0. In M. Franklin, editor, Advances in Cryptology - Crypto 2004, 24th Annual International Cryptology Conference, Santa Barbara, California, USA, August 15-19, 2004. Proceedings, volume 3152 of Lecture Notes in Computer Science, pages 407 - 425. Springer-Verlag, 2004. [9] L. S. Choon, A. Samsudin, and R. Budiarto, “Lightweight and cost-effective MPEG video encryption, in Proc. of Information and Communication Technologies”, From Theory to Applications, 2004, pp. 525–526. [10] B. Canvel, A. Hiltgen, S. Vaudenay, and M. Vuagnoux,”Password interception in a SSL/TLS channel”, In D. Boneh, editor, Advances in Cryptology - Crypto 2003, 23rd Annual International Cryptology Conference, Santa Barbara, California, USA, August 17-21, 2003, Proceedings, volume 2729 of Lecture Notes in Computer Science, pages 583-599. Springer-Verlag, 2003. [11] ] M. Hellman, R.Merkle, R. Schroeppel, L.Washington, W. Diffie, S. Pohlig, and P. Schweitzer, “Results of an initial attempt to cryptanalyze the NBS Data Encryption Standard”. Technical Report SEL 76-042, Department of Electrical Engineering, Stanford University, 1976. [12] M. Bellare and B. Yee, “Forward-Security in PrivateKey Cryptography”, Topics in Cryptology - CT-RSA 03, Lecture Notes in Computer Science Vol. 2612, M. Joye ed, Springer-Verlag, 2003. [13] H. Gilbert and M. Minier, “New results on the pseudorandomness of some block cipher constructions”, In M. Matsui (Ed.) Fast Software Encryption - FSE 2001, Lecture Notes in Computer Science 2355, Springer-Verlag, 2002, 248– 266. [14] P. Rogaway, M. Bellare, J. Black and T. Krovetz, “OCB: A block-cipher mode of operation for efficient authenticated encryption”, Proceedings of the 8th ACM Conference on Computer and Communications Security (CCS), ACM, 2001.
Page 64
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[15] J. Kelsey, Bruce Schneier, David Wagner, and C. Hall, “Cryptanalytic Attacks on Pseudorandom Number Generators, Fast Software Encryption”, Fifth International Workshop Proceedings (March 1998), Springer-Verlag, 1998, pp. 168-188. [16] Mitsuru Matsui, “The First Experimental Cryptanalysis of the Data Encryption Standard”. In Advances in Cryptology Proceedings of CRYPTO ’94, Lecture Notes in Computer Science 839, Springer-Verlag, 1994. [17] Alfred J.Menezes, Paul C. van Oorschot, and Scott A. Vanstone, Handbook of Applied Cryptography, CRC Press, 1997. [18] Christof Paar and Jan Pelzl, "The Advanced Encryption Standard", Chapter 4 of "Understanding Cryptography, A Textbook for Students and Practitioners". Springer, 2009. [19] Diffie, Whitfield and Martin Hellman, "Exhaustive Cryptanalysis of the NBS Data Encryption Standard" IEEE Computer 10(6), June 1977, pp74–84. [20] John Kelsey, Stefan Lucks, Bruce Schneier, Mike Stay, David Wagner, and Doug Whiting, Improved Cryptanalysis of Rijndael, Fast Software Encryption, 2000 pp213–230.
Velammal College of Engineering and Technology, Madurai
Page 65
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Secure Multiparty Computation Based Privacy Preserving Collaborative Data Mining J.Bhuvana#1,T.Devi#2 #
School of Computer Science and Engineering, Bharathiar University Coimbatore-46, India 1
[email protected] 2 [email protected]
Abstract— In today’s competitive world, businesses are competing to stay ahead of their competitors by predicting the future based on intelligent analysis of available data in decision making. Towards this, various techniques such as reporting, online analytical processing, analytics, data mining, business performance management, text mining and predictive analytics are used. Data Mining is the process of processing large volumes of data, searching for patterns and relationships within that data. In case of voluminous business data being analysed, collaborative data mining techniques are practised. In collaborative data mining, preserving data privacy is a biggest challenge. This paper gives a brief description of Data Mining, Privacy Preserving in Data Mining (PPDM), introduces Secure Multiparty Computation (SMC) based Privacy Preserving Collaborative Data Mining (PPCDM). Keywords— Privacy, Security, Privacy Preserving Data Mining, Privacy Preserving Collaborative Data Mining, Secure Multiparty Computation.
I. INTRODUCTION Recent advances in data collection, data dissemination and related technologies have initiated a new era of research where existing data mining algorithms should be reconsidered from the point of view of privacy preservation. Privacy refers to the right of users to conceal their personal information and have some degree of control over the use of any personal information disclosed to others. To conduct data mining, one needs to collect data from various parties. Privacy concerns may prevent the parties from directly sharing the data and some types of information about the data. The way in which multiple parties conduct data mining collaboratively without breaching data privacy is a challenge. In recent years, the research community has developed numerous technical solutions for the privacy preserving data mining. However, Clifton has pointed out “the notion of privacy that satisfies both technical and societal concerns is unknown” [3]. Security is a necessary tool to build privacy, but a communication or transaction environment can be
Velammal College of Engineering and Technology, Madurai
very secure, yet totally unprivate. Security and privacy are closely related technologies, Privacy is about informational self-determination, the ability to decide what information about one goes where. Security offers the ability to be confident that those decisions are respected. Data Mining is an information extraction activity whose goal is to discover hidden facts contained in databases. Using a combination of machine learning, statistical analysis, modeling techniques and database technology, data mining finds patterns and relationships in data and infers rules that allow the prediction of future results. Privacy preserving data mining allows multiple data holders to collaborate to compute important information while protecting the privacy of other information using the technological tools such as cryptography, data perturbation and sanitization, access control, inference control and trusted platforms. To use the data mining algorithms, all parties need to send their data to a trusted central place to conduct the mining. However, in situations with privacy concerns, the parties may not trust anyone. This type of problem is called Privacy Preserving Collaborative Data Mining (PPCDM) [10]. For each data mining problem, there is a corresponding PPCDM problem. Collaborative data mining is categorised into homogeneous collaboration and heterogeneous collaboration. In homogeneous collaboration, each party has the same sets of attributes and in heterogeneous collaboration, each party has different sets of attributes. Fig.1 shows how a traditional data mining problem could be transformed to PPCDM problem. Security and privacy are related but different [11]. In the privacy preserving data mining, the system needs secured channels to protect network attackers. The system categorizes the protection into two layers as shown in Fig.2. One is the protection against the collaborative parties; the other is protection against network attackers. Without loss of generality, attacks from collaborative parties are considered inside attacks, these parties are called inside attackers; attacks outside the collaborative parties are considered outside attacks, the attackers who conduct the attacks are called outside attackers. To protect against outside attackers, the system need
Page 66
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
to rely on secure channels. The research work by Dolve et al [5] assumes that the communications between collaborative parties are encrypted by a non-malleable encryption scheme.
The focus of this research is to prevent inside attackers from knowing private data in collaborative data mining. Prevention of inside attackers is different from prevention of outside attackers in that the inside attackers usually have more knowledge about private data than outside attackers. Furthermore, the goal of collaborative data mining is to obtain valid data mining result. However, the result itself may disclose the private data to inside attackers. Therefore, the system cannot hope to achieve the same level of protection in privacy preserving collaborative data mining as in general secure communications which protect against the outside attackers. However, the system would like to prevent the private data from being disclosed during the mining stage [5]. Particularly, in this work, the privacy preservation means that multiple parties collaboratively get valid data mining results while disclosing no private data to each other or any party who is not involved in the collaborative computations. The authors are currently carrying out research on Secure Multiparty Computation based Privacy Preserving Collaborative Data Mining with the specific objectives of designing and developing an algorithm for Secure Multiparty Computation based Privacy Preserving Collaborative Data Mining and to analyse efficiency of the algorithm by scaling up with various factors such as number of parties involved in the computation, encryption key size and size of data set.
Fig.1 Privacy Preserving Non-Collaborative and Collaborative Data Mining .
Fig.2 Inside Attackers vs Outside Attackers
Velammal College of Engineering and Technology, Madurai
II. LITERATURE REVIEW In privacy-preserving data mining, to protect actual data from being disclosed, one approach is to alter the data in a way that actual individual data values cannot be recovered, while certain computations can still be applied to the data. This is the core idea of randomization-based techniques. The random perturbation technique is usually realised by adding noise or uncertainty to actual data such that the actual values are prevented from being discovered. Since the data no longer contains the actual values, it cannot be misused to violate individual privacy. Randomization approaches were first used by Agrawal and Srikant [1] to solve the privacy-preserving data mining problem. Following the idea of secure multiparty computation [6], Lindell and Pinkas [7] introduced a secure multi-party computation technique for classification using the ID3 algorithm, over horizontally partitioned data. Encryption is a well-known technique for preserving the confidentiality of sensitive information. Comparing with other techniques described, a strong encryption scheme can be more effective in protecting the data privacy. One of the encryption schemes is the Homo-morphic encryption which allows certain computations performed on encrypted data without preliminary decryption operations. Homo-morphic encryption is a very powerful cryptographic tool and has been applied in several research areas such as
Page 67
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
electronic voting and on-line auction. Privacy preserving data mining has generated many research successes. However, as Clifton [3] pointed out. The proposed system present a definition of privacy and use Homo-morphic encryption and digital envelope technique to achieve collaborative data mining without sharing the private data among the collaborative parties. In particular, the research provides a solution for k-nearest neighbor classification with vertical collaboration. III. SECURED MULTIPARTY COMPUTATION MODEL FOR COLLABORATIVE DATA MINING
i=1
Support A =
Tot. Nodes
∑ Database Size (i) i=1
Tot. Nodes
∑ Support Count AB (i) i=1
Support AÆB = Tot. Nodes
∑ Database Size (i)
P. Secured Multiparty Computation (SMC) The basic idea of SMC is that a computation is secure if at the end of the computation, no party knows anything except its own input and the results. One way to view this is to imagine a trusted third party; everyone gives their input to the trusted third party, who performs the computation and sends the results to the participants. The concept of SMC was first introduced by Yao [13] and it has been proved that for any function, there is a SMC solution [12]. The approach used is: the function F to be computed is first represented as a combinatorial circuit, and then the parties run their private function in their node on the network. Every participant gets corresponding input values and out values for every node. SMC not only preserves individual privacy, it also preserves leakage of any information other than the final result. However, traditional SMC method requires a high communication overhead; they do not scale well with the network size. Q. Association Rule Mining Computing Association rules without disclosing individual data items is straightforward. Distributed association rule mining techniques can discover rules among multiple nodes [14]. Mining association rules from vertically partitioned data, where the items are partitioned and each itemset is split between nodes, can be done by extending the apriori algorithm. Most of the apriori algorithm can be done locally at each of the nodes. The crucial step involves finding the support count of an itemset. The support determines how often a rule can be applied to a given data set and the confidence determines how frequently items in B appear in transactions that contain A. One can compute the global support and confidence of an association rule AÆB knowing only the local support of A, B and the size of each node’s data using the following formulae [1]:
i=1
Support AÆB Confidence
Velammal College of Engineering and Technology, Madurai
= Support A
R. Privacy Formalization A privacy-oriented scheme S preserves data privacy if for any private data T; the following holds [11]: |Pr (T |PPDMS) − Pr(T )| ≤ € Where, PPDMS: Privacy-Preserving data mining scheme. Pr(T | PPDMS): The probability that the private data T is disclosed after a Privacy-Preserving data mining scheme has been applied. Pr(T): The probability that the private data T is disclosed without any Privacy-Preserving data mining scheme being applied. Pr(T | PPDMS) - Pr(T): The probability that private data T is disclosed with and without Privacy-Preserving data mining schemes being applied. To achieve privacy-preserving data mining, reduce the whole algorithm to a set of component privacy oriented protocols. The privacy preserving data mining algorithm preserves privacy if each component protocol preserves privacy and the combination of the component protocols does not disclose private data. In the secure multiparty computation literature, a composition theorem describes a similar idea. A privacyoriented component protocol CP preserves data privacy if for any private data T, the following is held:
Tot. Nodes
∑ Support Count A (i)
AÆB
|Pr(T |CP) − Pr(T )| ≤ € Where,
Page 68
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
CP: Component protocol. Pr (T | CP): The probability that the private data T is disclosed after a Privacy-Preserving component protocol has been applied. Pr(T | CP) - Pr(T): The probability that private data T is disclosed with and without a Privacy-Preserving component protocol. S. Digital Envelope A digital envelope is a random number (or a set of random numbers) only known by the owner of private data[2] . To hide the private data in a digital envelope, to conduct a set of mathematical operations between a random number (or a set of random numbers) and the private data. The mathematical operations could be addition, subtraction, multiplication, etc. For example, assume the private data value is A. There is a random number R which is only known by the owner of A. The owner can hide A by adding this random number, e.g., A + R. T. Privacy-Preserving Collaborative Data Mining Problems Multiple parties, each having a private data set which is denoted by D1, D2, · · · and Dn respectively, want to collaboratively conduct a data mining task on the concatenation of their data sets. Because they are concerned about their data privacy or due to the legal privacy rules, neither party is willing to disclose its actual data set to others. P1 has a private data set D1, P2 has a private data set D2, · · · and Pn has a private data set Dn. The data set [D1 ∪ D2 ∪· · · ∪ Dn] forms a dataset, which is actually the concatenation of D1, D2, · · · and Dn. The n parties want to conduct k-nearest clustering over [D1 ∪ D2 ∪ · · · ∪ Dn] to obtain the mining results satisfying the given constraints. There are two types of collaborative models. In the vertical collaboration, diverse features of the same set of data are collected by different parties. In the horizontal collaboration, diverse sets of data, all sharing the same features, are gathered by different parties. U. Privacy-Preserving (K-NN)Classification
K-Nearest
Neighbor
Classification introduces a privacy constraint that is not easily addressed in the SMC model. If the target class in the training data is private information, the classifier could be used to scratch the privacy. One solution is not to build a global classifier. Instead, the parties collaborate to classify each instance. Therefore, the problem becomes how to share a classifier among the various distributed nodes while still supporting the classification function.
Velammal College of Engineering and Technology, Madurai
Currently many researchers developing methods for secure distributed K-NN classification; two different solutions for different level of security. One is security based on cryptographic assumptions and another one is finding roots of the large degree polynomials. The later ones is computationally hard. The K-NN classification is an instance based learning algorithm that has been shown to be very effective for a variety of problem domains. The objective of K-NN classification is to discover k nearest neighbors for a given instance, then assign a class label to the given instance according to the majority class of the k nearest neighbors. The nearest neighbors of an instance are defined in terms of a distance function such as the standard Euclidean distance. IV. PERFORMANCE EVALUATION The proposed system discloses the support values for each tested item set to all the participants. This discloses may cause for full disclosure in some cases. The research presents a new approach that does not disclose the support values during the execution and discloses less amount of information in general. In any site there is no external information about any other database. The collusion between parties should not be beyond a given threshold value. The various parties follow the protocol honestly. However, any attempt to deviate from the protocol will not result in disclosed information. They may try to use a correct protocol to infer information, but shall show that this will not be helpful for them. The parties may store intermediate or final results. The analysis of privacy is in terms of what can be inferred from those stored results only done with the help of a third untrusted party (this party is not trusted with the database, but it is trusted with computations). The parties change roles at the end of the phase. When information between the parties is shared, then only information in which some attributes are real (in one of the databases) is of use. That is, a fake transaction whose corresponding Transaction Identification (TID) in the other database is empty, is not considered at all. Its important to note that when each site computes large itemsets it doesn’t know whether the attributes corresponding to his real transaction, are real or not. V. CONCLUSIONS Worldwide markets are competitive and in order to capture markets, businesses are practising business intelligence techniques that involve team based data mining. Data mining helps in decision making and future prediction of market. In modern business world, collaboration is highly essential for carrying out business successfully. In this paper, data mining, privacy preserving data mining, secure multiparty computation techniques and the privacy problem of privacy preserving collaborative data mining are discussed. The authors are currently working on design of privacy preserving
Page 69
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
collaborative data mining algorithms using the concepts of secure multiparty computation. ACKNOWLEDGMENT I record my sincere thanks to Bharathiar University for providing necessary financial assistance to carry out my research. REFERENCES [1] R. Agrawal and R. Srikant, “Privacy-Preserving data mining”, in Proc. ACM SIGMOD, May 2000, pages 439–450. [2] D. Chaum, “Security without identification Communication”, ACM, 28(10), pp. 1030–1044, Oct. 1985. [3] C. Clifton, “What is privacy? critical steps for Privacy-Preserving data mining”, in IEEE ICDM Workshop on Security and Privacy Ascepts of Data Mining, Houston, Texas, USA, November, 2005, pp. 27-30. [4] T. Cover and P. Hart “Nearest neighbor pattern classification”, iIn IEEE Transaction of Information Theory, Vol. 13, pp. 21- 27, January, 1968. [5] Dolev, D. Dwork, and M. Naor, “Non-malleable cryptography”, in Proc. twenty-third annual ACM symposium on Theory of computing, New Orleans, Louisiana, United States, 1991,pp: 542 – 552. [6] O. Goldreich, “The foundations of cryptography”, Vol. 2, Cambridge University Press, 2004. [7] Y. Lindell and B. Pinkas, “Privacy preserving data mining”, Advances in Cryptology - Crypto2000, Lecture Notes in Computer Science, volume 1880, 2000. [8] P. Paillier, “Public-key cryptosystems based on composite degree residuosity classes”, Advances in Cryptography - EUROCRYPT ’99, Prague, Czech Republic, 1999, pp 223-238. [9] R. Rivest, L. Adleman, and M. Dertouzos, “ On data banks and privacy homomorphisms”, Foundations of Secure Computation, eds. R. A. DeMillo et al., Academic Press, pp. 169-179., 1978. [10] J. Zhan, “Privacy Preserving Collaborative Data Mining”, PhD thesis, Department of Computer Science, University of Ottawa, 2006. [11] J. Zhan, “Using Cryptography for Privacy Protection in Data Mining Systems”, Springer-Verlag Berlin Heidelberg, 2007. [12] O. Golreich (1998) “Secure Multiparty Computation (working draft)” [online]. Available: http://www.wisdom.weizmann.ac.il/home/oded/public html/foc.html [13] A.C. Yao, “Protocols for secure computations”, Proc. 23rd Annual IEEE Symposium on Foundations of Computer Science, 1982. [14] Agrawal, R. and Shafer, J.C,“Parallel mining of association rules,” IEEE Transactions on Knowledge and Data Engineering, Vol. 8, No. 6, 1996, pp.929-969.
Velammal College of Engineering and Technology, Madurai
Page 70
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Towards Customer Churning Prevention through Class Imbalance M.Rajeswari#1, Dr.T.Devi#2 #
School of Computer Science and Engineering, Bharathiar University, Coimbatore, India 1
[email protected] 2 [email protected]
Abstract— Technologies such as data warehousing, data mining, and campaign management software have made Customer Relationship Management (CRM) a new area where firms can gain a competitive advantage. Particularly through data mining a process of extracting hidden predictive information from large databases, organisations can identify their valuable customers, predict future behaviors, and enable firms to make proactive, knowledge-driven decisions. Data Mining along with Customer Relationship Management plays a vital role in today’s business environment. Customer churn, a process of retaining customer is a major issue. Prevention of customer churn is a major problem because acquiring new customer is more expensive than holding existing customers. In order to prevent churn several data mining techniques have been proposed. One among such method is solving class imbalance which has not received much attention in the context of data mining. This paper describes Customer Relationship Management (CRM), customer churn and class imbalance and proposes a methodology for preventing customer churn through class imbalance. Keywords— Data Mining, Customer Relationship Management, Churn, Class Imbalance.
I. INTRODUCTION World-wide businesses are competing to capture and retain customers and towards this Customer Relationship Management (CRM) are practiced. Organisations turn to CRM to enable them to be more effective in acquiring, growing and retaining their profitable customers. Customer relationship management has emerged as one of the demanding ways for the firms and it provides an effective way for customer satisfaction and retention [12]. In order to carry out CRM, the historical data about customers and their behavior need to be analysed. Towards such analysis, data mining techniques are used nowadays. Due to increase in competition between the companies they realized that customers are their valuable assets and retaining existing customers is the best way to survive. But it is analysed that it is more profitable to satisfy existing customers than to look for a new one [3]. Customer attrition,
Velammal College of Engineering and Technology, Madurai
also known as customer churn, customer turnover, or customer defection, is a business term used to describe loss of clients or customers. Recent studies analysed the importance of predicting as well as preventing customer churn in customer relationship management. Customer churn is often a rare event in service industries, but of great interest and great value, until recently, however, class imbalance has not received much attention in the context of data mining. A class imbalance is a term where one class occurs much more often than the other. Now, as increasingly complex real-world problems are addressed, the problem of imbalanced data is taking centre stage and class imbalance concept comes under rarity. Many researchers are concentrating on rarity that relates to rare classes, or, more generally, class imbalance. This type of rarity requires labeled examples and is associated with classification problems. The data set used to detect oil spills from satellite images provides a good example of a rare class. Because only 41 of the 937 satellite images contain oil slicks, it can be said that oil slicks are rare (i.e., a rare class). Several problems arise when mining rare classes and rare cases. A case corresponds to a region in the instance space that is meaningful with respect to the domain under study and the rare case is a case that covers a small region of the instance space and covers relatively few training examples, for example if bird is taken as a class, non-flying bird is rare case because only very few birds come under this category [22]. In order to prevent customer churn, researchers have proposed number of problems that arise when mining rare classes and rare cases. The problems include: improper evaluation metrics, lack of data: absolute rarity, relative lack of data: relative rarity, data fragmentation, inappropriate inductive bias and noise. To overcome these problems, several methods such as more appropriate evaluation metrics, non-greedy search techniques, using a more appropriate inductive bias, knowledge/human interaction, learn only the rare class, segmenting the data, accounting for rare items, cost-sensitive learning, sampling and other methods such as boosting, two
Page 71
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
phrase rule induction, place rare cases into separate classes are followed [22]. Among these methods non-greedy search techniques, using a more appropriate inductive bias, learn only the rare classes are at conceptual stage only [5]. The rest of this paper is organized as follows: In Section 2, an overview of Data Mining is presented. Section 3 discusses about Customer Relationship Management and particularly about Customer Churn. An introduction to Class imbalance usage in predicting customer churn is described in Section 4. Methods followed in order to solve class imbalance are detailed in Section 5 and finally Conclusions are given in Section 6. II. DATA MINING (DM) Organizations are implementing data warehousing technology, which facilitates enormous enterprise-wide databases and the amount of data that organizations possess is growing at a phenomenal rate and the next challenge for these organizations is how to transform it into useful information and knowledge. Data Mining is one of the technologies used for meeting this challenge; data mining is the exploration and analysis, by automatic or semiautomatic means, of large quantities of data in order to discover meaningful patterns and rules. Data mining is defined as the process of extracting interesting and previously unknown information from data, and it is widely accepted to be a single phase in a complex process known as Knowledge Discovery in Databases (KDD). This process consists of a sequence of the following steps: data cleaning, data integration, data selection, data transformation, data mining, pattern evaluation and knowledge presentation [23]. Thus data mining is the search of valuable information in large volumes of data. This combines the efforts of human, who design the databases, describe the problems and set the goals, and computers who process the data looking for patterns that match this goal. Data mining sits at the common frontiers of several fields including database systems, artificial intelligence, statistics, machine learning, and pattern recognition or data visualization. In order to ensure that the extracted information generated by the data mining algorithms is useful, additional activities are required, like incorporating appropriate prior knowledge and proper interpretation of the data mining results. The basic objective is to construct a model for one situation in which the answer or output is known and then apply that model to another situation in which the answer or output is desired. The best applications of the above techniques are integrated with data warehouses and other interactive, flexible business analysis tools. There are many reasons by which all business environment started to analyze the performance of data mining in order to recover from all problems [9]. Databases today are huge containing more than 1,000,000 entities/records/rows, from 10 to 10,000 fields/attributes/
Velammal College of Engineering and Technology, Madurai
variables, gigabytes and terabytes. Databases are growing at an unprecedented rate and decisions must be made rapidly with maximum knowledge. Currently business to business marketers and service retailers are the major users of data mining [1]. Data mining involves many different algorithms to accomplish different tasks. All these algorithms attempt to fit a model to the data. The algorithms examine the data and determine a model that is closest to the characteristics of the data being examined. The created data mining model can be either predictive or descriptive in nature (e.g. Fig. 1). A Predictive Model – makes a prediction about values of data using known results found from different data. Predictive modeling may be made based on the use of other historical data. A Descriptive Model – identifies patterns or relationships in data. Unlike the predictive model, a descriptive model serves as a way to explore the properties of the data examined, not to predict new properties.
Fig. 1 Data Mining Models and Tasks [10]
Clustering, Summarization, Association rules, and Sequence discovery are usually viewed as descriptive in nature. The individual data mining tasks are combined to obtain more sophisticated data mining applications. There are several data mining techniques available and used. Mostly used techniques are associations, classifications, sequential patterns and clustering. The basic aim of an association is to find all associations, such that the presence of one set of items in a transaction implies the other items. Classification develops profiles of different groups. Sequential patterns subject to a user–specified minimum constraint. Clustering segments a database into subsets or clusters. The classifications of the data mining techniques are: user–guided or verification–driven data mining and discovery–driven or automatic discovery of rules. Most of the techniques of data mining have elements of both the models. Data mining is used in several applications such as retail, banking, telecommunication and other relevant applications. For example, when telecommunication companies around the world face competition which is forcing them to aggressively market special pricing programs aimed at retaining existing
Page 72
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
customers and attracting new ones. Knowledge discovery in telecommunication includes Call detail record analysis and Customer loyalty, Call detail record analysis: Telecommunication companies accumulate detailed call records. By identifying customer segments with similar use patterns, the companies can develop attractive pricing and feature promotions and Customer loyalty: Some customers repeatedly switch providers, or “churn”, to take advantage of attractive incentives by competing companies. The companies can use data mining to identify the characteristics of customers who are likely to remain loyal once they switch, thus enabling the companies to target their spending on customers who will produce the most profit. Similarly knowledge discovery applications are emerging in variety of industries such as customer segmentation and manufacturing [8]. III. CUSTOMER RELATIONSHIP MANAGEMENT (CRM) The way in which companies interact with their customers has changed dramatically over the past few years. A customer’s continuing business is no longer guaranteed. As a result, companies have found that they need to understand their customer better, and to quickly respond to their needs. Customer Relationship Management is the term used for the business practice allowing businesses that have more than a few customers to better serve and manage the interactions with those customers. CRM is defined as “a comprehensive strategy and process of acquiring, retaining and partnering with selective customers to create superior value for the company and the customer. It involves the integration of marketing, sales, customer service, and the supply chain functions of the organization to achieve greater efficiencies and effectiveness in delivering customer value” [18]. CRM is defined by a framework consisting of four elements: Know, Target, Sell, and Service. CRM requires the firm to know and understand its markets and customers. This involves detail customer intelligence in order to select the most profitable customers and identify those longer worth targeting. CRM also entails development of the offer: which product to sell to which customers and through which channel. In selling, firms use campaign management to increase the marketing department’s effectiveness. Finally, CRM seeks to retain its customers through services such as call centers and help desks. CRM is essentially a two-stage concept: the task of the first stage is to master the basics of building customer focus. This means moving from a product orientation to a customer orientation and defining market strategy from outside-in and not from inside-out. The focus should be on customer needs rather than product features. Companies in the second stage are moving beyond the basics; they do not rest on their laurels but push their development of customer orientation by integrating CRM across the entire customer experience chain, by leveraging technology to achieve real-time customer
Velammal College of Engineering and Technology, Madurai
management, and by constantly innovating their value proposition to customers [8]. CRM consists of four dimensions: Customer Identification, Customer Attraction, Customer Retention/Churn, and Customer Development. Customer Management System contains these four dimensions in closed cycle and used to create a better understanding of their customers and create a long term relationship with the organization [2]. The three stages in customer life cycle are acquiring customers, increasing the value of the customer and retaining good customers. Data mining can improve profitability in each stage through integration with operational CRM systems or as independent applications (e.g. Fig. 2). Data mining helps to accomplish such a goal by extracting or detecting hidden customer characteristics and behaviors from large databases. The generative aspect of data mining consists of the building of a model from data [7]. CRM is not only applicable for managing relationships between businesses and consumers, but even more crucial for business customers. In B2B environments, transactions are more numerous, custom contracts are more diverse and pricing schemes are more complicated. CRM strategies, such as customised catalogues, personalised business portals, and targeted product offers, can help smooth this process and improve efficiencies for both companies [8].
Fig. 2 Three simple business goals for CRM [14]
From the perspective of customer behaviour, Reference [6] suggests that B2B buyers choose a supplier with whom they can develop a relationship; one they can go back to as required and one on which they feel they can depend. Once they have chosen a supplier, having invested this time and effort, they are more likely to stay with that supplier for longer. This invokes the equal importance of deploying CRM in both recruiting new customers and maintaining existing customers. V. Customer Churn Attrition or churn is a growing problem in many industries and is characterised by the art of customer switching companies,
Page 73
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
usually to take advantage of “a better deal”. One of the most important issues for companies is customer retention. A persistent change in both the cost and revenue sides is customer churn, the movement of customers from provider to provider in search of better and cheaper products and services. Specifically churn is the gross rate of customer loss during a given period. Churn is a term used to represent the loss of a customer, in order to prevent churn one should acquire more loyal customers initially and then identify customers who are most likely to churn. This type of predictive churn modeling is applied in the field of banking, mobile telecommunication and life insurances. Churn can be shown as follows: Monthly churn = (C0 + A1 - C1) / C0 C0 is the number of customers at the start of the month, A1 is the number of customers at the end of the month, C1 is the gross new customers during the month. There are several reasons for a particular customer to churn such as Price, Service Quality, Fraud, Lack of carrier responsiveness, Brand disloyalty, Privacy concerns, Lack of features, new technology or product introduced by competitors. Minimizing customer churn provides a number of benefits, such as: Minor investment in acquiring a new customer, higher efficiency in network usage, Increase of added-value sales to long term customers, Decrease of expenditure on help desk, Decrease of exposure to frauds and bad debts and higher confidence of investors. Churn can be defined and measured in different ways: “Absolute” Churn is number of subscribers disconnected, as a percentage of the subscriber base over a given period, “Line” or “Service” Churn is number of lines or services disconnected, as a percentage of the total amount of lines or services subscribed by the customers, “Primary Churn” is number of defections and “Secondary Churn” is drop in traffic volume, with respect to different typology of calls. Measuring churn is getting more and more difficult, growing tendency for Business users to split their business between several competing fixed network operators. Carrier selection enables Residential customers to make different kind of calls with different operators, Carrier pre-selection and Unbundling of the Local Loop makes it very difficult to profile customers according to their “telecommunication needs”. IV. CHURN PREDICTION USING CLASS IMBALANCE A dataset is imbalanced if the classification categories are not approximately equally represented. Recent years brought increased interest in applying machine learning techniques to difficult "real-world" problems, many of which are characterized by imbalanced data. Additionally the distribution of the testing data may differ from that of the training data, and the true misclassification costs may be unknown at learning time. Predictive accuracy, a popular choice for evaluating
Velammal College of Engineering and Technology, Madurai
performance of a classifier, might not be appropriate when the data is imbalanced and/or the costs of different errors vary markedly [17]. The class imbalance problem corresponds to domains for which one class is represented by a large number of examples while the other is represented by only few. Class imbalance falls under rarity which deals with both rare classes (i.e. class imbalance) and rare cases. Fig. 3 shows an artificial domain with two classes, A and B, where A is the rare (minority) class and B is the common (majority) class. Holding with established conventions, the rare class is designated the positive class and the common class is designated the negative class. The true decision boundaries are displayed with solid lines while the learned boundaries are displayed with dashed lines. The labeled examples are represented in the figure using the “+” and “-” symbols. The five subconcepts associated with class A are labeled A1A5. Subconcepts A2-A5 correspond to rare cases, whereas A1 corresponds to a fairly common case, covering a substantial portion of the instance space. The majority class is comprised of two subconcepts, B1 and B2. Subconcept B1 is a very general case that covers most of the instance space. Subconcept B2, on the other hand, corresponds to a rare case, demonstrating that common classes may contain rare cases. However, it is expected that rare classes, by their very nature, to contain a greater proportion of rare cases (a very rare class cannot contain any common cases). Fig. 3 also shows that rare cases A3, A4, and A5 cause small disjuncts to be formed in the learned classifier (rare case A2 is “missed” completely by the classifier). Rare cases, like rare classes, can be considered the result of a form of data imbalance and have in fact been referred to as within-class imbalances [13].
Fig. 3 Graphical representation of rare classes and rare cases [22]
To predict customer churn many approaches are carried out and one such approach is to handle imbalance classes, in order to solve the problem of imbalance classes researchers have introduced various methods such as: Improper evaluation metrics, Lack of data: absolute rarity, Relative lack of data: relative rarity, Data fragmentation, Inappropriate inductive bias,
Page 74
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
and Noise. Improper evaluation metrics: often, not the best metrics are used to guide the data mining algorithms and to evaluate the results of data mining, Lack of data: absolute rarity: the number of examples associated with the rare class is small in an absolute sense, which makes it difficult to detect regularities within the rare class, Relative lack of data: relative rarity : objects are not rare in absolute sense, but are rare relative to other objects, which makes it hard for greedy search heuristics, and more global methods are, in general, not tractable, Data fragmentation : Many data mining algorithms, like decision trees, employ a divide-and-conquer approach, where the original problem is decomposed into smaller and smaller problems, which results in the instance space being partitioned into smaller and smaller pieces. This is a problem because regularities can then only be found within each individual partition, which will contain less data, Inappropriate inductive bias: Generalizing from specific examples, or induction, requires an extra-evidentiary bias. Without such a bias ‘‘inductive leaps” are not possible and learning cannot occur. The bias of a data mining system is therefore critical to its performance. Many learners utilize a general bias in order to foster generalization and avoid overfitting. This bias can adversely impact the ability to learn rare cases and rare classes, Noise: Noisy data will affect the way any data mining system behaves, but interesting is that noise has a greater impact on rare cases than on common cases [22]. V. PROPOSED WORK Customer churn prevention as a part of customer relationship management approach is high on the agenda, and it is found that almost each and every company is forced to implement churn prediction model to detect possible churners before they effectively leave the company. In order to predict churn, more and more data mining techniques are applied and fortunately for the companies involved, churn is often a rare object, but of great interest and great value, there is a need for understanding how to model rare events. ‘‘The models developed in marketing are typically applied to situations where the events of interest occur with some frequency (e.g., customer churn, customer purchases). These models can break down when applied to setting where the behaviour of interest is rare. For example, when modelling the correlates of customer acquisition in a lowacquisition rate setting, the performance of the familiar logit model is often unacceptable. There may be opportunity to gain valuable insights from the statistics literature on the modelling of rare events” [11]. Due to the increase in complex real-world problems, the problem of imbalance data is taking center stage and many researchers are found to solve such problems using various techniques and methods. There are six data mining problems related to rarity, and lists ten methods to address them, they are improper evaluation metrics, lack of data: absolute rarity, relative lack of data:
Velammal College of Engineering and Technology, Madurai
relative rarity, data fragmentation, inappropriate inductive bias and noise. To overcome these problems several methods are followed to deal with rarity such as more appropriate evaluation metrics, non-greedy search techniques, using a more appropriate inductive bias, knowledge/human interaction, learn only the rare class, segmenting the data, accounting for rare items, cost-sensitive learning, sampling and other methods such as boosting, two phrase rule induction, place rare cases into separate classes [22]. Among those methods non-greedy search techniques, using a more appropriate inductive bias, learn only the rare classes are at conceptual stage. Non-Greedy Search Techniques: Due to several disadvantages in greedy method, Reference [22] has suggested carrying out non-greedy techniques to handle imbalance classes, one among such non-greedy techniques is genetic algorithm. To solve the problem genetic algorithm makes use of candidate solution instead of single solution and then employs stochastic operators. It gives a conclusion that genetic algorithms work better with attribute interaction and thus avoids getting stuck into local maxima, which all together makes genetic algorithms suitable finding rarity. This would give a clear idea that for what reason genetic algorithms are being increasingly used for data mining. Several systems have relied to the power of genetic algorithms to handle rarity. Some of them used a genetic algorithm to predict very rare events while and other researches tried to use a genetic algorithm to discover “small disjunct rules” [21]. Using a More Appropriate Inductive Bias: Several attempts have been taken to improve the performance of data mining systems with respect to rarity by choosing a more appropriate bias. The simplest approach involves modifying existing systems to eliminate some small disjuncts based on tests of statistical significance or using error estimation techniques. The hope is that these will remove only improperly learned disjuncts [20]. Unfortunately, this approach was shown not only to degrade performance with respect to rarity, but also to degrade overall classification performance. Many researches worked on this concept with some techniques and obtained better results, like the usage of C4.5 algorithm. But fortunately the result showed only mixed success, because they tried out using only overall classification accuracy instead of small disjuncts. Learn Only The Rare Classes: When learning a set of classification rule for all classes, it is found that several rare classes are ignored and hence, the only solution is to predict rare classes. This technique is proved using Support vector machines and neural networks, and one data mining technique that utilizes this recognition-based approach is Hippo. Ripper is a rule induction system that utilizes a separate-and-conquer approach to iteratively build rules to cover previously uncovered training examples. Each rule is grown by adding conditions until no negative examples are covered. It normally
Page 75
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
generates rules for each class from the most rare class to the most common class. Given this architecture, it is quite straightforward to learn rules only for the minority class—a capability that Ripper provides. VI. CONCLUSION Churn is often considered as a rare object and in order to handle these rare objects, class imbalance (i.e. rare classes) have been utilised by many researchers. This paper gives ideas about various data mining techniques that have been applicable for customer relationship management concept particularly in the prevention of customer churn by using class imbalance. The authors are currently designing a methodology for preventing customer churn through class imbalance.
ACKNOWLEDGMENT I record my sincere thanks to Bharathiar University for providing necessary financial assistance to carry out my research. REFERENCES [1] Armstrong, G., and P. Kolter, Principles of Marketing, Prentice Hall New Jersey, 2001. [2] Au, W. H., Chan, K. C. C., & Yao, X, “A novel evolutionary data mining algorithm with applications to churn prediction,” IEEE Transactions on Evolutionary Computation, vol. 7, pp. 532–545, 2003. [3] Berry, M. J. A., & Linoff, G. S, Data mining techniques second edition - for marketing, sales, and customer relationship management, Wiley, 2004. [4] Berson, A., Smith, S., & Thearling, K, Building data mining applications for CRM, McGraw-Hill, 2000. [5] Burez, J., Van den Poel. D, “Handling class imbalance in customer churn prediction,” Expert Systems with Applications, vol. 36, pp. 4626–4636, 2009. [6] Bush, R., The Interactive and Direct Marketing Guide, The Institute of Direct Marketing, Middlesex, 2002, Chapter 3.6. [7] Carrier, C. G., & Povel, O, “Characterising data mining software. Intelligent Data Analysis,” vol. 7, pp. 181–192, 2003. [8] Chris Rygielski, Jyan-Cheng Wang, David C.Yen, “Data Mining and Customer RElatonship Management,” Technology in Society, vol. 24, pp. 483-502, 2004. [9] Fayyad, U.M. (2003). Editorial. SIGKDD Explorations, 5(2). [10] Garver. M. S, “Using Data Mining for Customer Satisfaction Research. Marketing Research” vol. 14, no. 1, pp. 8–12, 2002. [11] Gupta, S., Hanssens, D., Hardie, B., Kahn, W., Kumar, V., Lin, N., et al, “Modeling customer lifetime value,” Journal of Service Research, 2006, 9 (2), 139–155. [12] Jayanthi Ranja., Vishal Bhatnagar, “Critical Success Factors For Implementing CRM Using Data Mining,” Journal of Knowledge Management Practice, vol. 9, 2008. [13] Japkowicz, N, “Concept learning in the presence of between class and within-class imbalances,” In Proceedings of the Fourteenth Conference of the Canadian Society for Computational Studies of Intelligence, 2001, pages 67-77, SpringerVerlag. [14] Jorg-Uwe Kietz, “Data Mining for CRM and Risk Management,” Knowledge discovery services and knowledge discovery applications, 2003.
Velammal College of Engineering and Technology, Madurai
[15] Kracklauer, A. H., Mills, D. Q., & Seifert, D, “Customer management as the origin of collaborative customer relationship management. Collaborative Customer Relationship Management - taking CRM to the next level,” 2004, 3–6. [16] Neslin, S., Gupta, S., Kamakura, W., Lu, J., & Mason, C, “Detection defection: Measuring and understanding the predictive accuracy of customer churn models. Journal of Marketing Research,” vol. 43(2), pp. 204–211, 2006. [17] Nitesh V. Chawla, “Data Mining For Imbalanced Datasets: An Overview,” pp. 853-867, 2005. [18] Parvatiyar, A., & Sheth, J. N, “Customer relationship management: Emerging practice, process, and discipline,” Journal of Economic & Social Research, vol. 3, pp. 1–34, 2001. [19] Regielski, C., Wang, J.C. & Yen, D.C, “Data Mining Techniques for Customer Relationship Management,” Technology in Society, vol. 24, pp. 483502, 2002. [20] Weiss, G. M, “Learning with rare cases and small disjuncts,” In Proceedings of the Twelfth International Conference on Machine Learning, pages 558-565, Morgan Kaufmann, 1995. [21] Weiss, G. M, “Timeweaver: a genetic algorithm for identifying predictive patterns in sequences of events,” In Proceedings of the Genetic and Evolutionary Computation Conference, pages 718-725, Morgan Kaufmann, 1999. [22] Weiss, G. M. (2004), “Mining with rarity: A unifying framework. SIGKDD Explorations,” vol. 6 (1), pp. 7–19, 2004. [23] Wirth, R. and Hipp, J, “CRISP-DM: Towards a standard process model for data mining,” In Proceedings of the 4th International Conference on the Practical Applications of Knowledge Discovery and Data Mining, 2000, pages 29-39, Manchester, UK.
Page 76
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Designing health care forum using semantic Search engine & diagnostic ontology #1
#2
Prof.Mr.V.Shunmughavel #1, Dr.P.Jaganathan #2, Department of CSE, K.L.N. College of Information Tech. Pottapalayam, 630611. Professor& Head, Department of Computer Applications, P.S.N.A. College of Engg & Tech, Dindigul. [email protected], [email protected]
Abstract. Healthcare Forum is a coalition of doctors, hospitals, health plans and state associations, that have joined together to improve the health care system. It should follow the best practices and evidence-based guidelines, with a strong emphasis on patient safety. Chronic diseases are on the rise and the population is ageing. Added to this, advancing technologies, new medications, and new procedures and providers have a wide array of tools to manage in order to deliver high quality health care. With the advent of semantic web, the fast dissemination of ontologies to represent and solve information causes a deep impact on knowledge retrieval as a whole. Ontology is increasingly seen as a key factor for enabling interoperability across heterogeneous systems and semantic web applications. The main objective of this research paper is to develop an effective functioning forum which integrates the unique association thinking of humans with an intelligent diagnostic Ontology to devise an automated semantic search engine which provides recommended solutions for patients, doctors and research activities in the medicine worldwide.
Keywords: Ontology, Semantic Search 1.
INTRODUCTION
In general any Forum is a group creativity technique designed to generate a large number of ideas for the solution to a problem. Although forum has become a popular group technique, researchers have generally failed to find evidence of its effectiveness for enhancing either quantity or quality of ideas generated. Because of such problems as distraction, social loafing, evaluation apprehension, and production blocking, brainstorming groups are little more effective than other types of groups, and they are actually less effective than individuals working independently. For this reason, there have been numerous attempts to improve traditional forum design
Velammal College of Engineering and Technology, Madurai
techniques or replace it with more effective variations of the basic technique. A good example to quote for the current forum sessions is the simple forum available over the internet. In these forums questions will be posted by users and answers for those questions will be posted by various users. Until now, the best answer for a particular question will be judged by a human. There is no guarantee that the answer selected as the best answer to be the perfect solution always. The thinking and decision making capability of human mind differs from person to person. So on the whole, opinion differs. In order to overcome the demerits in the current traditional forum sessions, in this research a novel idea is proposed for choosing the best recommended solution to a question. In general in any forum, first a user posts a question. Now users who are interested in that question can view that question and cast their answers. It is not necessary that the answer given by a user should be the perfect solution. Instead any number of solutions can be posted by users for a particular question. These solutions will now be available for everyone to view. In addition to this, the research focuses on providing a recommended solution using Ontology. With this aim, this work is organized as follows: Section 2 discusses the related work. Section 3 discusses the diagnostic ontology building methodologies. Section 4 discusses the proposed architecture of Health Care Forum. Finally Section 5 presents conclusion and future scope of this work. 2.
RELATED WORK
The main objective of the website [9] is to deliver updated knowledge for Practicing Anesthesiologists and extend helping hand at the time of facing critical cases and also to
Page 77
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
create awareness amongst them. Semantically enhanced information extraction system providing automatic annotation with references to classes in the ontology and instances in the knowledge base is presented. Based on these annotations, we perform IR-like indexing and retrieval, further extended by using the ontology and knowledge about the specific entities [10]. Web Ontology Language (OWL) is a language for defining and instantiating web ontologies (a W3C Recommendation). OWL is designed for use by applications that need to process the content of information instead of just presenting information to humans. This is used for knowledge representation and also is useful to derive logical consequences from OWL formal semantics [11]. Matthew Horridge[12]introduces the Protege-OWL plug-in for creating OWL ontologies. It gives a brief overview of the OWL ontology language and focuses on building OWLDL ontology and by using a description logic reasoner to check the consistency of the ontology and automatically compute the ontology class hierarchy. The vision of a semantic Web is extremely ambitious and would require solving many long-standing research problems in knowledge representation and reasoning, databases, computational linguistics, computer vision, and agent systems [13]. In OWL Web Ontology Language Guide [14], The OWL Web Ontology Language is intended to provide a language that can be used to describe the classes and relations between them that are inherent in Web documents and applications. This document demonstrates the use of the OWL language to 1. formalize a domain by defining classes and properties of those classes, 2. define individuals and assert properties about them, and 3. reason about these classes and individuals to the degree permitted by the formal semantics of the OWL language. Oscar Corcho [7] presents how to build ontology in the legal domain. Graciela Brusa [8] presents a discussion on the process and product of an experience in developing ontology for the Public Sector whose organization requires a strong knowledge management. Particularly, this process has been applied to develop ontology for Budget Domain. Pablo Castells [15] assumes a knowledge base has been built and associated to the information sources (the document base), by using one or several domain ontologies that describe concepts appearing in the document text. Our system can work with any arbitrary domain ontology with essentially no restrictions, except for some minimal requirements, which basically consist of conforming to a set of root ontology classes. The concepts and instances in the KB are linked to the documents by means of explicit, non embedded annotations to the documents.
Velammal College of Engineering and Technology, Madurai
3.
Building Diagnostic Ontology
This section describes the process of an experience in developing a Diagnostic Ontology for Health Care Domain (Fig.1) [7] [8]. 3.1.
Specification: The Ontology Goal and Scope
Fig. 1. Specification The scope limits the ontology, specifying what must be included and what must not. It is an important step Fig. 1 for minimizing the amount of data and concepts to be analyzed, especially for the extent and complexity of the diagnostic semantics. In successive iterations for verification process, it will be adjusted if necessary. 3.2.
Specification: Domain Description
In this analysis, the application to formulating the diagnostic ontology and its related documentations were studied and revised. Furthermore, meetings with a group of experts were carried out. This group was conformed by Doctors, and software engineers who bring informatics supports for these tasks. As it can be seen, the group of experts was very heterogeneous. In addition, they do not have much time to assign the meetings. This group was the support for knowledge acquisition during the ontology development. Then, we have to define different intermediate representations to communicate the knowledge acquired to the experts considering the background of each one and the time of meetings. Then, a brief description of the domain is presented.
Page 78
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3.3.
Specification: Motivating Scenarios
The motivation scenarios show problems that arise when people need information that the system does not provide. Besides, the scenario description contains a set of solutions to these problems that includes the semantic aspects to solve them. In order to define motivation scenarios and communicate them to the involved people, templates have been used. These templates were based on those proposed to specify case uses in object oriented methodology. The template describes: the name of the scenario, people who participate in the scenario, a brief scenario description, and a list of possible terms related to the scenario. Since this template shows the most important information in a concise way, it is useful when the experts do not have a lot of time to analyze the scenarios. 3.4. Specification: Ontology Granularity and Type According to the level of conceptualization and granularity, the ontology proposed here is domain ontology. Domain ontology describes the vocabulary related to a specific domain and, the ontology objective is to facilitate the discussion among the Health Care Forum members. 3.5. Conceptualization: Domain Conceptual Model
defined concepts. The UML class diagram can be used to express concepts in terms of classes and relationships among them. Although UML in its standard form is not suitable for semantic representation, the information modeled in the UML class diagram was the base for building the ontology term glossary, trying to include other concepts by means of generalization and specialization techniques. 3.5. Conceptualization: Identification of Classes, Relations and Attributes We used Concepts classifier trees to analyze hierarchies and attributes, binary relations, axioms and instances tables. For determining classes, we identified those terms of independent existence from the key terms list and the glossary. Once the conceptual model of the ontology has been created, the next step is to define relevant instances 3.7. Implementing the Diagnostic Ontology with PROTEGE
Fig. 3. Implementation
Fig. 2. Conceptualization In this step Fig. 2, a list of the most important terms was elaborated. The core of basic terms is identified first and then they are specified and generalized if necessary. Then with these concepts as reference, the key term list was defined. To properly understand the conceptual aspects in the context, a Unified Modeling Language (UML) [5] diagram was elaborated with the main relations among
Velammal College of Engineering and Technology, Madurai
In order to implement Fig. 3 the diagnostic ontology, Protege tool is an opt tool because of the fact that it is extensible and provides a plug-and-play environment that makes it a flexible base for rapid prototyping and application development. Protege ontologies can be exported into different formats including Relational Description Framework Schema and Web Ontology Language (OWL). Particularly, it is proposed to implement the Diagnostic Ontology in OWL. To compare the ontology implementation with its conceptualization, graphics using the OWLViz and Ontoviz plug-ins will be generated and compared with UML diagrams. On the one hand, OWLViz enables the class hierarchies in OWL Ontology to be viewed, allowing comparison of the asserted class hierarchy and the inferred class hierarchy.
Page 79
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
4.HEALTH CARE FORUM ARCHITECTURE & ENVIRONMENT This system aims at constructing a new Health Care Forum model, integrating semantic based diagnostic ontology module using semantic search engine. As depicted in Fig. 4, the architecture and the environment have several components. It also provides an overall high-level description of the interactions between the main components of the system. These parts and their interactions are described as follows:
concepts appearing in the document text. The concepts and instances in the KB are linked to the documents by means of annotations to the documents. Documents are annotated with concept instances from the KB by creating instances. Weights are computed automatically, based on the frequency of occurrence of the instances in each document. The annotations are used by the retrieval process. The Knowledge Base records every client’s domain knowledge, which comprises idea instances and the relationships between these instances. That is, this study assumes the existence of a given set of participants and their relevant knowledge. 4.3.
Semantic Search Engine Module
Semantic Search Engine Module is the prime module which acts like interface between the Data Module and Diagnostic Ontology. In this module Jena framework has been used for the implementation of RDF Data Query Language (RDQL) [3]. Jena is a Java toolkit which provides an API for creating and manipulating RDF models. Jena sources can be retrieved at http://jena.sourceforge.net/. RDQL is an implementation of an SQL-like query language for RDF. The RDQL query is executed against the knowledge base, which returns a list of instance tuples that satisfy the query. The semantic search engine extracts recommended solutions from the diagnostic ontology module with the help of instance tuples. These recommended solutions will be sent to the forum blackboard. Fig. 4: Proposed Architecture of Health Care Forum 4.1.
User Interface Module
This module has two components: Open Problem and Participants Component [1]. An “open problem” is an initial topic or issue given to the system. “Participants” means the registered users attending the Health Care Forum session. Both the components represent the given open problem, the registered users and posted questions with related topic.
4.2.
Data Module
This comprises two parts: Document Base and Knowledge Base(KB). The document base stores fundamental information about registered users, posted questions and replies. A knowledge base has been built and associated to the information sources (the document base), by using one or several domain ontologies that describe
Velammal College of Engineering and Technology, Madurai
4.4.
Diagnostic Ontology Module
Diagnostic Ontology is already created with the help of Domain Experts and implemented by using Protege tool and kept for use. Very often the Diagnostic Ontology is updated with recent findings and developments and used by Health Care Forum. 5. CONCLUSION Building domain ontologies is not a simple task when domain experts have no background knowledge on engineering techniques and/or they don’t have much time to invest in domain conceptualization. Sharing the best practice on ontology building can be useful for the whole community. In our work, we have also shown how ontologies could develop merging different methodologies and software engineering. Particularly this approach has been used to define Domain ontology for Medical Diagnostic System which could be used by Doctors and Experts of various domains belongs to Medicine. The proposed architecture for Health Care forum integrates the unique association thinking of humans with an
Page 80
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
intelligent diagnostic Ontology to devise an automated and suitable solution for Patients, Doctors and Research activities in the medicine worldwide. As future work, we plan to complete the implementation of our approach with practical evaluation. 6.
REFERENCES
[1] Soe-Tsyr Yuan and Yen-Chuan Chen (2008) “Semantic Ideation Learning for Agent-Based E-Brainstorming”, IEEE Transactions on Knowledge And Data Engineering, Vol. 20, No. 2. [2] A.R. Dennis, A. Pinsonneault, K.M. Hilmer, H. Barki, R.B. Gallupe, M. Huber, and F.Bellavance (2005) “Patterns in Electronic Brainstorming: The Effects of Synergy and Social Loafing on Group Idea Generation” Int’l J. e-Collaboration, vol. 1, no. 4, pp. 38- 57. [3] W.-L.Chang and S.-T. Yuan (2004) “iCare Home Portal: A Quest for Quality Aging e-Service Delivery,” Proc. First Workshop Ubiquitous Seaborne A., RDQL - A Query Language for RDF, W3C Member Submission. http://www.w3.org/Submission/2004/SUBM-RDQL-20040109/ [4] Smith,M., Welty.C., McGuinness.D. (2004) OWL Web Ontology Language Guide, W3C Recommendation 10 http://www.w3.org/TR/owl-guide/ [5] UML (2006) Unified Modeling Language. http://www.uml.org/ Brickley, D., Guha, R.V. (2004) RDF Vocabulary Description Language 1.0: RDF Schema. W3C Recommendation. http://www.w3.org/TR/rdf-schema/ [6] Caliusco M. L. (2005), “A Semantic Definition Support of Electronic Business Documents in e–Colaboration”, PhD thesis. UTN - F.R.S.F. Santa Fe, Argentina. [7] Corcho O, Fernández-López M, Gómez-Pérez A, LópezCima A. (2005), “Building legal ontologies with METHONTOLOGY and WebODE. Law and the Semantic Web, Legal Ontologies, Methodologies, Legal Information Retrieval, and Applications”, Australian Computer Society, Inc. T Australasian Ontology Workshop (AOW 2006), Hobart, Australia. Conferences in Research and Practice in Information Technology, Vol. 72. [8] Graciela Brusa, “A Process for Building a Domain Ontology: an Experience in Developing a Government Budgetary Ontology” Dirección Provincial de Informática, San Martín 2466, Santa Fe (Santa Fe), Argentina, [email protected] [9] www.isakanyakumari.com [10] www.ontotext.com/kim/ semanticannotation.html [11] Goutam Kumar Saha, ACM Ubiquity, v.8, 2007 [12] Matthew Horridge, Holger Knublauch, Alan Rector, Robert Stevens, Chris Wroe, “A Practical Guide To Building OWL Ontologies Using The Prot´eg´e-OWL”, Plugin and CO-ODE Tools Edition 1.0, August 27, 2004. [13] Ian Horrocks, Ontologies and the Semantic Web, communications of the acm|december 2008 | vol. 51| no. 12. [14] http://www.w3.org/TR/owl-guide/ [15] Pablo Castells, Miriam Ferna´ndez, and David Vallet An Adaptation of the Vector-Space Model for Ontology-Based Information Retrieval, IEEE Transactions On Knowledge And Data Engineering, sVol. 19, No. 2, February 2007.
Velammal College of Engineering and Technology, Madurai
Page 81
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
An Enhancing the Life Time of Wireless Sensor Networks Using Mean Measure Mechanism P.Ponnu Rajan #1, D.Bommudurai*2 #
Department of Computer Applications, K.L.N College of Information Technology, Pottapalayam, Sivagangai District, Tamilnadu, India. 1
[email protected]
*
Department of Computer Science and Engineering, K.L.N College of Information Technology, Pottapalayam, Sivagangai District, Tamilnadu, India. 2
[email protected]
Abstract— Wireless sensor Network function with many routing protocols from the LEACH to High power short distance routing protocol. Still there are some unresolved issues in routing protocol in wireless sensor networks. In this paper the proposed algorithm is Distance based energy aware routing named as DER. The distance based Energy aware routing protocol that selects the route to the base station based on mean measure and the node that is nearest to the base station with in the radius. The energy of the node with in the range is calculated as mean measures or threshold. By comparing the threshold with the energy level of the node the transmission is take place to the base station. The distance-based energy aware routing will efficiently reduces the time delay, number of nodes and energy consumption in WSN. The energy consumption of the node in the network is well balanced and prolonging the survival time of the wireless sensor network. Keywords— Wireless Sensor Networks; Mean measure; Energy aware routing protocol
I. INTRODUCTION In wireless sensor network [1, 2] the sensors are deployed randomly or densely in deterministic or non-deterministic environment. The sensors node is small in nature with limited power. The energy is vital for wireless sensor network because it is not easy to recharge the nodes because it has been deployed in a hostile environment. So conservation of the energy efficiently is the primary motive in a wireless sensor networks. The sensor node consumes energy for sensing the activity, processing and transmitting. The energy required for transmitting the data is high compared to sensing and processing the data. The energy [3,4] spent for transmitting a single bit of data over 100 meters is equal to processing 3000 instructions. WSNs have been widely employed in several applications such as habitat monitoring, disaster supervising, defence application, and commercial application. An important challenge in designing of a wireless sensor network is very limited bandwidth and energy than in wired network environment. The innovative techniques are needed to eliminate energy inefficiency that shortens the lifetime of the network and efficient use of the limited bandwidth. So it is highly desirable to find new methods for energy efficient route
Velammal College of Engineering and Technology, Madurai
discovery and relaying of data from sensor node to the base station. DER is the significant energy aware routing mechanisms. In this scheme, the sensed data are transmitted to the base station based on assumed transmission range by the sensor nodes. In the transmission range, the energy level of the sensor nodes are aggregated together and mean measure is calculated. The target node is finding based on the criteria such that the node which one is nearest to the base station in the range. And the energy level of the node is greater than the threshold level. This scheme effectively reduces the delay and save network energy compared to some of the previous schemes. This protocol highly useful in military based application especially in cross border terrorisms for detecting an object .The detected events are it quickly to the base station for further processing or actions. The remainder of this paper is organized as follows. In section 2, some previous related works, which motivate the research, are briefly reviewed and discussed. The design of the proposed protocol DER is detailed in section 3. The algorithm specified in the section 4 and section5 shows the performance evaluation of the protocol. II. RELATED WORKS Low Energy Adaptive Clustering Hierarchy [5,6] (LEACH) is operated on two phases, the setup phase and steady phase. In the setup phase, the clusters are organized and cluster heads are selected. In the steady phase, the actual transfer to the base station takes place. LEACH randomly selects a few sensor nodes as cluster heads and rotates its role to evenly distribute the energy load among the sensor in the network life time. It holds some of the weaknesses are lots of cluster heads are selected and cluster heads are not selected in a distributed manner leads to unbalanced energy consumption of the nodes and prolonging lifetime of the network get shortly reduced. HPSD [7,8] (High power short distance protocol) is a nodebased protocol that selects route to the base station based on closest node that has the highest battery power relative to its surrounding neighbours. When an event occurs the source node sends the data to the base station using the shortest path. It consists of 10 randomly
Page 82
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
deployed sensor nodes. The base station is fixed and located in the left of middle corner. Initially, node N2 selects the next neighboured node (Node N7) that has the shortest distance to the base station, because all nodes have the same power level. Hence the path will be N7, N6, N9, N11, N13, and N15. Later on that node N7 changes its power level from high to medium due to participating in some routing task. And let node N3 has new data to send to the base station. Then this source node has three possible neighbours to use (Node5, Node7 and node4). That Node7 already participated in the previous routing and power level is medium. Therefore the source node will select either Node 5 or Node 4 because it posses high energy. The process is repeated until it reaches the base station. It posses some of the drawbacks are time delay and more number of nodes required for transmission and energy consumption for more node processing and additional overhead for sending control packets to its neighbours. N1
N4
N13
N11 N9 N6
N15
N7
the base station. Finally the base station calculates distance based on time and speed .The summarized data’s are broadcast it again to all nodes in the network by the base station. Equation for calculating the Distance: Distance= (Speed*Time) Where • Distance=Distance between the Node and the Base station. • Speed=Speed of wave in air. • Time=Time elapsed of a message to reach from the base station to the node. Constructing route table: After identification of node position, the base station broadcast the routing information. The routing table is maintained at each node for taking routing strategy. The routing table contains three fields such as Node-ID, energy level of the node, distance from the base station. The routing strategy is taken by analyzing the above fields in the routing table. The Table I shows the routing table maintained at each node in the sensor network.
N 2
TABLE I ROUTING TABLE
Node-ID N10 N1 4
N 12
N8
N5 N3
Base station
Sensor Field
Fig.1. HPSD Protocol operation
III. PROPOSED METHODOLOGY W. DER: Protocol Operation The Distance based Energy Aware Routing protocol operations are to increase the node lifetime by distributing energy load among sensor nodes. This protocol operated on three phases: 1) Initialization phase 2) Forwarding phase 3) Updating phase 1) Initialization Phase: Once the sensors are deployed in the environment, it is not so easy to re deploy the sensor in the same field. The energy is vital for sensor to survive in the environment. So the energy should be conserve efficiently in the sensing field. Identification of node position: The sensor deployed in the environment randomly. To know its position first the base station broadcast [7] the initialization packet to all the nodes in the sensor network. On receiving such packet, it calculates the time taken to reach every node in the network. Each node in the network stored its time. Then the node transmits its attributes (N_ID, Time and Energy) to
Velammal College of Engineering and Technology, Madurai
N1 N2 ,, ,, Nn BS
Distance
25Meters 37Meters ,, ,, NMeters -
Energy
1Joules 0.97Joules ,, ,, NJoules High Energy
2) Forwarding Phase: After the event triggered by the sensor nodes, the source node route the packet to the base station by considering the routing table. Successive Transmission: In successive transmission, the energy level of the sensor nodes gets changes in the sensor field. Here the source node takes the transmission range of 50m.The target node is predicted based on the criteria: a) The node that is nearest to the base station from the nodes with in the range. b) The Energy level of the node should be greater than the threshold level. Calculation of Threshold level: The energy level of the nodes with in the range is calculated on the basis of mean measures. n ⎛ EN (i ) + EN (i +1) + .......EN (n ) ⎞ ⎟⎟ Th = ∑ ⎜⎜ i =1 ⎝ No. of nodes in the range ⎠
Where, • •
Th is the Threshold level. EN(i) is the Energy level of node.
Page 83
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
If the energy level of the node greater than the threshold level, then the data to be transmitted to the node. Suppose if the energy level of the node lesser than the threshold level, then previous least distance node from the routing table should be taken and compared with its threshold level. If the condition is met, then make transmission to the node. During the successive transmission, the energy level of nodes gets changed. Suppose the Node N2 act as source node after sense the event. It takes the transmission range of 50 meters. The Node N4, N5 and N7 are located with in the range and mean threshold is calculated. Then Node N2 decided to take transmission to N7 that is least distance to the base station but the energy level is greater than the mean threshold level .so it act as target node. Next the N7 take transmission range of 50mts.The Node N6, N8, N11, and N10 are located with in the range and mean threshold is calculated. The Node N7 transmits the data to Node N10 that is least distance to the base station, but the energy level of the node lesser than the mean threshold. So it takes the previous least distance node N11 and compares energy level with the mean threshold. Suppose if the energy level is more than the mean threshold takes it as target node N11. The same procedure is repeated by N11 to transmit the data to Node 15 and finally to the base station. Sensor Field N14
N4 N11
N9
N15
N2
N 7
N10
N1
N 6 N8 N12 Base station
N5
N3
N13
Sensed Data
Node-ID
Target Node Energy
Fig.3. DER Updating Phase
IV. DER IN DETAIL The proposed algorithm is Distance based Energy aware routing Protocol. The procedure for forwarding phase is after the event triggered the source node takes the transmission range of 50 meters. From the source node, the transmission range of 50 meters should be taken. Within that range, find the node, which is nearest to the base station, and compare that node’s energy level with the threshold level. Suppose if energy of the node is lesser than the threshold level, then the previous least distance node from the routing table is taken and compared with the threshold level until the node energy is greater than the threshold level. If the energy of the node is greater than the threshold level, then take that node as target node. The process is repeated until the information reaches the base station. 1. Const Tr=50; 2. For each (Node n in Transmission range) 3. Loop beginning 4. If (n. Energy level is not same) //Node energy level is not uniform with in the range 4. Get_least_distance_node ( ); 5. End If 6. Loop End //If the node is least distance to the base station 7. Get_least_distance_node ( ) 8. Begin 9. For each (Node n in Transmission range) 10. Loop Beginning 11. Nn=Minimum_dist (); 12. Get_target_node (Nn); 13. Loop End 14. End
Fig.2. DER Forwarding phase
3) Updating Phase: After every transmission of information from source to base station, the energy level of the node in the sensor fields get changed. So it is essential for updating routing information at each node after every transmission of packets. Here the energy required for updating the routing information at each node lies on the hands of the base station. a) During transmission, the target nodes in every range not only send the sensed data in addition to that it sends energy level and Node-ID to the base station. b) The base station broadcast the messages to all nodes in the network. c) After receiving such message the nodes updates its routing information. The above process is repeated after every transmission to update the routing table at every node in the sensor network.
Velammal College of Engineering and Technology, Madurai
// If the Mean threshold greater than basic threshold and // Energy level of node greater than the mean threshold 15. Get_target_node (Nn) 16. Begin 17. For each (Node n in Transmission range) 18. Loop Beginning 19. EN= EN1+EN2… +ENn; 20. MTh=EN/n; 21. Const BTh=0.002j; 22. Loop End 23. If ((MTh >= BTh) && (ENn >= MTh)) 24. target node=Nn //Route to target node 25. Else 26. Get_next_minimum_dist (); 27. If (target. BS==true)
Page 84
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
28. //Route to base station 29. End If 30. End If 31. End //Finding the Minimum Distance Node 32. Minimum_dist () 33. Begin 34. For each (Node n in Transmission range) 35. For each (Node m in Transmission range) 36. Loop Beginning 37. If (N [n] < M [n]) 38. t=N[n] 38. N [n]=M [n] 39. M [n]=t 40. End If 41 Loop End 42. End 43. Get_next_minimum_dist () 44. Begin 45. For each (Node m in Transmission range) 46. Loop Beginning 47. Nn=M [n--] 48. Get_target_node (Nn) 49. Loop End 50. End
Y. B. Comparing HPSD with DER Table IV shows the comparison between the High Power Short Distance Protocol and Distance based Energy Aware Routing Protocol. TABLE III PARAMETERS USED IN THE SIMULATION
Parameters Number of nodes Network dimension Initial Energy Dead node Sense radius Deployment Base station location Data packet size Transmission Range Topology
Values 50 50*50 m2 1 Joule < 0.002 Joule 20 Meters Random Middle of left side 53 bytes (variable length) 50 Meters Static topology TABLE IV
S.No. 1)
HPSD Source node selects the target node from the nearest neighbor nodes.
2)
Energy level mention as High, Medium, and Low Energy level. More number of nodes required for transmitting a packet
Fig. 3. Pseudo code of DER
In the fig.3, the pseudo code describes the forwarding phase in DER. The Table II shows the variable used in the pseudo code. TABLE II VARIABLE MEANING
Name Tr Nn EN1... ENn EN BTh MTh BS
Usage Transmission range Minimum distance node Energy level of nodes in the range Total energy of the nodes in the range Base Threshold Mean Threshold Base Station
4)
5)
Time delay is more for transmitting a packet.
DER Select the target node based on the criteria: a) In the range, Node that nearest to base station. b) The node holds the energy level greater than the threshold level. Threshold Energy level is calculated based on mean measure. Less number of nodes required for transmission, due to assumed transmission range. Time delay is less for transmitting a packet.
HPSD VERSUS DER
X. A. Simulation environment and parameters Using NS2 [Network Simulator] to simulate DER Protocol from the following aspects: total energy consumption [9, 10, 11, and 12] of the node in the network and the Time delay in data transmission. The parameters used in the simulation are set as shown in Table III. It is assumed that the transmission frequency range is set dynamically in the applications
Z. C. Simulation Result Fig.4 shows the simulation results for total energy consumption for three compared schemes. The energy consumption of DER protocol is the least among three protocols (DER, LEACH and HPSD). As LEACH protocol adopts one- hop communication, the death of nodes will lead to the increasing of communication distance, which causes the energy consumption of LEACH protocol increase fast after 300s.The additional cost is required by transferring of control packets to the neighbourhood node leads to the energy cost of HPSD is more than the DER protocol. Figure.5 shows the simulation results of the time consumption for two schemes. In the HPSD Protocol, the time consumption for transmitting the packet is high compared to DER. Because the number of nodes required for transmitting a
Velammal College of Engineering and Technology, Madurai
Page 85
V. PERFORMANCE EVALUATION In this section, it describes the simulation environment and its parameter and simulation result.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
packet is more compared to DER. In DER protocol by taking the assumed transmission frequency the number of nodes required to reach the base station is shortly reduced lead to reduce the time delay in the sensor network.
Fig.4. Average Energy Consumption
Fig.5. Time Consumption
VI. CONCLUSION & FUTURE WORK The Distance based Energy Aware Routing protocol (DER) selects routes based on energy aware of the nodes in the network by mean measure approach. The distance based energy aware routing efficiently reduces the time and energy in WSN. The energy consumption of the node in the network is well balanced and prolonging the survival time of the wireless sensor network. The performance of DER is analyzed through simulation, the result show that DER can increase the lifetime of the network compared to other conventional protocol. For future work dynamic readjustment of the transmission frequency of the target node, and it leads to reduce the time and energy of the nodes in the sensor networks. In broadcasting of packets to the nodes, DER needs improvement in security aspects.
[22] Routing in wireless sensor networks. Mayank Saraogi. www.cs.utk.edu/~saraogi/594paper.pdf by M Saraogi – 2005. [23] A survey of wireless sensor network, EURASIP Journal on Wireless Communications and Networking 2005:5, pp 774–788. [24] Zheng, Yugui Qu, Baohua Zhao, Data Aware Clustering for Data Gathering in Wireless Sensor Networks, International Conference on Networks Security, Wireless Communications and Trusted Computing, 2009, vol. 1, pp.192-195, 2009. [25] Shemshaki, M.; Shahhoseini, H.S., Energy Efficient Clustering Algorithm with Direct Path Supports 2009 International Conference on Signal Processing Systems Volume, Issue, 15-17 May 2009 Page(s): 277 – 281. [26] Al-Karaki, J.N.; Al-Malkawi, I.T, On energy efficient routing for wireless sensor networks, International Conference on Volume, Issue, 16-18 Dec. 2008 Page(s): 465-469. [27] Hey, L.A.Power Aware Smart Routing in Wireless Sensor Networks Next Generation Internet Networks, 2008. NGI 2008 Date: 28-30, Apri2008, On page(s): 195-202. [28] Weizheng Ren, Lianming Xu, Zhongliang Deng, A LowPower Routing Protocol for Wireless Sensor Networks Based on Data Aggregation, International Symposium on Intelligent Information Technology Application Workshops, 2008 IEEE. [29] Fatma Bouabdallah, Nizar Bouabdallah, and Raouf Boutaba, On Balancing Energy Consumption in Wireless Sensor Networks, IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 58,NO. 6, JULY 2009. [30] Hong Luo; Luo, J.; Liu, Y.; Das, S.K, Adaptive Data Fusion for Energy Efficient Routingin Wireless Sensor Networks, IEEE Transactionson Volume 55, Issue 10, Oct. 2006 Page(s): 1286 – 1299. [31] Asudeh, A, Movaghar, A., MEHR: Multi-Hop EnergyAware Hierarchical Routing for Wireless Sensor Network Onpage(s): 1-6. [32] Fariborzi, H.; Moghavvemi, M., EAMTR: energy aware multi-tree routing for wireless sensor networks Communications, IET, Volume 3, Issue 5, May 2009 Page(s): 733-739. [33] WeizhengRen, LianmingXu, ZhongliangDeng, A LowPower Routing Protocol for Wireless Sensor Networks Based on Data Aggregation Intelligent Information Technology Application Workshops, 2008. IITAW '08. Date: 21-22Dec.2008, Onpage(s): 661-664.
REFERENCES [20] Routing Techniques in Wireless Sensor Networks: A. Survey. Jamal N. Al-Karaki. Ahmed E. Kamal. – 2004. [21] Energy Conservation in Wireless Sensor Networks: a Survey. Giuseppe Anastasi, Marco Conti, Mario Di Francesco, Andrea Passarella, Volume 7, Issue 3 (May 2009) Year of Publication: 2009 Pages 537-568.
Velammal College of Engineering and Technology, Madurai
Page 86
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Study of Similarity Metrics for Genomic Data Using GO-Ontology Annalakshmi V#1, Priyadarshini R#2, Bhuvaneshwari V#3 #
School of Computer Science and Engineering, Bharathiar University Coimbatore-641 046, Tamil Nadu, India 1
[email protected] [email protected] 3 [email protected]
2
Abstract— Bioinformatics is the science of managing, mining and interpreting information from biological sequences and structures. The Gene Ontology (GO) is a major bioinformatics initiative to unify the representation of gene and gene product attributes across all species. Similarity is quantity that reflects the strength of relationship between two objects or two features. In this paper, a study on similarity metrics for genomic data is done. The modeling of data for measuring the measure is explained in detail. The comparison of traditional similarity and ontology based similarity measures are discussed and compared. Keywords— Bioinformatics, Gene Ontology, Similarity Measures, GO terms, genes.
I. INTRODUCTION Data mining, the extraction of hidden predictive information from large databases, is a powerful new technology with great potential to help companies focus on the most important information in their data warehouses [1]. Data mining techniques are the result of a long process of research and product development. Data mining is a component of a wider process called Knowledge discovery from databases. Bioinformatics is the science of organizing and analyzing biological data that involves collecting, manipulating, analyzing, and transmitting huge quantities of data. Bioinformatics and data mining provide exciting and challenging researches in several application areas especially in computer science. Bioinformatics is the science of managing, mining and interpreting information from biological sequences and structures [2]. Data are collected from genome analysis, protein analysis, microarray data and probes of gene function by genetic methods. The Gene Ontology (GO) is a major bioinformatics initiative to unify the representation of gene and gene product attributes across all species. The ontology covers three domains such as, cellular component, molecular function and biological process. GO is used to explore and analyze the data by using the GO term object. A gene is the basic unit of heredity in a living organism. All living things depend on genes. Genes hold the information to build and maintain an organism's cells and pass genetic traits to offspring. GO represents terms within a Directed Acyclic Graph (DAG) consisting of a number of terms, represented as nodes within the graph,
Velammal College of Engineering and Technology, Madurai
connected by relationships, represented as edges. Terms can have multiple parents, as well as multiple children along the \is-a' relationships, together with part-of relationships [6]. Similarity is defined as the ratio of the shared information between the objects to the total information about both objects [18]. Measure of semantic similarity for the knowledge component of bioinformatics resources should afford a biologist a new tool in their collection of analyses. The similarity has been computed by four general approaches: the set-based approach, the graph-based approach, the vectorbased approach and the term based approach. In the set-based approach an annotation is viewed as a ‘bag of words’. Two annotations are similar if there is a large overlap between their sets of terms. A graph-based approach views similarity as a graph-matching procedure. Vector-based methods embed annotations in a vector space where each possible term in the ontology forms a dimension. Term-based approaches compute similarity between individual terms and then combine these similarities to produce a measure of annotation similarity [19]. In this paper we have done a detailed study on the similarity metrics using GO ontology for genomic and proteomic databases using annotations. The paper is organized as follows. Section 2 provides the literature study of the various semantic similarity measures. Sections 3 describe similarity metrics using gene ontology. In section 4 provides modeling of dataset for similarity measure. The implemented results for the yeast dataset are analyzed and validated within the go ontology functionalities. The final section draws the conclusion of the paper. II. REVIEW OF LITERATURE The literature related to the various similarity measures based on the GO ontology is discussed. Semantic Similarity measures have been proposed for comparing concepts within ontology. Lord [6] has proposed a semantic similarity measure based on the information content of each term. This is defined as the number of times each term, or any child term, occurs in the corpus, which is expressed as a probability. Resnik [3], [4] developed a measure of semantic similarity for "is-a" ontology based on the information content of the lowest common ancestor (LCA) of two terms. The more frequently a term occurs, i.e., the higher its probability of occurring, the
Page 87
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
lower its information content [5]. One node based approach to determine the conceptual similarity is called the information content approach. If the LCA of two terms describes a generic concept, these terms are not very similar and this is reflected in a low information content of their LCA. This measure considers how specific the LCA of the two terms is but disregards how far away the two terms are from their LCA [5]. Lin [Lin, 1998] has presented a semantic similarities of two GO terms based on the annotation statistics of their common ancestor terms [9]. Jiang and Conrath, [1997] proposed a semantic similarity measure based on hybrid approach, i.e. it combines information content and conceptual distance [10]. Lord et al. Resnik, Lin, and Jiang's have provided similarity methods based on the information content of the GO terms. The two approaches target semantic similarity from quite different angles. Jiang and Conrath have proposed the edge-based distance method is more intuitive, while the node-based information content approach is more theoretically sound. Both have inherent strength and weakness [7]. Similarity measures that satisfy the only condition of reflexive, i.e., every object are most similar to it. Chabalier and Mosser have presented the vector Space model measures, they consider only m*n matrix representation, where m is the total number of genes in the corpus and n is the total number of GO terms. Each row in the matrix represents a gene vector of its annotations. Each vector is binary valued, with 1 represents the presence of the GO term in the gene’s annotation and 0 representing its absence [11]. The major contribution of this approach is the possibility of using a weighting scheme over the annotation.Chabalier and Mosser have proposed the Cosine similarity; the measure can be calculated by using the vector for each gene in the pair. This similarity first generates a weight, for each GO term based on the frequency of its occurrence in the corpus [20]. Huang et al developed the Kappa statistics are used to measure cooccurrence of annotation between gene pairs. Kappa statistics are used for continuous and non-categorical data. It can specifically detect the gene-gene functional relationships [12]. Lee et al has proposed the Term Overlap (TO) Measure score for two genes are calculated as the number of terms that occur in the intersection set of the two gene product annotation sets. The Normalized Term Overlap (NTO), in which the term overlap score is divided by the annotation set size for the gene with the lower number of GO annotations [13]. III.
SIMILARITY METRICS USING GENE ONTOLOGY
measures such as, Resnik, Lin’s, Jiang and Conrath measure. These measures are fully focused on the Information Content (IC). Group wise similarity is calculated among the group of genes. This method consists of the following measures such as, Cosine similarity, Kappa Statistics and Term Overlap (TO). These measures consider only matrix representation of gene products. AA. Node-based (Information Content) Approach One node based approach to determine the conceptual similarity is called the information content approach (Resnik 1992, 1995). This method contains the following measures such as, Resnik, Lin’s, Jiang and Conrath measure. 1) Lord Measure: Lord measure calculates the similarity between two terms by using only the IC of the t1: The information content of a GO term ti is:
Eq., (1) Where p(ti) is the probability of a term occurring in the corpus:
Eq., (2) Where the corpus is the set of annotations for all genes under consideration. "Root" represents one of the three root ontology terms and freq (root) is the number of times a gene is annotated with any term within that ontology. freq(ti) is given by:
Eq., (3) The Lord measure calculates the information content of term ti. 2) Resnik’s Measure: Resnik's measure calculates the similarity between two terms by using only the IC of the lowest common ancestor (LCA) shared between two terms t1 and t2:
Eq., (4)
Gene ontology is a collection of controlled vocabularies that describes the biology of a gene product. It consists of approximately 20,000 terms arranged in three independent ontology: Biological Process, Cellular Component, and Molecular Function. Similarity measures are used to find the similarity between genes. We can measure the similarity in two ways. There are Pair wise similarity and Group wise similarity. The pair wise similarity measures are calculated by using gene pairs. This method contains the following
Where LCA denotes the lowest common ancestor between ontological terms T1 and T2. This measure only accounts for the commonality between terms.
Velammal College of Engineering and Technology, Madurai
Page 88
3) Lin’s Measure: Lin's measure of similarity takes into consideration the IC values for each of terms t1 and t2 in addition to the LCA shared between the two terms and is defined as follows.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Eq., (5) Which has the advantage that it maps onto values on the interval [0, 1] unlike Resnik's measure which maps onto the interval [0, ∞]. Lin's measure also accounts for both the commonality and difference between terms. Resnik's measure does have the desirable property that terms close to the root of the ontology have a low similarity however. This is not the case for Lin's measure. 4) Jiang & Conrath Measure: Jiang and Conrath proposed an IC based semantic distance, which can be transformed into a similarity measure.
Eq., (6) This measure considered probability of term t1, t2 and LCA of the two terms. The highest score for Lin and Jiang is 1, and Resnik's measure has no upper bound. 5) Term Overlap: The term overlap score for two genes is then calculated as the number of terms that occur in the intersection set of the two gene product annotation sets.
Eq., (7) Normalized term overlap (NTO), in which the term overlap score is divided by the annotation set size for the gene with the lower number of GO annotations.
Eq., (8) Traditional cardinality-based similarity measures such as Jaccard and Dice [21] are computed similarly to NTO, but use the union or sum, respectively, of the two gene annotation set sizes as the normalizing factor. BB.
absence (resp. presence) of a term (along a particular dimension) in an annotation. This has the advantage that standard clustering techniques on vector spaces such as kmeans can be applied to group similar terms. What is required is a means of measuring the size of vectors. This can be achieved by embedding terms in a metric space (usually Euclidean). The most common method of measuring similarity between vectors of terms is the cosine similarity
Eq., (9) Where vi represents a vector of terms constructed from an annotation (group of terms) Gi. |·| corresponds to the size of the vector and • corresponds to the dot product between two vectors. The source of descriptiveness, commonality and difference is the same as the situation for set-based approaches. CC.
Edge-based (Distance) Approach
The edge based approach is a more natural and direct way of evaluating semantic similarity in a Go Ontology. It estimates the distance (e.g. edge length) between nodes which correspond to the terms/genes being compared. Given the multidimensional concept space, the conceptual distance can conveniently be measured by the geometric distance between the nodes representing the concepts. Obviously, the shorter the path from one node to the other, the more similar they are. In a more realistic scenario, the distances between any two adjacent nodes are not necessarily equal. It is therefore necessary to consider that the edge connecting the two nodes should be weighted. Most of these are typically related to the structural characteristics of a hierarchical network. Some conceivable features are: local network density (the number of child links that span out from a parent node), depth of a node in the hierarchy, type of link, and finally, perhaps the most important of all, the strength of an edge link. This method contains the following measures. Sussna (1993) considered the first three factors in the edge weight determination scheme. The weight between two nodes c1 and c2 is calculated as follows:
Eq., (10) Given
Vector-based Approach
Vector-based methods embed ontological terms in a vector space by associating each term with a dimension. Usually a vector is binary consisting of 0's and 1's where 0 denotes the
Velammal College of Engineering and Technology, Madurai
Eq., (11)
Page 89
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
where ®r is a relation of type r, ®r' is its reverse, d is the depth of the deeper one of the two, max and min are the maximum and minimum weights possible for a specific relation type r respectively, and nr(x) is the number of relations of type r leaving node x. In determining the overall edge based similarity, most methods just simply sum up all the edge weights along the shortest path. To convert the distance measure to a similarity measure, one may simply subtract the path length from the maximum possible path length (Resnik 1995):
Eq., (12) Where dmax is the maximum depth of the taxonomy, and the len function is the simple calculation of the shortest path length (i.e. weight = 1 for each edge). DD. Set based Approach Set based methods for measuring the similarity of annotations are based on the Tversky ratio model of similarity, which is a general model of distance between sets of terms. It is represented by the formula
EE. Graph based Approach Ontology is a directed, acyclic graph (DAG) whose edges correspond to relationships between terms. Thus it is natural to compare terms using methods for graph matching and graph similarity. We may consider the similarity between annotations in terms of the sub-graph that connects terms within each annotation. Annotation similarity is then measured in terms of similarity between two graphs. Graph matching has only a weak correlation with similarity between terms. It is also computationally expensive to compute, graph matching being an NP-complete problem on general graphs. 1) Improving Similarity Measures by Weighting Terms: Set, vector and graph-based methods for measuring similarity between annotations can be improved by introducing a weighting function into the similarity measure. For example, the weighted Jaccard distance can be formulated as:
Eq., (16) Where, as before, G1 and G2 are annotations or sets of terms describing data (e.g. a gene product), Tx is the xth term from a set of terms and m(Tx) denotes the weight of Tx.
Eq., (13) Where G1 and G2 are sets of terms or annotations from the same ontology and ƒ is an additive function on sets (usually set cardinality). For α=β=1 we get the Jaccard distance between sets:
Eq., (14) And for α=β=1/2 we get the Dice distance between sets
Instance-Based Weights: One approach to assigning weight to an ontological term is to measure how informative a term is in describing data. A method of measuring information is to analyze a term's use in a corpus against the general use of ontological terms in the same corpus. Information is measured using the surprisal function:
Eq., (17) Where p(Ti) corresponds to the probability of a term Ti or its taxonomic descendants occurring in a corpus. Other Weighting Approaches: Other measures of information can be used not necessarily relying on corpus data. One measure [14] relies on the assumption that how the ontology is constructed is semantically meaningful:
Eq., (15) In this situation the source of descriptiveness of an annotation is its set of terms. Each term and its set of associated instances are considered independent of other terms. The commonality and difference between annotations is modeled as set intersection and difference of sets of terms respectively. Setbased approaches return a similarity of zero if they do not share common terms ignoring the fact that terms may be closely related. Because of the atomic nature of terms in the set-based approach the monotonicity property does not apply .
Where desc(Ti) returns the number of descendants of term Ti and numTerms refers to the total number of terms in the ontology.
Velammal College of Engineering and Technology, Madurai
Page 90
Eq., (18)
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
FF. Term-based Approaches In term-based approaches similarity between pairs of terms from each annotation are computed. These weightings are then combined in order to characterize the similarity between annotations as a whole. There are several ways to combine similarities of pairs of terms such as the min, max or average operations. Term-based approaches depend on a function s(Ti, Tj) where Ti and Tj are terms from two annotations G1 and G2 respectively. s(Ti, Tj) provides a measure of distance/similarity between these two terms. Once distances has been measured between all possible pairs of terms they are then aggregated using an operation such as max or the average of all distances. For example:
Graphical Measures of Term Similarity: The simplest approach to measuring similarity between ontological terms using the graph structure is to measure the shortest path distance between terms in the graph [15, 16]. A more refined use of graph distance as a basis for a measure of term similarity is found in the Wu-Palmer measure of similarity [17]. It uses the idea that the distance from the root to the lowest common taxonomic ancestor (LCTA) measures the commonality between two terms while the sum of the distance between the LCTA and each term measures the difference between two terms. Combining these aspects results in the formula:
Eq., (20)
Eq., (19) More sophisticated term based approaches combine multiple measures of term similarity and aggregate similarity values using more complex functions.
Where T1 and T2 are the two terms being compared, Tlcta is the term that corresponds to the lowest common taxonomic ancestor between T1 and T2. Troot denotes to root node of the ontology (assuming that the ontology has only one root). dist(Ti, Tj) denotes the graph distance between terms Ti and Tj. The 2 * dist(Tlcta, Troot) component of the denominator serves to normalize the measure.
The various approaches with author reference are given in Table 1. TABLE I. COMPARISON OF VARIOUS APPROACHES AND MEASURES
Approach Node-based Approach
Measures The similarity value is defined as the information content value of specific gene pairs.
Author Resnik, Lin’s, Jiang and Conrath. Meeta Mistry and Paul Pavlidis
Edge-based Approach
This method estimates the distance between nodes which correspond to the terms/genes being compared. This method for measuring the similarity of annotations is based on the Tversky ratio model of similarity, which is a general model of distance between sets of terms.
Jiang JJ, Conrath DW
Set-Based Approach
Graph-Based Approach
Term-Based Approach
Vector-Based Approach
This method is measuring in terms of similarity between two graphs, and then measuring similarity between annotations can be improved by introducing a weighting function into the similarity measure This method is measuring similarity of between pairs of terms from each annotation are computed. This method is measuring similarity of matrix representation of genes.
Velammal College of Engineering and Technology, Madurai
Huang da W, Sherman BT, Tan Q, Collins JR, Alvord WG, Lempicki RA, Jeffrey Hau, William Lee, Brendan Sheehan Brendan Sheehan, Aaron Quigley, Benoit Gaudin and Simon Dobson, Lee JH, Kim MH, Lee YJ Rada R, Mili H, Bicknell E, Bletner, Wu-Palmer, Lee JH, Kim MH, Lee YJ Resnik P , Lord PW, Stevens RD ,Meeta Mistry and Paul Pavlidis, Brendan Sheehan
Ref No [6], [4], [3], [13],[19], [21],[22] [7],[8], [10] [12],[18], [19]
[13],[16], [19] [15],[16], [17] [4],[6], [13],[19]
Page 91
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IV.
MODELING DATASET FOR SIMILARITY MEASURE
GO annotation site contains various data sets. Yeast data set is one of the types of data set. We can download the yeast data set form the Go annotation website. Yeast microarray data contains about 6400 Genes (e.g. YALO51W, YALO54C) with their corresponding Yeast values (0.1650, 0.2720). The file “yeastgenes.sgd” was obtained from the GO annotation site. SGDann is a master structure of Yeast microarray data. It contains the parameters namely SGDaspect, SGDgenes, SGDgo. SGDaspect contains the corresponding functionality for the SGDgenes. SGDgenes contains the whole genes. Similarly SGDgo contains the corresponding Goid for SGDgenes. The entire microarray data are mapped with this structure (SGDgenes) and the microarray data is splitted based on the functionality of genes using GO ontology. For each gene the corresponding functionality of genes are found (Biological Process, Cellular Component, and Molecular Function). All the genes with same functionality are grouped for the corresponding Yeast values from the microarray dataset. Yeast data set contains 6400 genes. The sample genes for yeast data set can be shown in Fig. 1.
The similarity measure can be calculated based on the gene pair. So we are taking the gene from the yeast micro array data. Consider one gene name is ‘YKL050C’. This gene contains the following Goids (3674, 5575, 8150). This can be mapped from the entire SGDgenes and find out the goid values. The gene and goid can be shown in Fig. 3.
Fig. 3 Gene for ‘YKL050C’ & Goid values
Each descendant can have their own genes. The genes can be finding out by using the goid values. Descendants contain the corresponding goid for the genes. The respective genes can be shown in Fig. 4.
Fig. 1 Genes from the yeast data set Fig. 4 Genes for the descendant’s goid
SGDann structure is the entire yeast micro array data set. It contains the attributes such as, Database, DB_Object_ID, DB_Object_Symbol, Goid, Aspect, and DB_Object_Type. It can be shown in Fig. 2.
Descendant means children for the respective goid. We can easily get the children (descendants) from the GO ontology using the get descendant method. For example consider one goid ‘3674’. It contains 8675 descendants. The sample descendants can be shown in Fig. 5.
Fig. 2 SGDann Data Structure
Velammal College of Engineering and Technology, Madurai
Page 92
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
methods. Using this method we can’t get the accurate result for the semantic similarity for analyzing gene functionalities for genes. The semantic measures calculated using traditional approaches are shown in Table 2. TABLE II. RESULTS OF TRADITIONAL METHODS
Genes YKL050C
Fig. 5 Descendant for the goid ‘3674’
Similarly we can get the genes from the yeast dataset and can be arranged in the structure. For example we consider the six genes namely YKL050C, YGL114W, YBR241C, YIL166C, YBL095W, YOR390W. And the corresponding goid, descendants can be shown in Fig. 6.
Fig. 6 Multiple Genes & descendants structure
YGL114W
YBR241C
Euclidean Jaccard Cosine Euclidean Jaccard Cosine Euclidean Jaccard Cosine
YKL050C
YGL114W
YBR241C
0 0 0 21.1896 0.7143 0.0061 11.0905 0.7143 0.0020
21.1896 0.7143 0.0061 0 0 0 21.9089 0.8571 0.0065
11.0905 0.7143 0.0020 21.9089 0.8571 0.0065 0 0 0
For the above we consider the genes for ‘YKL050C’, ‘YGL114W’, and ‘YBR241C’ for traditional method. In this table we calculated the values for the Euclidean, Jaccard and Cosine similarity. The GO ontology based semantic similarity measures are calculated using the discussed approaches above. These measures identify the relation among genes based on their children and parents. The Table 3 provides with the result of the approaches based on ontology. In table3 we consider the genes for ‘YKL050C’, ‘YGL114W’, and ‘YBR241C’ for calculating the semantic similarity measures based on the genes functionality. In this table we are used Lin, Jiang and Conrath measures. TABLE III. RESULTS FOR GO ONTOLOGY FUNCTIONALITY
Genes YKL050C Lin Jiang YGL114W Lin Jiang YBR241C Lin Jiang
YKL050C 0 0 1.5728 0.9039 1.5749 0.9037
V.
Fig. 7 Graphical representation for ‘YKLO50C’ gene with 3 Go Terms
A. Experimental Results The traditional method for semantic similarity measure is calculated based on the content comparison. The basic types of traditional methods are Euclidean, Jaccard and Cosine
Velammal College of Engineering and Technology, Madurai
YGL114W 1.5728 0.9039 0 0 1.6721 0.8248
YBR241C 1.5749 0.9037 3.6721 0.8248 0 0
CONCLUSION
In this paper we have done a detailed study of the semantic similarity measures for finding the relationships among genes. The modeling of the dataset for the measure calculation is provided in detail. The results on implementation we have found that gene ontology based similarity measures finds the accurate association of genes based on Go Ontology. And also we compared the traditional measures and gene ontology based similarity measures. We conclude from our study that the gene ontology based semantic measures can be used for accessing genes for various factors than traditional based measures.
Page 93
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
REFERENCES [34] Arun. K. Pujari, Data Mining Techniques, Universities press (India) Limited 2001, ISBN-81-7371-380-4. [35] Daxin Jiang, Chun Tang, and Aidong Zhang, “Cluster Analysis for Gene Expression Data: A Survey, IEEE Transactions on knowledge AND Data Engineering”, Vol 16, No. 11, November 2004. [36] Resnik P: “Using Information Content to Evaluate Semantic Similarity in a Taxonomy” Proc 14th Int'l Joint Conf Artificial Intelligence, 1995, Vol 1, p.448-453. [37] Resnik P: “Semantic Similarity in a Taxonomy: An InformationBased Measure and its Application to Problems of Ambiguity in Natural Language” J Artif Intell Res 1999, 11:95-130. [38] Schlicker A, Domingues FS, Rahnenfuhrer J, Lengauer T: “A new measure for functional similarity of gene products based on Gene Ontology” BMC Bioinformatics 2006, 7:302. [39] Lord PW, Stevens RD, Brass A, Goble CA: “Semantic similarity measures as tools for exploring the gene ontology” Pac Symp Biocomputing 2003:601-612. [40] Jiang JJ, Conrath DW: “Semantic Similarity Based on Corpus Statistics and Lexical Taxonomy” ROCLING X: 1997; Taiwan 1997. [41] Tatusova TA, Madden TL: “BLAST 2 Sequences, a new tool for comparing protein and nucleotide sequences” FEMS Microbiol Lett 1999, 174(2):247-250. [42] Guangchuang Yu: “GO-terms Semantic Similarity Measures” October 28, 2009. [43] Francisco M. Couto, Mário J. Silva, Pedro Coutinho: “Implementation of a Functional Semantic Similarity Measure between GeneProducts” DI–FCUL TR–03–29, 2003. [44] Chabalier J, Mosser J, Burgun A: “A transversal approach to predict gene product networks from ontology-based similarity” BMC Bioinformatics 2007, 8:235. [45] Huang da W, Sherman BT, Tan Q, Collins JR, Alvord WG, Roayaei J, Stephens R, Baseler MW, Lane HC, Lempicki RA: “The DAVID Gene Functional Classification Tool: a novel biological module-centric algorithm to functionally analyze large gene lists” Genome Biol 2007, 8(9):R183. [46] Meeta Mistry1 and Paul Pavlidis*2: “Gene Ontology term overlap as a measure of gene functional Similarity” Bioinformatics 2008, 9:327. [47] Veale N, Seco JHT: “An Intrinsic Information Content Metric for Semantic Similarity in WordNet” ECAI 2004 2004:1089-1090. [48] Rada R, Mili H, Bicknell E, Bletner M: “Development and Application of a Metric on Semantic Nets” IEEE Transactions on Systems, Man, and Cybernetics 1989, 19:17-30. [49] Lee JH, Kim MH, Lee YJ: “Information Retrieval Based on Conceptual Distance in IS-A Hierarchies” Journal of Documentation 1993, 49:188-207. [50] Wu Z, Palmer M: “Verb semantics and lexical selection” In 32nd. Annual Meeting of the Association for Computational Linguistics New Mexico State University, Las Cruces, New Mexico; 1994:133-138. [51] Jeffrey Hau, William Lee, John Darlington: “A Semantic Similarity Measure for Semantic Web Services” Imperial College London, 180 Queens Gate, London, UK. [52] Brendan Sheehan*, Aaron Quigley, Benoit Gaudin and Simon Dobson: “A relation based measure of semantic similarity for Gene Ontology annotations” BMC Bioinformatics 2008. Vol 9,9:468. [53] Chabalier J, Mosser J, Burgun A: “A transversal approach to predict gene product networks from ontology-based similarity” BMC Bioinformatics 2007, 8:235. [54] Popescu M, Keller JM, Mitchell JA: “Fuzzy measures on the Gene Ontology for gene product similarity” IEEE/ACM Trans Comput Biol Bioinform 2006, 3(3):263-274. [55] Lin D: “An information-theoretic definition of similarity” In 15th International Conf on Machine Learning Morgan Kaufmann, San Francisco, CA; 1998:296-304
Velammal College of Engineering and Technology, Madurai
Page 94
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hybrid PSO based neural network classifier and decision tree for brain MRI mining #
Dr.V.Saravanan#1 , T.R.Sivapriya*2 Dr.N.G.P Institute of Technology, Coimbatore, India *
[email protected]
Lady Doak College, Madurai, India [email protected]
Abstract Artificial neural networks have been applied in a variety of realworld scenarios with remarkable success. In this paper, a hybrid PSO based back propagation neural network for classifying brain MRI is proposed. The results show that there is a marked difference while training BPN with PSO. Also a customized PSO is used to train the BPN which again results in better performance compared to the conventional PSO used in training. The Decision tree extracts rules from the trained network that aids in medical diagnosis. The neural model based on Particle Swarm Optimisation was trained with 150 samples (including all patients with mild and severe dementia). Additional hundred samples have been used for validation. The proposed algorithm outperforms the result of conventional training algorithms and is found to have 95% sensitivity and 96% accuracy. The samples were tested with a radiologist and psychiatrist by means of blind folded study. When compared with the experts, the algorithm achieved good accuracy with higher rate of reliability for the assessment of mild and severe dementia. Keywords --- Image mining, Neural networks, PSO, Decision tree
1.INTRODUCTION There has been a growing number of research applying ANNs for classification in a variety of real world applications. In such applications, it is desirable to have a set of rules that explains the classification process of a trained network The classification concept represented as rules is certainly more comprehensible to a human user than a collection of ANNs weights . This paper proposes a fast hybrid PSO based BPN classifier, whose results are mined by decision tree for the diagnosis of dementia. A neural network, by definition, is a massively parallel distributed processor made up of simple processing units, each of which has a natural propensity for storing experiential knowledge and making the knowledge available for use [11]. Neural networks are fault tolerant and are good at pattern recognition and trend prediction. In the case of limited knowledge, artificial neutral network algorithms are frequently used to construct a model of the data.
Velammal College of Engineering and Technology, Madurai
A. Artificial Neural Networks In the last few years artificial neural networks have been applied successfully to a variety of real-world problems. For example, neural networks have been successfully applied in the area of speech generation and recognition, vision and robotics, handwritten character recognition , medical diagnostics [9], and game playing. Neural network method is used for classification, clustering, feature mining, prediction and pattern recognition. It imitates the neurons structure of animals, bases on the M-P model and Hebbian learning rule, so in essence it is a distributed matrix structure. Through training data mining, the neural network method gradually calculates (including repeated iteration or cumulative calculation) the weights the neural network connected. Knowledge can be more easily recognized when presented in the form of images. For example, geophysical and environmental data from satellite photos, Web pages containing images, medical imaging including Computed Tomography (CT), Magnetic Resonance Imaging (MRI), is sources of useful information used in our daily life. B. The Artificial Neural Network Models for Data Mining The center piece of the analytic plan is modeling ANN architectures. Artificial neural network (ANN) modeling has been used for various tasks, including speech recognition, stock market prediction, mine detection, cancerous cell identification, and handwriting recognition. We use the ANN as a computational modeling tool, the center piece of the datamining approaches taken in this application. In practice, however, reports of ANN’s ability to surpass linear models for classification and prediction problems have been inconsistent (Doig et al., 1993; Baxt, 1995; Dybowski et al., 1996; Anderer et al., 1994; Duhet al., 1998), depending on data complexity, evaluation measures, and the user’s familiarity with the data as well as modeling.
Page 95
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
II. BACK PROPAGATION NEURAL NETWORKS The backpropagation algorithm for estimating parameters in neural networks has been the most popular in the medical literature [Reggia, 1993. One of the problems encountered by researchers utilizing the backpropagation algorithm is that low-frequency patterns (or rare patterns) may take a long training time to be recognized, because frequent patterns “dominate” the error.This problem is overcome by faster training using PSO. A hidden layer is used for ANN’s internal transformation: hidden neurons transform the “signals” from the input layer and generate signals to be sent to the next layer, so that MLP is often called the “feedforward” neural network. The input layer consists of input variables plus an input bias, which adjusts the weighted mean of the input variables. All input variables are regressed on the next layer, in the same fashion as multiple regression estimation.
Fig. 1 Backpropagation Network
A. BPN architecture design BPN architecture is selected based on testing and training of 7 different models.The steps are as follows: 1. Create a customized bpn model n1. 2. Train and test with new data 3. Check for the efficiency and ability to recognize new patterns 4. If the model n1 has comparatively higher efficiency and detects new patterns ,it is selected as the ideal model for diagnosis. III. PARTICLE SWARM OPTIMISATION Particle Swarm Optimisers (PSO) are a new trend in evolutionary algorithms, being inspired in group dynamics and its synergy. PSO had their origins in computer simulations of the coordinated motion in flocks of birds or schools of fish. As these animals wander through a three dimensional space, searching for food or evading predators, these algorithms make use of particles moving in an n-dimensional space to search for solutions for a variable function optimization problem. In PSO individuals are called particles and the population is called a swarm [1]. PSO are inspired in the intelligent behaviour of beings as part of an experience sharing community as opposed to an isolated individual reactive response to the environment
Velammal College of Engineering and Technology, Madurai
A. Training the BPN with PSO A Neural Network’s effectiveness is in its weights (assuming the topology is correct) Other methods exist to train NNs, but they can be slow and/or difficult to implement (GA,Simulated Annealing. Every edge (or weight) in the network is an element in our particle. The number of weights in the network equalizes the number of dimensions of the particles in PSO. The fitness value is how well the NN performs on the test data. The algorithms is run for a given number of iterations or until a minimum error is reached as follows: 1. Initialize a population size, positions and velocities of agents, and the number of weights and biases. 2. The current best fitness achieved by particle p is set as pbest. The pbest with best value is set as gbest and this value is stored. 3. Evaluate the desired optimization fitness function fp for each particle as the Mean Square Error (MSE) over a given data set. 4. Compare the evaluated fitness value fp of each particle withits pbest value. If fp < pbest then pbest = fp and bestxp = xp, xp is the current coordinates of particle p, and bestxp is thecoordinates corresponding to particle p’s best fitness so far. 5. The objective function value is calculated for new positions of each particle. If a better position is achieved by an agent, pbest value is replaced by the current value. if fp < gbest then gbest = p, where gbest is the particle having the overall best fitness over all particles in the swarm. 6. Change the velocity and location of the particle based on random number assigned. vid = vid + ϕ1*rnd()*(pid-xid) + ϕ2*rnd()*(pgd-xid); xid = xid + vid; • Where i is the particle,ϕ1,ϕ2 are learning rates governing the cognition and social components • g represents the index of the particle with the best pfitness, and d is the dth dimension. IV. DECISION TREE Decision tree induction is one of the simplest, and yet most successful forms of learning algorithm. It serves as a good introduction to the area of inductive learning, and is easy to implement. A decision tree takes as input an object or situation described by a set of attributes and returns a decision-the predicted output value for the input. The input attributes can be discrete or continuous. For now, we assume discrete inputs. The output value can also be discrete or continuous; learning a discrete-valued function is called classification learning; learning a continuous function is called regression. A. Evaluation of Rule set The constructed rule set is evaluated against the instance test set. Each instance is sequentially tested with rules in the rule set, until it matches one rule antecedent. The instance class is then compared to the matching rule predicted class. When all
Page 96
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
instances have been tested, the reliability of the rule set is computed regarding correctly classified instances and presented in a percentage value. CLASSIFY MRI
CLASS I
CLASS II
4. Test the network 5.Train the same BPN network using PSO 6. Test the network 7.Compare the performance 8. Apply data mining algorithm to produce object association rules from the weight vector of the trained Hybrid PSO based BPN. READ MRI IMAGE
PREPROCESSING
NORMAL
HIGLHLY DEMENTED
LESS DEMENTED
Fig 2. Decision Tree for Mining V. MRI DATASET OASIS provides brain imaging data that are freely available for distribution and data analysis. This data set consists of a cross-sectional collection of 416 subjects covering the adult life span aged 18 to 96 including individuals with early-stage Alzheimer’s Disease (AD). For each subject, 3 or 4 individual T1-weighted MRI scans obtained within a single imaging session are included. The subjects are all right-handed and include both men and women. 100 of the included subjects over the age of 60 have been diagnosed with very mild to mild AD. Additionally, for 20 of the non-demented subjects, images from a subsequent scan session after a short delay (less than 90 days) are also included as a means of assessing acquisition reliability in the proposed study.
FEATURE EXTRACTION
TRAINNG IN ANN
TESTING
CLASSIFY CC,VT,
EXTRACT RULES
Fig 2. System Overview
For each subject, a number of images are taken for analysis, including: 1) images corresponding to multiple repetitions of the same structural protocol within a single session to increase signalto-noise, 2) an average image that is a motion-corrected co registered average of all available data, 3) a gain-field corrected atlas-registered image to the 1988 atlas space of Talairach and Tournoux (Buckner et al., 2004), 4) a masked version of the atlas-registered image in which all non-brain voxels have been assigned an intensity value of 0, and 5) a gray/white/CSF segmented image (Zhang et al., 2001). VI. HYBRID PSO BASED BPN 1. Feature extraction. Segment images into regions identifiable by region descriptors This step is also called segmentation. a. extract cortex b. extract hippocampus c. ventricles d. basal ganglia 2. Feature reduction by PCA 3. Train conventional BPN with features to generate desired output
Fig 3. Brain MRI features Initially the conventional Particle Swarm Optimisation technique with 25 particles is used to train the backpropagation algorithm. Although the convergence was quick, performance improved further when the number of particles is reduced to 10. A. Training
Velammal College of Engineering and Technology, Madurai
Page 97
A multilayered back-propagation neural network was used (37 inputs from each of the 150 adolescents, comprising the input patterns, and the two binary outputs). The network was exposed to the data, and the parameters (weights and biases)
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
were adjusted to minimize error, using a back-propagation training algorithm. . The input layer has 10 neurons, where each neuron represents reduced pixel group. The number of neurons in the hidden layer is calculated based on the following formula: N3 =
((2/3)*(N1))+N2
Rules : IF CC=+++ AND HIP=+++ AND BG=+++ AND VT=++++ THEN DEMENTIA= HIGH IF CC= +++ AND HIP=+ AND BG=+ AND VT=+++ THEN DEMENTIA= MODERATE
N1 represents number of nodes in the input layer; N2 represents number of nodes in the output layer; N3 represents number of nodes in the hidden layer.
IF CC=+ AND HIP= -- AND BG = -- AND VT=+ THEN DEMENTIA= MILD
B. Validation
IF CC= -- AND HIP= -- AND BG = -- AND VT=-THEN DEMENTIA= NO TABLE III RESULTS OF CONVERGENCE
PSO-BPN
BPN
Iterati ons 175
Iterati ons 175
MRI
The ability of the neural network model (a 10 X 7 X 1 structure was used) to perform the classification was examined by setting aside 20% of the patterns (or observations) as validation (or testing) data. In this crossvalidation approach, the training involved repeatedly exposing the network to the remaining 80% of the patterns (training data) for several epochs, where an epoch is one complete cycle through the network for all cases. (Data were normalized before training.) A network trained in this manner is considered generalizable, in the sense that it can be used to make predictions. VII. RESULTS
200
TABLE I COMPARISON OF CONVERGENCE
225
575
BPN
Sensitivity %
85
HYBRID PSOBPN 92
Accuracy % 83 90 Specificity % 87 93 CC- Cerebral cortex HIP= Hippocampus VT – Ventricle BG – Basal Ganglia +++ Æ high increase ;+ Æ less increase -- Æ no increase
275 300 175 200
TABLE 2 COMPARISON OF EFFICIENCY
MEASURE
250 NORMAL
ITERATIONS >2000 800
CUSTOMISED PSOBPN 95 96 97
Velammal College of Engineering and Technology, Madurai
HIGHLY DEMENTED
TECHNIQUE BPN BPN-PSO WITH 25 PARTICLES BPN –PSO WITH 10 PARTICLES
225 250 275 300
GBest 0.247863299824 5176 0.247863290721 40684 0.247863290493 80179 0.247863290484 66602 0.247863290484 66071 0.247863290484 66071 0.247863299824 5176
200
0.247863290721 40684 0.247863290493 80179 0.247863290484 66602 0.247863290484 66071 0.247863290484 66071
200
225 250 275 300 175
225 250 275 300
MSE 0.003164 58/0 0.003122 28/0 0.003033 84/0 0.002969 5 0.002870 65 0.002763 27/0 0.003164 58/0 0.003122 28/0 0.003033 84/0 0.002969 5 0.002870 65 0.002763 27/0
Page 98
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the weights of the BPN correlated with the diagnosis of the experts.
All Data, Blue = known, Red = net 1 0.9 0.8
REFERENCES
0.7 0.6 out
0.5 0.4 0.3 0.2 0.1 0 0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
in
Fig. 3. Trained PSO-BPN
Fig 4. Converging speed of conventional BPN
HYBRID BPN VS BPN 35 30 TIME
25 20 15 10 5 0 1
2
3
4
5
6
7
8
9
10
NO. OF MRI IMAGES
Fig 5. Comparison of converging speed of Hybrid PSO-BPN and conventional BPN
VII. CONCLUSION The hybrid PSO based BPN effectively classifies the brain MRI images of dementia patients. It appears to prove for highly reliable for use by practitioners in the medical field. Mining large database of images poses severe challenges. However, the hybrid classifier is being offered s to overcome difficulties in diagnosis and treatment. The rules mined from
Velammal College of Engineering and Technology, Madurai
[1] Ashish Darbari, “Rule Extraction from Trained ANN: A Survey,” Technical Report, Department of Computer Science, Dresden University of Technology, Dresden, Germany, 2000. [2] Heon Gyu Lee, Ki Yong Noh, Keun Ho Ryu, “Mining Biosignal Data: Coronary Artery Disease Diagnosis using Linear and Nonlinear Features of HRV,” LNAI 4819: Emerging Technologies in Knowledge Discovery and Data Mining, pp. 56-66, May 2007. [3] Cristianini, N., Shawe-Taylor, J. “An introduction to Support Vector Machines”, Cambridge University Press, Cambridge, 2000. [4] Li, W., Han, J., Pei, J., “CMAR: Accurate and Efficient Classification Based on Multiple Association Rules”, In: Proc. of 2001 Interna’l Conference on Data Mining. 2001. [5] Parpinelli, R., Lopes, H. and Freitas, A., “An Ant Colony Algorithm for Classification Rule Discovery”, Idea Group, 2002. [6] F. H. Saad, B. de la Iglesia, and G. D. Bell, “A Comparison of Two Document Clustering Approaches for Clustering Medical Documents”, Proceedings of the 2006 International Conference on Data Mining (DMIN06), 2006. [7] Gerhard Münz, Sa Li, and Georg Carle, “Traffic anomaly detection using k-means clustering”, In Proc. of Leistungs-, Zuverlässigkeits- und erlässlichkeitsbewertung von Kommunikationsnetzen und Verteilten Systemen, 4. GI/ITG-Workshop MMBnet 2007, Hamburg, Germany, September 2007. [8] R. Reed, “Pruning algorithms-A survey,” IEEE Transactions on Neural Networks, vol. 4, pp. 740-747, 1993. [9] Han Jiawei, Micheline Kamber, “Data Mining: Concepts and Techniques,” Morgan Kaufmann Publisher: CA, 2001 [9] Evangelou, I. E.: Automatic Segmentation of Magnetic Resonance Images (MRIs) of the Human Brain using Self-Organizing Maps (SOMs). MS Thesis, Clarkson University, Potsdam, NY, USA (1999). [10] Fayyad, U. M., Piatetsky-Shapiro, G., Smyth, P., and R. Uthurusamy, R., (Eds.). Advances in Knowledge Discovery and Data Mining. AAAI Press, The MIT Press, CA, USA (1996). [11] J. S. D. Bonet. Image preprocessing for rapid selection in “Pay attention mode”. MIT, 2000.. [12] Baladi, P. & Hornik K., “Neural networks and principal component analysis: learning from examples and local minima “, Neural Networks, 2,5358.1989. [13] Hsinchun Chen, Sherrilynne S. Fuller, Carol Friedman, and William Hersh, "Knowledge Management, Data Mining, and Text Mining In Medical Informatics", Chapter 1, eds. Medical Informatics: Knowledge Management And Data Mining In Biomedicine, New York, Springer, pp. 3-34, 2005. [14] S Stilou, P D Bamidis, N Maglaveras, C Pappas, “Mining association rules from clinical databases: an intelligent diagnostic process in healthcare”, Stud Health Technol Inform 84: Pt2. 1399-1403, 2001. [15] T Syeda-Mahmood, F Wang, D Beymer, A Amir, M Richmond, SN Hashmi, "AALIM:Multimodal Mining for Cardiac Decision Support", Computers in Cardiology, pp. 209-212,Sept. 30 2007-Oct. 3 2007. [16] Anamika Gupta, Naveen Kumar, and Vasudha Bhatnagar, "Analysis of Medical Data using Data Mining and Formal Concept Analysis", Proceedings Of World Academy Of Science, Engineering And Technology,Vol. 6, June 2005. [17] Sellappan Palaniappan, Rafiah Awang, "Intelligent Heart Disease Prediction System Using Data Mining Techniques", IJCSNS International Journal of Computer Science and Network Security, Vol.8 No.8, August 2008. [18] Andreeva P., M. Dimitrova and A. Gegov, “Information Representation in Cardiological Knowledge Based System”, SAER’06, pp: 23-25 Sept, 2006. [19] Tzung-I Tang, Gang Zheng, Yalou Huang, Guangfu Shu, Pengtao Wang, "A Comparative Study of Medical Data Classification Methods Based on Decision Tree and System Reconstruction Analysis", IEMS, Vol. 4, No. 1, pp. 102-108, June 2005. [19] M. Zekic-Susac, B. Klicek. A Nonlinear Strategy of Selecting NN Architectures for Stock Return Predictions. Finance, Proceedings from the
Page 99
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
50th Anniversary Financial Conference Svishtov, Bulgaria, 11-12 April, 2002, pp. 325-355. [20] M. Zekic-Susac, N. Sarlija, M. Bensic. Small Business Credit Scoring: A Comparison of Logistic Regression, Neural Network, and Decision Tree Models, Proceedings of the 26th International Conference on Information Technology Interfaces, June 7-10, 2004, Cavtat, pp. 265-270.
Velammal College of Engineering and Technology, Madurai
Page 100
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
GAP: Genetic Algorithm based Power Estimation Technique for Behavioral Circuits Johnpaul C. I1, Elson Paul2, Dr. K. Najeeb3 Department of Computer Science & Engineering Government Engineering College, Sreekrishnapuram Palakkad, India-679517 1
[email protected] [email protected] 3 [email protected]
2
Abstract— The problem of peak power estimation in digital circuits is essential for analyzing the reliability and performance of circuits at extreme conditions. The power estimation problem involves finding input vectors that cause maximum dynamic power dissipation (maximum toggles) in circuits. In this paper, an approach for power estimation for both combinational and sequential circuits at behavioral level is presented. The basic intuition behind the approach is to use the genetic algorithm based approach to find the vector pairs for maximum toggling in the circuit. Peak dynamic power consumption is strongly dependent on the stream of inputs applied to the circuit. Without any information about the input stream, it is impossible to accurately estimate the power consumption of a design. Thus, for a power estimation technique it is necessary to use approximation methods. The proposed technique was employed on the ISCAS'85, ISCAS'89 and ITC benchmark circuits. Keywords— CMOS circuits, power dissipation, power estimation, automatic power vector generation, genetic algorithms
VI. INTRODUCTION The high transistor density, together with the growing importance of reliability as a design issue, has made early estimation of worst case power dissipation (peak power estimation) [1] in the design cycle of logic circuits an important problem. High power dissipation may results in decrease in performance or in extreme cases cause burn out and damage to the circuit. The increased transistor density in the processors can be well defined by Moore's law which describes a long term trend in the computing hardware, in which the number of transistors that can be placed inexpensively on an integrated circuit has doubled approximately in two years. Whenever technology advances with new high performance processors, transistor density will obey Moore's law. As more and more transistors are put on a chip, the cost to make each transistor decreases, but the chance that the chip will not work due to a defect increases. Moreover due to the complexity of their operations their interrelationship between the transistors increases. This will increase the processor complexity both in its design and complexity. In this context it is essential to formulate the
Velammal College of Engineering and Technology, Madurai
power estimation methods that will become more helpful in the phase of designing complex digital circuits. The power is an important factor in designing the circuits [2]. In the case of processors containing billions and billions of transistors, power dissipated is more. Hence power management is essential in designing the circuits of a processor to optimize its performance. As technology advances, the bus width of transistor circuits evolves in the order of 32nm, 22nm, 16nm, 11nm, the implementation of power management techniques is a crucial part in the production of complex hardware circuits. Moreover when the bus width is reduced to 11nm range, the number of transistors on a single die dramatically increases. Another impact due to this phenomenon is that the clock frequency increases exponentially with increase in the number of transistors. Another dilemma due to increase in power is thermal run away. The power in the circuit can be expressed in the form of heat dissipated [3]. Due to increase in the heat, current in the circuit increases which will further increase the heat. This will cause destruction in the circuit. Parameter variations cause many hurdles for the CPU designers. The most critical sources of parameter variations are in process characteristics across dies, supply voltage drop and temperature variation [4]. These are generally known as PVT variations. Parameter variations causes the worst case delay in flow of current and, hence effects performance and power [5]. This will further increase heat. In other words the peak power determines the thermal and electrical limits of components and the system packaging requirements [6]. The peak power consumption corresponds to the highest switching activity generated in the circuit under the test during one clock cycle [7]. The dynamic power dissipated in the combinational portion of the sequential circuit [8] (PR ) can be computed as shown in equation (1):
P R=
1 × V 2DD× f 2
∑ [N
g ×C g
]
(1)
Page 101
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
where the summation is performed overall gates g, and N(g) is the number of times gate g has switched from 0 to 1 or vice versa within a given clock cycle, C(g) is the output capacitance of gate g, VDD is the supply voltage and f is the frequency of operation. In the proposed method the assumption taken for power estimation is that the output capacitance for each gate is equal to the number of fanouts. Therefore, the total switching activity is the parameter that needs to be maximized for maximum dynamic power dissipation. Accurate estimation of maximum power consumption involves finding a pair of input vectors which when applied in sequence, maximize the value of PR given by equation (1), among all possible input vector pairs.
power dissipation in the digital circuits. Rest of the paper is organized as follows. Section 2 describes the related work done in this field of power optimization. Section 3 describes the frame work developed with genetic algorithm approach. It includes algorithm used for optimization and toggle vector generation. Section 4 includes the implementation details and experimental results on ISCAS'85, ISCAS'89 and ITC benchmark circuits. Section 5 concludes the whole work. PREVIOUS WORK There are different techniques that have been developed to solve the problems related to peak power estimation. Each of these methods have their own peculiarities and limitations.
Power estimation of digital circuits can be viewed under various abstraction levels. By carefully analyzing complex circuits under these abstractions [3], it will be more convenient to determine the power dissipation in each stages. The different abstraction levels includes system, behavioral, RT level, gate level, circuit level and device level. Each of these levels contains notable peculiarities of different parameters which include estimation accuracy, speed, estimation capacity, complexity, ease of abstraction etc. A designer can choose any of these abstraction levels for conducting a detailed study of power estimation in a systematic pattern [7]. The figure 1 shows the complexity of power estimation for different abstraction levels.
Fig. 2 Classification of Peak Dynamic Power Estimation (PDPE) methods The figure 2 shows different methods of peak dynamic power estimation. The different models that are used for the power estimation in circuits are SAT based approach, constraint based approach, simulation based approach, ATPG based approach, genetic algorithm based and macro-modeling.
Fig. 1 Power estimation vs. different abstraction levels VII.
A. OUTLINE AND CONTRIBUTION OF THIS PAPER
This paper proposes a methodology for peak dynamic power estimation in behavioral circuits using genetic algorithm [9]. The paper proposes the generation of an intermediate file from Verilog input file which will finally fed into the proposed algorithm for optimization process. This optimization process is to obtain an optimized pair of input vectors that produce maximum toggle in the corresponding circuits. This input vectors can be otherwise called as Genetic Power Virus. Power virus in the sense that it will result in the maximum
Velammal College of Engineering and Technology, Madurai
Recent advances in Boolean satisfiability (SAT) models [10] and algorithms have made it tempting to use satisfiability based techniques in solving various VLSI design related problems such as verification and test generation [11]. Usually SAT solvers are used to handle zero-one integer linear programming problems [11]. The constraints are typically expressed in conjunctive normal form (CNF) [12]. This method works well for small and medium sized circuits. But when the size of the circuit increases it will work slower. Simulation based power estimation method can be applied after completing the design of a circuit [13]. It provides an accurate way to estimate power consumption of the design because the power estimation based on the method which reflects the actual design behavior. Since simulation happens later in the design cycle, this power estimation is generally used to verify the power consumption of a device on board. This method is also used as a tool to estimate the power
Page 102
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
dissipation in the portions of a larger design when integrating smaller designs into larger FPGAs. In the case of gate level approach the simulation can be of the form that different gates are simulated using the programming methods and the test vectors are fed to the circuit as input to analyze the functionality [14]. Automatic test pattern generation (ATPG) based techniques try to generate an input vector pair that produces the largest switched capacitance in the circuit [15]. The power consumption by the vector pair is calculated by counting the total number of circuit elements switched from one state to another [14]. The ATPG based techniques are very efficient and generate a tighter lower bound than that generated by random vector generation. ATPG based techniques can handle only simple delay models such as the zero-delay and unitdelay models [10]. Macro-modeling is a method of inclusion of large number of desired properties into a model, there by the worst case analysis of a complex system can be generated [16]. It can also be used to analyze the possible outcomes form a model. Several macro-modeling methodologies have been proposed in the recent past. The majority of these approaches targets specifically register-transfer level power estimation. The fundamental requisites for a power macro-models are evaluation efficiency and flexibility. In every macro-modeling techniques evaluation efficiency is dictated by the need of performing numerous power estimates while the synthesis tool is used for evaluating design alternatives [16]. Flexibility is imposed by the fact that the same macro can be instantiated in completely different designs or operated under widely varying conditions. Genetic algorithm (GA) based approach of power estimation is to determine the input vector pairs that can provide maximum toggles in the digital circuits [9]. Different design methodologies are considered in this approach. GA efficiently produced promising results in circuit partitioning problem and also the circuit element placement problem [9]. The prime aim of GA is optimization of the results that causes power dissipation and to find out the optimal pair of vectors that produces the maximum toggles in digital circuits [2] [7]. GA based approach is highly efficient in the design phase of complex digital circuits such as processors, co-processors etc. ALGORITHM FOR TOGGLE VECTOR GENERATION A. PROBLEM DEFINITION Given a description of circuit either in gate or behavioral level abstraction. The problem is to find out an optimal input test vector pairs v1 and v2 when applied in sequence, leads to maximum switching activity in the circuit. B. GENERATION OF INTERMEDIATE
FILE FROM VERILOG INPUT
Velammal College of Engineering and Technology, Madurai
The algorithm for the generation of intermediate file from Verilog description described in algorithm 1. The algorithm contains separate functions to isolate the input lines, output lines, assign lines, wire lines from the Verilog file. After the generation of corresponding arrays, the formatted data in the arrays are used for generating the intermediate file. For this, a function called tree() is used. The tree() function will traverse the elements of the assign array and replaces the various operations such as NEGATION, XOR, AND etc. to numerical form. The conversion is on the basis that each operations has given a specific number. The corresponding output is written to an output file, which is further used in this framework for the generation of test vectors. Algorithm 1 (INTERMEDIATE FILE) input: Verilog description of behavioral circuit. output: Numerical representation. assign: Assign statements. queue: Enqueue the assignment statements in the order of execution in the Verilog file. repeat Stmt= dequeue(Queue). Tree(stmt). Convert the statements to numerical representation. until Queue= NULL A typical module of the behavioral Verilog file is given below. The algorithm for the generation of the intermediate file scans the input in such a way that whenever it finds a module call from the main module it will execute that module functions and return to the main module. In this way the modules are interconnected. module PriorityA(E, A, PA, X11), Input[8:0] E, A; Output[8:0] X1; Wire [8:0] Ab, EAb; assign Ab = ~ A; assign EAb = ~ (Ab & E); assign PA = ~ EAb; assign X1 = ~ EAb ^ (9(PA)); endmodule. A behavioral level module of circuit 432b.v is shown above. The algorithm 2 explains the procedure used by the proposed approach. The core functions used in the approach are crossover and mutation. These basic functions drive the algorithm to produce the optimized results. Algorithm 2 (GENETIC ALGORITHM) Create initial population. repeat Select parent1 and parent2 from the population. Offspring= crossover (parent1,parent2).
Page 103
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
are randomly complemented. Mutation may sometimes result in the generation of vectors with better fitness value.
Mutation (offspring). until Stopping criteria is reached.
C . EXPLANATION
OF GENETIC ALGORITHM FOR THE FRAMEWORK
Updation Create Initial Population A population of size P is generated randomly to create the initial population. Here the population elements are the strings of binary numbers containing zeros and ones which are equal to the number of input lines in the circuit [9]. The size of the population is to be large enough so that variance in the strings can be easily formed. From this initial population optimization process is performed for the production of the toggle vectors. Selection of parents Each individual or a string has a fitness value, which is a measure of quality of the solution represented by the individual [18] [19]. The formula which is used to calculate the fitness value of each individual
F i =T i
The creation of two offspring increases the size of the population by two. Since it is necessary to maintain a constant population size, two vectors will need to be eliminated from the population [9]. For this elimination process fitness values are used. In the population there are vectors with the lowest fitness value. Two of such vectors are removed and the newly generated offspring are added to the population. Stopping Criteria Stopping criteria uses a limit value W to determine when the algorithm stops. If there is no improvement after W generations, then the algorithm stops. No improvement means that there are no changes in the maximum fitness value of the population. The final solution is the individual with the highest fitness value [9], [19].
T w− T b 3
(2)
IMPLEMENTATION AND EXPERIMENTAL RESULTS
where Fi represents the fitness value of the individual, Ti represents the toggle count of individual i. Tw represents the smallest toggle count in the population and Tb is the largest toggle count in the population. The fitness values are used for sorting the strings in the ascending order. From the middle of the population size two individuals are selected for performing crossover. Crossover After two parents are selected, crossover is performed on the parents to create two offspring [9]. A string split point is randomly selected, and is used to split each parent string into two. The first offspring is created by concatenating the left half of the first parent and the right half of the second parent, while the second offspring is created by concatenating the left half of the first parent and the complement of the right half of the second parent. Mutation Mutation is the phenomenon that produces sudden change in the bit sequence of each string [9]. The mutation operation is performed on each string in such a way that a bit position is selected randomly and that bit is complemented. Mutation can be designed in another way that from the position say b bits
Velammal College of Engineering and Technology, Madurai
Fig. 3 Toggle vector generation The implementation details can be visualized from figure 3. The input to the test vector generator algorithm is in the form of Verilog file. The genetic algorithm is applied over both gate level circuits and behavioral level circuit. The proposed intermediate file is generated from the Verilog input file. This proposed intermediate file is the key file input to the GA. The above mentioned intermediate file contains the information of fanout and the output lines, input lines, wires etc. of circuit elements. The intermediate file of behavioral level and gate level circuits are entirely different from each other. A gatelevel intermediate file is shown in figure 3. The interpretation
Page 104
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of this file is that the specific circuit described contains total number of 1145 gates. Second line till end contains the interconnection to different gates. The first number in each line determines the type of the gate. A typical gate level intermediate file is given below: 1145 2 333 45 -1 3 43 34 675 -1 7 23 56 763 -1 8 89 47 -1 2 44 231 673 89 -1 2 54 33 43 -1 2 22 66 332 90 -1 3 84 431 221 643 -1 7 432 123 243 745 -1 3 67 -1 2 87 336 89 -1 In the case of behavioral level intermediate file [20], it contains the specific bit positions of each module and also the total bit count in the output. Under genetic algorithm, various operations of each module is simulated and record the changes after a pair of input vector is fed into the genetic algorithm. By calculating the number of bits changed due to the application of second optimized input vector [11], toggle count is obtained. Optimized in the sense that applying various steps of the GA in a specified manner [11]. As toggle count is obtained from this experiments, the power dissipation from each circuit can be understood. From the different experiments conducted on ISCAS'85, ISCAS'89 and ITC benchmark circuits, it can infer that the toggle count in behavioral level circuit is better promising result than the gate level [17]. 31 135 36 3 1 70 56 -1 0 4 17 3 1 70 78 0 0 136 144 -1 0 70 78 -1 0 0 8 3 1 54 54 0 1 136 136 -1 0 79 87 2 0 46 44 -1 0 78 87 4 1 136 144 -1 0 54 54 3 1 88 96 -1 0 0 8 3 1 97 105 1 0 136 144 -1 0 88 96 -1 0 18 26 3 1 106 114 0 0 136 144 -1 0 36 44 -1 97 105 3 1 55 55 0 1 136 136 -1 0 106 114 2 0 45 53 -1 0 106 114 4 1 136 144 -1 0 55 55 3 1 88 96 -1 0 0 8 3 1 115 123 1 0 136 144 -1 0 88 96 -1 0 27 35 Toggling in the gate-level circuits are shown in the table I contain various benchmark circuits, with their corresponding number of gates in the circuit. Genetic algorithm is applied to all these circuits with a population size P. The number of generations (iterations) is represented as W. The number of gates in each circuit is also shown. The output produced by the genetic algorithm will be the test vector pairs that produces maximum toggling. The optimization of the test vectors are done by the GA. The generation of test vector pairs is shown in the figure below. The selection of the
Velammal College of Engineering and Technology, Madurai
optimal pair of vector pairs is based on the maximum toggle count. The first two test vectors are the fittest solution that produces the maximum toggling in a given circuit. Since the maximum toggling is found out from this experiments, the notion of maximum power dissipation can also be known because they are directly proportional. Thus the peak power can be estimated.
Fig. 4 Toggle vector generation Circuit
No. of gates
P
W
Toggle count
In lines
c17.v
6
10
100
5
5
c432.v
160
10
100
76
36
c7552.v
3513
10
100
1603
207
b01.v
70
10
100
37
4
b02.v
41
10
100
10
3
b03.v
312
10
100
96
6
b04.v
998
10
100
219
13
b05.v
1007
10
100
484
3
b06.v
73
10
100
37
4
b07.v
682
10
100
142
3
b08.v
261
10
100
109
11
b09.v
273
10
100
91
3
b10.v
288
10
100
103
13
b11.v
793
10
100
234
9
b11n.v
1145
10
100
284
9
b12.v
1084
10
100
537
7
b13.v
577
10
100
230
12
b14.v
6437
10
100
2398
34
b141.v
6934
10
100
2766
34
b15.v
12292
10
100
3649
38
b20.v
13980
10
100
5251
34
b211.v
14882
10
100
5646
34
b221.v
21671
10
100
7749
34
Page 105
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
boothG.v
1145
10
100
658
10
IdctG.v
7550
10
100
3086
41
TABLE I TOGGLE COUNT OF GATE-LEVEL CIRCUITS
The results of toggling in behavioral circuits is also represented as the number of bits toggled. The input vector pair that produced maximum toggling is also calculated. In the case of behavioral and gate level this input vector pair is in the form of long sequences of binary strings which will be equal to the number of input lines in gate level and it will be equal to the number of bits in the input lines of behavioral circuits. CIRCUIT
TOTAL BIT
P
W
TOGGLE
c432b.v
135
10
100
57
c74L85b.v
13
10
10
11
c74L81b.v
52
10
100
20
c6288.v
32
10
100
22
TABLE II TOGGLE COUNT OF BEHAVIOR LEVEL CIRCUITS
Tables I and II show the experimental results conducted on different ISCAS'85, ISCAS'89 and ITC benchmark circuits. The programming language used for the implementation of the proposed approach is C. It proved to be more flexible for accessing hardware resources such as memory, file operations etc. The intermediate file production is performed using scripting language Perl. It is more convenient to use this language in order to extract sufficient information from an input file. The context of using this language is that the input to the peak power estimation program [1] is a Verilog file. Most of the cases the data extracted from the input file is with the help of regular expressions. CONCLUSION AND FUTURE WORK
The proposed method of power estimation is a novel approach in the field of power estimation. GA based power estimation in digital circuits not only generates the required toggle vectors but also optimizes it. Experimental results using the ISCAS'85, ISCAS'89 and ITC benchmark circuits shows that the proposed method could produce an optimized results on behavioral circuits. The GA based power estimation can be extended for power estimation in complex processors, for analyzing their behavioral peculiarities. Future work shall involve modification of this method for behavioral level power estimation of many core architectures. VIII.
REFERENCES
[56] [1] E. M. Rudnick M. S. Hsiao and J. H. Patel, “Peak power estimation of [57] VLSI circuits: New peak power measures,” in IEEE Transactions on [58] VLSI Systems, 2000, p. 435439.
Velammal College of Engineering and Technology, Madurai
[59] [2] E. M. Rudnick M. S. Hsiao and J. H. Patel, “Peak power estimation of [60] VLSI circuits: New peak power measures,” in IEEE Transactions on [61] Very Large Scale Integration Systems., 2000, pp. 435–439. [62] [3] C.T. Hsieh C.S. Ding, Q. Wu and M. Pedram, “Statistical estimation of [63] the cumulative distribution function for power dissipation in VLSI [64] circuits,” in Proceedings of Design Automation Conference, 1997, pp. [65] 371–376. [66] [4] Liao and Lepak K. M, “Temperature and supply voltage aware [67] performance and power modeling at micro architecture level,” in IEEE [68] Trans on Computer-aided Design of Integrated Circuits and Systems, [69] 2005, pp. 1042–1053. [70] [5] M. Hunger and S. Hellebrand, “Verification and analysis of self[71] checking properties through ATPG,” in Proc. 14th IEEE Int. On-Line [72] Testing Symp (IOLTS08), 2008, pp. 25–30. [73] [6] S. Sheng A. P. Chandrakasan and R. W. Broderson, “Lowpower CMOS [74] digital design,” Journal of Solid-State Circuit, vol. 27, no. 4, pp. 473– [75] 483, April 1992. [76] [7] H. Takase K. Zhang, T. Shinogi and T. Hayashi, “A method for [77] evaluating upper bound of simultaneous switching gates using circuit [78] partition,” in Asia and South Pacific Design Automation Conf. [79] (ASPDAC), 1999, p. 291. [80] [8] M. Pedram C-Y. Tsui and A. M. Despain, “Efficient estimation of [81] dynamic power dissipation under a real delay model,” in IEEE [82] Transactions on Computer-Aided Design of Integrated Circuits and [83] Systems, 1993, pp. 224–228. [84] [9] Theodore W. Manikas and James T. Cain, “Genetic algorithms vs [85] simulated annealing: A comparison of approaches for solving the circuit [86] partitioning problem,” in IEEE Trans. on Computer-aided Design, 1996, [87] pp. 67–72. [88] [10] P. Schneider and U. Schlichtmann, “Decomposition of Boolean [89] functions for low power based on a new power estimation technique.,” [90] In Proceedings of 1994 International Workshop on Low Power Design, [91] 1994, pp. 123–128. [92] [11] S. Devdas F. Fallah and K. Keutzer, “Functional vector generation for [93] HDL models using linear programming and Boolean satisfiability,” in [94] IEEE Transactions on Computer-Aided Design of Integrated Circuits [95] and Systems, August 2001, pp. 994–1002. [96] [12] S. Ravi L. Lingappan and N. K. Jha, “Satisfiability-based test [97] generation for non separable RTL controller-datapath circuits,” in IEEE [98] Transactions on Computer-Aided Design of Integrated Circuits and [99] Systems., 2006, pp. 544–557.
Page 106
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[100] [13] Q. artas, Simulation Based Power Estimation, Altera Corporation, [101] 2004. [102] [14] A. K. Chandra and V. S. Iyengar, “Constraint solving for test case [103] generation: a technique for high-level design verification,” in In [104] Proceedings on International Conference on Computer Design: VLSI [105] in Computers and Processors., 1992, pp. 245–248. [106] [15] I. Ghosh and M. Fujiita, “Automatic test pattern generation for [107] functional register transfer level circuits using assignment decision [108] diagrams,” in IEEE Trans on Computer-aided Design of Integrated [109] Circuits and Systems, 2001, pp. 402–415. [110] [16] Noel Menezes Chandramouli Kashyap, Chirayu Amin and Eli Chiprou, [111] “A nonlinear cell macromodel for digital applications,” in in IEEE [112] Transactions on Computer-Aided Design of Integrated Circuits and [113] Systems, 2007, p. 678685. [114] [17] C. Y. Wang and K. Roy, “Cosmos: A continuous optimization [115] approach for maximum power estimation of CMOS circuits,” in [116] Proceedings of ACM/IEEE International Conference on Computer [117] Aided Design, 1997, pp. 45–50. [118] [18] Thomas Weise, Global Optimization Algorithms: Theory and [119] Application, www.it-weise.de/projects/book.pdf, Germany, 2006. [120] [19] Alex Fraser and Donald Burnell, Computer Models in Genetics, [121] McGraw-Hill, New York, 2002. [122] [20] R. Vemuri and R. Kalyanaraman, “Generation of design verification [123] tests from behavioral VHDL programs using path enumeration and [124] constraint programming,” in in IEEE Transaction on VLSI Systems, 1995, pp. 201–214.
Velammal College of Engineering and Technology, Madurai
Page 107
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Human Action Classification Using 3D Star Skeletonization and RVM Classifier Mrs. B. Yogameena#1, M. Archana*2, Dr. (Mrs) S. Raju Abhaikumar#3 #
Deptartment of Electronics and Communication, Affiliated to Anna University Thiagarjar College of Engineering, Madurai – 625015, Tamil Nadu, India. 1
[email protected] [email protected]
3
*
Deptartment of Electronics and Communication, Affiliated to Anna University 94/16, Irulandi st, Natraj nagar, Madurai – 625016, Tamil Nadu, India. 2
[email protected]
Abstract - This paper presents a real-time video surveillance system which classifies normal and abnormal action in crowds. Detection of individual’s abnormal action has become a critical problem in the events like terrorist attacks. The aim of this work is to provide a system which can aid in monitoring human actions, using 3D star skeleton technique, which is a suitable skeletonization for human posture representation and it reflects the 3D information of human posture. This 3D star skeleton technique can offer more accurate skeleton of posture than the existing star skeleton techniques as the 3D data of the object is concerned. Using these 3D features as input to the Relevance Vector Machine (RVM) classifier, the different human actions from a stationary camera is classified as normal or abnormal for a video surveillance application. Keywords – Video surveillance, Abnormal action, 3D star skeletonization, Relevance Vector Machine.
I. INTRODUCTION Security of citizens in public places such as Hotels, Markets, Airports and Train stations is increasingly becoming a crucial issue. The fundamental problem in visual surveillance system is detecting human presence, tracking human motion, analysing the activity and asses abnormal situations automatically. Based on this motivation, crowd modelling technology has been under development to analyse the video input which is constantly crowded with humans, as well as to ready to act against abnormal activities emerge. The aim of the paper is to classify human normal and abnormal actions in crowds using 3D star skeletonization for a surveillance system which is based on RVM. Human action and posture recognition is a significant part on human centred interface which is coming up to the resent issues now-a-days. In posture recognition application, the skeletal representation [1] captures the essential topology of the underlying object in a compact form which is easy to understand. There are three existing methods for skeletal construction such as distance transformation [2]-[5], voronoi diagram [6]-[8] and thinning [9]-[12]. These methods can generate an accurate skeleton but are computationally
Velammal College of Engineering and Technology, Madurai
expensive and cannot be used in real time. To complement these, a 2D star skeleton was proposed [13], [14], [15]. However it also has some limitations to describe the 3D information of the posture. Hence the 3D star skeletonization method is proposed for the feature extraction for the classifier. Then the normal and abnormal human action is classified in the real time surveillance applications to analyze the unusual activity of an individual from the normal ones. Many methods based on computer vision have been proposed in the literature to classify people’s action earlier. Ardizzonee et al. [16] has proposed a pose classification algorithm using support vector machine to classify different poses of the operator's arms as direction commands like turn-left, turnright, go-straight, and so on. In [17] the authors proposed a method that describes a system which delivers robust segmentation of the hands using a combination of colour and motion analysis, and the implementation of a multi-class classification of the hand gestures using a SVM ensemble. Cipolla and Philip et al. [18] estimate the pose of an articulated object from a single camera using relevance vector machine. In [19] Hui-Cheng Lian et al. presented a novel approach to multi-view gender classification considering both shape and texture information to represent facial image and the classification is performed using SVM. These papers dose not concentrate on full body pose. Hence Relevance Vector Machine is considered which utilizes only the relevance vectors for the training [15]. In this paper, an improved 3D star skeleton [20] is proposed which is based on the 3D information of the human posture whose purpose is to recognize human action as normal or abnormal. The 3D features are given as input to the Relevance Vector Machine for classification of human action. Experimental results demonstrate that the approach is robust in classifying human abnormal action. The remainder of this paper is organized as follows: Section 2 describes the methodology. Section 3 describes the 3D star skeletonization. Section 4 describes the RVM learning system for classification. Section 5 describes the experimental results. Finally, the conclusion is presented.
Page 108
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Video Sequence
Background Subtraction
3D Star Skeletonization
Classification Using Relevance Vector Machine
Classified Poses Fig 1: The block diagram of the Video Surveillance System I. METHODOLOGY The overview of the system is shown in fig.1. In the foreground detection stage, the blob detection system detects the foreground pixels using a statistical background model. Subsequently foreground pixels will be grouped into blobs. The 3D star skeleton features are obtained for the foreground image sequences. Using these features the Relevance Vector Machine classification scheme classifies the given image sequence into normal action or abnormal action sequence.
II.
BACKGROUND SUBTRACTION AND PROJECTION In this work, background subtraction is accomplished in real-time using the adaptive mixture of Gaussians method proposed by Atev et al. [21]. There are some of the practical issues concerning the use of the existing algorithm based on mixtures of Gaussians for background segmentation in outdoor scenes, including the choice of parameters [[22], [15]]. The proposed system analyses the choice of different parameter values and their performance impact is obtained to get robust background model. In addition, the motivation for dopting this method stems from its simplicity and efficiency in meeting with sudden global illumination changes based on the contrast changes over time. Subsequently, the individual is to be identified for further analysis leading to action classification. An extracted blob in a frame, representing an individual is subjected to action analysis described in subsequent sections. If there exists more than one blob, but with connectivity, there is likelihood to be considered as single entity. This results in the identification of a group as “individual.” This makes recognition of individual’s action in a crowd more difficult. Therefore, a geometric projection on the blob is proposed to separate an individual from the group for analysing his or her actions. The blob is projected to head and ground plane from the camera view point leading to intersected area in world coordinates. Such projection shown in Figure 2 eliminates the variation of area with the distance from the camera so that it identifies only humans [15]. The success of human identification lies on segmentation of individual human in a given frame as a single blob. However, there is a chance of multiple blobs representing an individual human due to over segmentation.
Velammal College of Engineering and Technology, Madurai
But, since the projection of a blob is accomplished from head plane to ground plane, any discontinuity in a blob representing an individual is achieved by linking discontinuous blobs covered by bounding rectangle. III. 3D STAR SKELETONIZATION Star skeleton, which is a “star” fashion, is a kind of representative features to describe a human action. 3D star skeleton technique extracts the 3D information of the image sequence. It is simple, real-time and robust technique. There are 4 steps involved id 3D star skeletonization as described below. A. Projection Map Generation The definition of a projection map is to project the 3D world onto the 2D image plane orthographically. A pixel value on the projection map generally contains 3D information of the target object, such as nearest boundary, furthest boundary and the thickness of the object.To generate the projection map, and initially the 8 projection plane is created. Then a virtual camera is set to the projection plane and perpendicular vectors are drawn along the viewing direction. These hit points along the viewing direction are obtained and is saved in a buffer. From this the maximum hit points are obtained. The depth information of the human posture is obtained from these projection maps. The 8 projection map is used as the input to construct the 3D star skeleton. Each of the projection map used has a different silhouette of human action according to where the projection map is generated and each pixel value means a distance from view point to a voxel of model surface, depth information. Through these properties star skeleton can diminish the loss of 3D human posture information. B. Candidate Extraction Candidate extraction means detection of gross extremities of posture in the projection map. The extremities of boundary in the projection map are called the “candidate”. Candidate is likely to be the final extremity of 3D star skeleton as a feature. As the existing star-skeleton is generated by the extremities which are from 2D boundary points, there are some limitations to describe a 3D human posture. To overcome these limitations, the 3D information losses of human posture is reduced by using the 3D boundary, in which depth, and x, y co-ordinates data of the posture in the projection map are included.Candidates are found from the distance from each boundary trace in clockwise or counter-clockwise order. The distance function has noise component of the boundary, which can be reduced by the noise reduction process executed using smoothing filters. Consequently, candidates are detected by finding zero crossings in the smoothed distance function.
Page 109
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1)
Candidate Extraction Process Boundary of posture in the projection map is given as input to the candidate extraction system. The centroid of the candidate of the posture boundary in a projection map (xc, yc, zc) are extracted using the following expressions.
xc =
1 N ∑ xi N j =1
1 Nb
(2)
Nb
∑z i =1
M rotation
( xi − xc ) + ( yi − yc ) + ( zi − zc ) 2
2
2
(4)
Then the distance function is declared as a one dimensional discrete function as d(i)=di. But this distance function is corrupted by noise of the boundary. This noise is reduced by a smoothing filter. The smoothed distance function is . C. dCoordinate Transformation (i ) In transformation phase, the candidates of each projection map are transformed into one particular coordinate. This is done because the coordinate systems of the extracted candidates are not identical. All candidates exist in their own projection map coordinate system. To construct 3D star skeleton, the candidates should be in the same coordinate. To transform all candidates into one particular coordinate, 3 phases are done: reflection, translation, rotation process. Reflection is to fit the two opposite projection map coordinates to one coordinate. The applied reflection matrix is,
M reflection
⎛ −1 ⎜ 0 =⎜ ⎜0 ⎜ ⎝0
0 0 w ⎞ ⎟ 1 0 0 ⎟ 0 −1 255 ⎟ ⎟ 0 0 1 ⎠
(5)
Where, w is the width of projection map and maximum values of z in the projection map is 255. To rotate the candidates to one particular coordinate, the axis of all candidates should be translated to one identical axis, which is
Velammal College of Engineering and Technology, Madurai
⎛ ⎜1 ⎜ ⎜0 =⎜ ⎜0 ⎜ ⎜0 ⎝
w⎞ 2⎟ ⎟ 1 0 0 ⎟ 256 ⎟ 0 −1 ⎟ 2 ⎟ 0 0 1 ⎟⎠ 0
0
−
(6)
The rotation of the 4 couples of projection map coordinates, which are fitted at reflection process, into one particular coordinate is performed by the process of rotation. The rotation matrix is given as,
(3)
i
Where, Nb is the number of boundary points. (xi, yi, zi) is a boundary point of the projection map. The distance di from each boundary point (xi, yi, zi) to the centroid (xc, yc, zc) is obtained using,
di =
M translation
(1)
1 N yc = ∑ yi N j =1
zc =
performed by the translation process. Translation matrix performs the axis shift operation which is given by,
⎛ cos θ ⎜ 0 =⎜ ⎜ − sin θ ⎜ ⎝ 0
0 sin θ 1 0 0 cos θ 0 0
0⎞ ⎟ 0⎟ 0⎟ ⎟ 1⎠
(7)
Where, θis the variable which is changed according to the projection map. D. Clustering Candidates and 3D Star Skeleton Construction To determine the extremities as features of the 3D star skeleton, the transformed candidates should be classified. After transformation process the transformed candidates are scattered in one particular coordinate indicating the some part of posture locates near others generated from the same part, not exist at one specific position together due to the thickness of human body. Hence, all the candidates are to be classified into several groups. The mean of each cluster becomes the extremities of star skeleton. ‘K’ mean clustering algorithm is used to classify the candidates. According to which the candidates are divided into k groups, defining K as 5. The centroid is the average of centroids on all the projection maps. Finally, the 3D star skeleton is constructed by connecting the extremities with the centroid. The features of 3D star skeleton are generated by calculating the distance from extremities to the centroid of 3D star skeleton. In some cases the candidates might not locate at the landmark points, such as feet or hands. For example, one knee point can screw up the foot canter if it is included in that cluster leading to error in clustering process. Hence K-Mean Clustering Algorithm is modified for fitting to the proposed method. To reduce the error candidates, noises are filtered from the cluster by using standard deviation. After clustering process, if some candidate is far from the mean of its cluster, it is removed and the mean of that cluster is recalculated. Thus the 3D star skeleton is constructed. IV. RELEVANCE VECTOR MACHINE Action classification is a key process for analyzing the human action. Computer vision techniques are helpful in automating this process, but cluttered environments and consequent
Page 110
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
occlusions often makes this task difficult [24]. There are numerous methods for incremental model based pose estimation where a model of an articulated structure (person) is specified [25]-[26]. Many types of neural networks are used for a binary classification problem like individual’s activity classification as normal or abnormal. By training the systems, the difference between normal and abnormal human actions, the computational action models built inside the trained machines can automatically identify whether the action is normal or abnormal. The action classification system proposed in this paper is trained for both normal and abnormal actions so that testing becomes a two class hypothesis problem. SVM is classical training algorithm because it has stronger theory-interpretation and better generalization than the other neural networks mentioned earlier. The decision function of the SVM classification system cannot be much sparser i.e., the number of support vectors can be much larger. This problem can be partially overcome by the state of the art model RVM. The proposed Relevance Vector Machine (RVM) classification technique has been applied in many different areas of pattern recognition, including functional neuro images analysis [27], facial expressions recognition [28] and pose estimation [29]. The RVM is a Bayesian regression framework, in which the weights of each input vector are governed by a set of hyper parameters. These hyper parameters describe the posterior distribution of the weights and are estimated iteratively during training. Most hyper parameters approach infinity, causing the posterior distributions of the corresponding weights to zero. The remaining vectors with non-zero weights are called relevance vectors. RVM does not need the tuning of a regularization parameter and also the inversion of a large matrix is not required during the training phase. This makes this methodology appropriate for large datasets. In this paper, Relevance Vector Machine technique is used for the classification of human action such as normal or abnormal. A. Classification of poses using multi-class Relevance Vector Machine The extracted 3D features are inputted to Relevance Vector machine for learning to classify poses. In our framework, we are focused our attention on the pose of the whole human body. This kind of gestures can be described very well through the analysis of the body contour. Moreover, to provide a detailed description of a shape, one has to take into account the whole body contour. The image features were shape-contexts descriptors of silhouette points and pose estimation was formulated as a one-to-one mapping from the feature space to pose space, as shown in Fig.4. The pose of an articulated object, a full human body, is represented by a parameter vector x given in eqn (8)
X = wk φ ( z ) + ξ K
k
w - is the weight of the basis function ф(z) - is the vector of the basis function k
ξ - is the Gaussian noise vector In order to learn about the multiple RVM technique an expectation maximization algorithm has been adopted. It is used to minimize the cost function of the multiple RVM regression shown in (9)
( )
Lk = ∑ C k( n ) Yk( n )
T
( )
S k Yk( n )
(9)
( )
Yk( n ) = x( n ) − wkφ Z ( n ) Here, Y
(n)
k
(10)
-is the output with n sample points belongs to the
mapping function k. k
W - is the weight of the basis function. (n)
Φ(z ) –is the design matrix of vector of the basis function. k
S – is the diagonal covariance matrix of the basis function. C
k
(n)
- is the probability that the sample point n belongs
to the mapping function k. V. RESULTS and DISCUSSION: The efficiency of the proposed algorithm has been evaluated by carrying out extensive work on the simulation of the algorithm using Matlab7.9. The proposed method process about 24 frames of any size each on PC with AVI file format. The Weizmann dataset for walking, running, bending and jumping are taken and CAVIAR dataset for a person lifting his hand is taken. In order to test for complexity, IBM dataset having multiple people walking and CMU database containing two persons fighting with each other in one sequence and a person pulling the hands of the other person in one sequence are taken. The video sequence is converted into frames and the foreground frames are obtained using GMM. As mentioned in Table.1. Dataset DS I, has a person bending down along with two other foreground blobs is detected. In DS II, there are multiple people with two groups, one group of people in the camera’s viewpoint and the other group of people away from camera’s viewpoint. Hence the people away from camera’s viewpoint are easily identified as individuals and the remaining as group. In DS III five foreground blobs are obtained including the running person’s blob. In all the three DS the car is considered to be a static background.
(8)
Where, x – is the input for the system
Velammal College of Engineering and Technology, Madurai
Page 111
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE 1: DATASETS FROM DIFFERENT OUTDOOR SEQUENCES Data set College Campus
Benchmark Dataset
IBM Dataset Weizmann Dataset
CMU Dataset CAVIAR
Indoor/outdoor
Sequence Length
Frame Size
Indoor/outdoor
1628
240X320
Indoor/outdoor
864
240X320
Running human in a group-DS-III
outdoor
1320
240X320
A person carrying a bar in a group DS- IV Two persons Walking
outdoor
920
240X320
Indoor
781
240X320
Eli-Walk
Outdoor
645
240X320
Eli-Run
Outdoor
712
240X320
Moshe-Bend
Outdoor
786
240X320
Eli-Jump
Outdoor
855
240X320
Indoor
1234
240X320
An individual bends down while most of walking-DS-I(DS represents Data Set) A person waving hand in a group-DS-II
Person A fights with person B Person A pulls the person B
Indoor
1065
240X320
An individual with his hand lifted up
Indoor
1187
240X320
Then the 3D star skeleton obtained for the different datasets using the projection map system. For the individual blob like walking, running and jumping the skeletal features are obtained even more clearly than the 2D star skeleton. For the bending action DS I, the skeleton features vary from the Weizmann dataset in 2D star skeleton. In the Weizmann dataset the skeleton points were depicted as a short human walking, were as it is a skeleton of a man “bending”. This problem is overcome in 3D star skeletonization. The bending person is identified correctly. In 2D skeleton of the CMU fighting datasets the person who holds the stool is skeletonised including the stool. But in 3D star skeletonization the stool is identified separately from the person holding it. Similarly for the CMU pulling dataset also the skeleton features are obtained clearly for the two individuals than the 2D star skeleton. Finally, the skeleton features are fed to the Relevance Vector Machine to classify the abnormal action from the normal action which is indicated in red color as shown in Figure (2.a.), Figure (2.b.) & Figure (2.c.) and also for benchmark datasets as shown in Figure (4).
a)One person bending down
Velammal College of Engineering and Technology, Madurai
b) One person waving hand
c) One person running in the crowd and another person half bending Fig.2. Results of classified abnormal actions for College Campus dataset
Relevance Vector Machine uses suitable kernel for the task of classification. Gaussian kernels are used in this proposed method. A ten-fold cross validation (CV) has been applied in the training dataset, to determine the fine-tuning parameters of the RVM classifier model for optimal performance. For each dataset, 80% of the sequence has been used for training and the remaining 20% for testing. A best error level of 6.89% is obtained by using the Gaussian Kernel. Gaussian Kernel is the best with around 100% for training, 94% for CV and 96% for testing the given feature vectors. The error rate of the Gaussian Kernel is lower than that of other kernels in terms of classification rate. The results of RVM classification is shown in table.2.
Page 112
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE 2: THE RESULTS OF RVM CLASSIFICATION Actions
Normal Abnormal
Benchmark Datasets
Running Carrying Bar Bending Waving hand IBM Eli-Walk Eli-Run MosheBend Eli-Jump CMU1 CMU2 CAVIAR
Vectors RVM Multi class RVM with skeleton Points 11 18 23 12 16 12 15 18 20 16 19 14 12
and false negative. Higher error level generally leads to poor classification of actions. From Figure (3) it is inferred that, as the number of iterations increases the better convergence of relevance vector is achieved for the experimental database. When the 2D star skeleton features were inputted to the RVM classifier the abnormal actions except the Weizmann bending and CMU pulling were classified. Weizmann bending was classified as normal action and in CMU pulling dataset both individuals were considered as walking. But this is overcome by using 3D star skeleton features as input to the RVM classifier. The Weizmann bending and CMU pulling datasets are classified as abnormal actions as shown in Figure (4).
VI. CONCLUSION In this paper, a novel and real-time video surveillance system for classifying human normal and abnormal action is described. Initially, the foreground blobs are detected using adaptive mixtures of Gaussians which is robust to illumination changes and shadows. Then the projection map system is generated to get the 3D information of the object in the foreground image. Then the 3D star skeleton features are extracted. These features reduce the training time and also improve the classification accuracy. The features are then learnt though a relevance vector machine to classify the individual’s actions into normal and abnormal behaviour. The number of relevance vectors obtained is less and it does not require the tuning of a regularization parameter during the training phase. The error rate is also reduced by selecting the appropriate Gaussian kernel which also reduces the computational complexity. The distict contribution of this proposed work is to classify the actions of individuals. The proposed system is able to detect abnormal actions of individuals such as running, bending down, waving hand while others walk, and people fighting with each other with high accuracy.
Fig 3: Relevance vector obtained after training input feature vectors
The percentage of error level considered for both training and testing sequences of the dataset are in terms of true positive
Velammal College of Engineering and Technology, Madurai
Page 113
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Dataset
Original Frame
Background Subtracted Image
Skeleton Image
Classified Image
IBM dataset
Weizmann dataset Eli-Walking
Eli-Running
MosheBending
EliJumping CMU Fighting
CMU Pulling
CAVIAR
Fig 4: Results for benchmark datasets
Velammal College of Engineering and Technology, Madurai
Page 114
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
ACKNOWLEDGMENT We sincerely thank our management and our respectful principal for providing us with all the facilities we required for carrying on our research. We also thank them for their support and encouragement in all aspects for our research. REFERENCES [1] Nicu D. Cornea, Deborah Silver, Patrick Min, “ Curve-Skeleton Properties, Application, and Algorithms,” IEEE Trans. Visualization and Computer Graphics, vol.13, 2007, pp. 530-548. [2] Gunilla Borgefors, “Distance transformations in digital images,” Computer Vision, Graphics, and Image Processing, vol. 34, 1986. pp. 344-371. [3] Gunilla Borgefors, “ Distance transformations in arbitrary Dimensions,” Computer Vision, Graphics, and Image Processing, vol. 27, 1984. pp. 321-345. [4] Gunilla Borgefors, “ On digital distance transforms in three dimensions,” Computer Vision and Image Understanding, vol. 64, 1996, pp. 368-376. [5] Frank Y.Shih and Christopher C.Pu, “A skeletonization algorithm by maxima tracking on Euclidean distance transform”, J.Pattern Recognition, vol.28, 1995, pp. 331-341. [6] Franz Aurenhammer, “Voronoi diagrams – A Survey of a fundamental geometric data structure,” ACM Commputing Survey, vol. 23,1991, pp. 345-405. [7] Jonathan W. Brandt and V. Ralph Algazi, “ Continuous skeleton computation by Voronoi diagram,” CVGIP: Image understanding, vol. 55, 1991, pp. 329-338. [8] Kenneth E. Hoff III, Tim Culver, John Keyser, Ming Lin and Dinesh manocha, “Fast computation of generalied Voronoi diagrams using graphic hardware,” in Proc. 26th annual Conf. Computer graphics and interactive tech-nique, 1999, pp. 277-286. [9] Kalman Palagyi, Erich Sorantin, Emese Balogh, Attila Kuba, Csongor Halmail, Balazs Erdohelyi, and Klaus Hausegger, “A Sequential 3D Thinning Algorithm and Its Medical Applications,” in Proc. 17th international Conf.IPMI, vol. 2082, 2001, pp. 409-415. [10] Kalman Palagyi and Attila Kuba, “ A 3D 6-subiteration thinning algorithm for extracting medial lines,” Pattern Recognition Letters, vol. 19, 1998, pp. 613-627. [11] Kalman Palagyi and Attila Kuba, “ Directional 3D thinning using 8 subiterations,” in proc. 8th international Conf. DGCI, vol. 1568, 1999, pp. 325-336. [12] Ta-Chin Lee, Rangasami L. Kashyap and Chong-Nam Chu, “ Building skeleton models via 3-D medical Surface/axis thinning algorithms,” CVGIP : Graphical Models and Image Processing, vol. 56, 1994, pp. 462-478. [13] H. Fujiyoshi and A.J Lipton, “Real-time human motion analysis by image skeletonization,” 4th IEEE Workshop on Application of Computer Vision, 1998, pp. 15-21. [14] Hsuan-Sheng Chen, Hua-Tsung Chen, Yi-Wen Chen and Suh-Yin Lee, “Human Action Recognition Using Star Skeleton,” in Proc. 4th ACM international workshop on video surveillance and sensor networks, 2006, pp.171-178. [15] B. Yogameena, S. Veeralakshmi,E. Komagal, S. Raju, and V. Abhaikumar,“ RVM-Based Human Action Classification in Crowd through Projection and Star Skeletonization,” in EURASIP Journal on Image and Video Processing, vol. 2009, 2009. [16] E. Ardizzone, A. Chella, R. Phone, “ Pose Classification Using Support Vector Machines,” International Joint Conference on Neural Networks, vol. 6, 2000. [17] D.F. Llorca, F. Vilarino, Z. Zhouand G. Lacey, “A multi-class SVM classifier ensemble for automatic hand Washing quality assessment,” Computer science Dep, Trinity College Dublin, Rep, of Ireland. [18] Cipolla and Philip et al,“ Pose estimation and tracking using Multivariate regression,” pattern recognition,2008.
Velammal College of Engineering and Technology, Madurai
[19] Hui – Cheng Lian and Bao – Liang Lu, “ Multi - View Gender Classification using Local Binary Patterns and Support Vector machines,” verlag, 2006, pp. 202-209. [20] Sungkuk Chun, Kwangjin Hong, and Keechul Jung, “ 3D Star Skeleton for Fast Human Posture Representation,” Proceedings of World Academy of Science, Engineering and Technology, vol. 34, 2008. [21] S. Atev, O. Masoud, N.P. Papanikolopoulos, “ Practical mixtures of Gaussians with brightness monitoring,” Proc. IEEE Int. Conf. Intel. Transport. Syst. pp.423–428, October, 2004. [22] C. Stauffer and W.E.L Grimson., “Adaptive background mixture models for real time tracking,” In proceedings of the IEEE Int’l conf.Computer Vision and Pattern recognition pp. 246 -252 1999. [23] G. Tzikas, A. Likas, N. P. Galatsanos, A. S. Lukic and M. N. Wernick, “Relevance vector machine analysis of functional neuroimages,” IEEE intel. Symposium on Biomedical Imaging, vol.1, pp. 1004-1007, 2004. [24] H.C.Lian, B.L. Lu, “Multi-view Gender Classification Using Local Binary Patterns and Support Vector Machines,”-verlag -pp. 202–209, 2006. [25] R. Chellappa, A. K. Roy-Chowdhury and S. K. Zhou, “Recognition of Humans and Their Activities Using Video,” First edition, Morgan and Claypool publishers, 2005. [26] H. Ren and G. Xu, “Human action recognition in smart classroom,” IEEE Proc. Int. Conf. on Automatic Face and Gesture Recognition, pp. 399404, 2002. [27] A. Mittal, L. Zhao, L.S. Davis, “ Human Body Pose Estimation Using Silhouette Shape Analysis,” – proceedings of the IEEE International Conference on Advanced Video and Signal Based Surveillance (AVSS'03), 2003 [28] D Rita Cucchiara , A. Prati, R. Vezzani, University of Modena and Reggio Emilia, Italy, “Posture Classification in a Multi-camera Indoor Environment,” – International conference on image processing..vol 1, 2005. [29] G. Tzikas, A. Likas, N. P. Galatsanos, A. S. Lukic and M. N. Wernick, “Relevance vector machine analysis of functional neuroimages,” IEEE intel. Symposium on Biomedical Imaging, vol.1, pp. 1004-1007, 2004.
Page 115
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Relevance Vector Machine Based Gender Classification using Gait Appearance Features Mrs. B. Yogameena#1, M. Archana*2, Dr. (Mrs) S. Raju Abhaikumar#3 #
Deptartment of Electronics and Communication, Affiliated to Anna University Thiagarjar College of Engineering, Madurai – 625015, Tamil Nadu, India. 1
[email protected] [email protected]
3
*
Deptartment of Electronics and Communication, Affiliated to Anna University 94/16, Irulandi st, Natraj nagar, Madurai – 625016, Tamil Nadu, India. 2
[email protected]
Abstract - Automatic gait recognition is important task in security surveillance, biomechanics, purloining behaviors and biometrics. The background subtraction for detecting foreground information, the reliable extraction of characteristic gait features from image sequences and their classification are three important issues in gait recognition. The person has to be found in the image plane first, using robust background subtraction by modifying the parameters in Gaussian Mixture Model (GMM), assuming a static camera.This paper describes the representation of consistent gait appearance features and Relevance Vector Machine (RVM) based gender classification. Experimental results on benchmark datasets demonstrate the proposed gender classification method is robust and efficient. A comparative study between RVM based classification and Support Vector Machine (SVM) based classification is also presented. Keywords - Gait recognition, Relevance Vector Machine, appearance features, background subtraction
I. INTRODUCTION Automatic visual surveillance systems could play an important role in supporting and eventually replacing human observers. To become practical, this system needs to distinguish people from other objects and to recognize individual persons with a sufficient degree of reliability, depending on the specific application and security level. These applications required the task of estimating the gait and automatically recognising the gender for high level analysis. The first step in the automatic gender classification system is background subtraction. Piccardi et al. have reviewed a number of background subtraction approaches [1]. Wren et al. [2] have proposed a statistical method, in which a single Gaussian function was used to model the distribution of background. Later Mittal et al. have proposed a novel kernel based multivariate density estimation technique that adapts the bandwidth according to the uncertainties [3].Yet there are issues like the robustness to illumination changes, the effectiveness in suppressing shadows and the smoothness of foreground’s boundary which need to be addressed in indoor and outdoor environments [4].
Velammal College of Engineering and Technology, Madurai
There have been a number of appearance-based algorithms for gait and activity recognition. Cutler and Davis [5] used self-correlation of moving foreground objects to distinguish walking humans from other moving objects such as cars. D.F. Llorca et al., [6] proposed a method that describes a system which delivers robust segmentation of the hands using a combination of color and motion analysis, and the implementation of a multi-class classification of the hand gestures using a SVM ensemble. Cipolla and Philip have been estimated the walking nature of an individual of an articulated object from a single camera using relevance vector machine [7]. In [8] Hui-Cheng Lian et al., presented a novel approach to multi-view gender classification considering both shape and texture information to represent facial image and the classification is performed using SVM. Nixon, et al., [10] used principal component analysis of images of a walking person to identify the walker by gait. Shutler, et al., [11] used higher-order moments summed over successive images of a walking sequence as features in the task of identifying persons by their gait. Johnson and Bobick [12] utilized static parameters of the walking figure, such as height and stride length, to identify individuals. However it has been reported that the feature extraction is an important issue for classification algorithm [9]. The consistency of the number of features for a machine learning classification algorithm also plays a vital role in gender classification. There are many types of neural networks that can be used for a binary classification problem, such as Support Vector Machines, Radial Basis Function Networks, Nearest Neighbor Algorithm, and Fisher Linear Discriminant and so on. Machine learning techniques can be considered as linear methods in a high dimensional feature space nonlinearly related to the input space. Using appropriate kernel functions, it is possible to compute the separating hyper plane which classifies the two classes. An automated system based on SVM classifier that classifies gender by utilizing a set of human gait data has been described by Yoo et al. [9]. The decision function of the SVM classification system can not be much sparser i.e., the number of support vectors can be much larger. SVM need the tuning of a regularization parameter
Page 116
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
during the training phase. Consequently a machine learning technique which minimizes the number of active kernel functions to reduce computation time is required ([18]-[20]). The objective of this paper is to classify gender using appropriate features and classification technique to improve the classification accuracy. Hence a novel system is proposed to classify the gender. First the foreground blobs are detected using a background subtraction technique which is to be robust to illumination changes and shadows. Then appropriate features and RVM are used to classify the gender and thereby reducing the computational complexity by selecting an appropriate kernel also by improving the classification accuracy. Experimental results demonstrated that the proposed approach is robust in classifying gender. The remainder of this paper is organized as follows: Section 2 describes the methodology. Section 3 describes background subtraction, section 4 presents the formulation of centroid, section 5 represents the elliptical view of body posture, section 6 presents the extraction of feature and section 7 depicts the experimental results and the classification of gender using the Relevance Vector Machine extraction method. Finally, the conclusion is presented. II. METHODOLOGY Figure 1 shows an outline of the proposed system. Each image from the camera is forwarded to the pre-processing module where the background subtracted image sequences are obtained. Then consistent features have been selected using silhouette segmentation and ellipse fitting. Finally, the Relevance Vector Machine classification scheme classifies the gender as male or female for the given image sequence.
III.BACKGROUND SUBTRACTION The first stage of video surveillance systems seeks background to automatically identify people, objects, or events of interest in various changing environments. The difficult part of background subtraction is not the differencing itself, but maintenance of a background model and its associated statistics. In this work, background substraction is accomplished in real-time using the adaptive mixture of Gaussians method proposed by Atev et al [22]. There are some of the practical issues concerning the use of the existing algorithm based on mixtures of Gaussians for background segmentation in outdoor scenes, including the choice of parameters [23]. The proposed system analyses the choice of different parameter values and their performance impact are obtained to get robust background model. In addition to that, the notion for adapting this method is because of its simplicity and also an efficient method for coping with sudden global illumination changes based on the contrast changes over time. It describes K Gaussian distributions to model the surface reflectance value and is represented by eqn (1). k
P ( X t ) = ∑ ω i ,t *η ( X t , μ i ,t , ∑ i, t ) Where K is the number of distributions,
ωi,t is an estimate of the weight of the ith Gaussian, μ is the mean,
∑ i, t
is the covariance matrix of the ith Gaussian, and
η is a Gaussian probability density function. where, η(X , μ , ∑i,t) = t i,t
Background subtraction
Silhouette segmentation using centroid
Ellipse Fitting
……. (1)
i =1
1 n
1 (2π) 2 |∑| 2
− 1 (Xt −μt )T ∑−1(Xt −μt ) l 2
..... (2)
Assume the covariance matrix is of the form
∑
k
= σ k2 I
…… (3)
This means that the red, green and blue reflectance components of the surface are independent and have the same variances, which can reduce costly matrix computation. The weight is adjusted as shown in eqn (4)
ω ω t =(1 − α )ω t −1 +α ( M t ) i, t
Feature Extraction .
Classification using RVM III.
…… (4)
Where α is the learning rate and
M t is 1 for the model which is matched and 0 for others. After ordering the Gaussians, the first B distributions are chosen as the background model as shown in eqn (5)
Classified Gender
Velammal College of Engineering and Technology, Madurai
Page 117
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
b
B = arg min b (∑ ω k ⟩T )
.…. (5)
k =1
Where, T is a measure of minimum models for background. The background subtracted image obtained using adaptive GMM is shown in Fig. 2.
Fig.2.An example of the background subtracted image of weizmann dataset
IV. CENTROID For each silhouette of a gait, the centroid (xc, yc) is determined by using the following eqns (6 ),(7) and is shown in Fig 3.
xc =
1 N
N
∑x j =1
1 yc = N
….. (6)
i N
∑y j =1
i
….. (7)
Where (xc, yc) represent the average contour pixel position, (xi, yi) represent the points on the human blob contour and there are a total of N number of points on the contour [16].
The frontal-parallel view of the silhouette is divided into the front and back sections (except for the head region) by a vertical line at the silhouette centroid. The parts above and below the centroid are each equally divided in the horizontal direction, resulting in 7 regions that roughly correspond to: head/shoulder region, front of torso, back of torso, front thigh; back thigh, front calf/foot, and back calf/foot [12]. These seven regions are chosen to be very easy to compute in contrast to methods that locate the joints and segment the body at those joints. These regions are by no means meant to segment the body parts precisely. V. ELLIPTICAL VIEW After getting the centroid of the full body region and the seven segments of a silhouette, for each of the 7 regions, an ellipse is fitted to the portion of foreground object visible in that region as shown in Figure 5. The fitting of an ellipse to an image region involves computing the mean and the covariance matrix for the foreground pixels in the region. Let I(x, y) be the binary foreground image of a region to which the ellipse is to be fitted [17]. Assume that the foreground pixels are 1 and the background pixels are 0, then the centroid of the region, is calculated using the following relations,
1 ∑ I ( x, y ) x N x, y 1 y = ∑ I ( x, y ) y N x, y x=
…….. (8)
……. (9)
where N is the total number of foreground pixels:
N = ∑ I ( x, y )
......... (10)
x, y
The covariance matrix of the foreground region is then obtained by, Fig.3.An example figure for pointing the centroid of an weizmann dataset
Then the silhouette is proportionally divided into 7 parts as shown in Fig 4.
⎡( x − x ) 2 ( x − x)( y − y )⎤ ⎡a c ⎤ 1 ⎥ ⎢b b ⎥ = N .∑ I ( x, y ).⎢ 2 x, y ⎣ ⎦ ⎣⎢( x − x)( y − y ) ( y − y ) ⎦⎥ ............(11) The covariance matrix can be decomposed into eigenvalues λ1, λ2 and eigenvectors v1, v2 which indicate the length and orientation of the major and minor axes of the ellipse which is obtained by eqn (12).
⎡λ1 0 ⎤ ⎡a c ⎤ ⎢c b ⎥ [v1 v 2 ]= [v1 v 2 ] ⎢0 λ ⎥ ⎣ ⎦ 2⎦ ⎣
……. (12)
Fig 4: An example figure for the segmented silhouette image of Weizmann dataset
Velammal College of Engineering and Technology, Madurai
Page 118
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The elongation of the ellipse, l, is given by
l=
λ1 λ2
.
…. (13)
and the orientation α of the major axis is given by the following relation
⎛ v .x ⎞ α = angle (v1 ) = arccos⎜⎜ 1 ⎟⎟ ⎝ v1 ⎠
Frame 113
Frame 125
Frame 145
….. (14)
where x is the unit vector [1, 0].
Frame 124
Frame 132
Frame 129
Fig.6: Human gait image from Weizmann dataset
Fig 5: An elliptical view for the segmented silhouette image of Weizmann dataset
I. FEATURE EXTRACTION After calculating the orientation and elongation of the human walking figure, an ellipse is fitted according to the features extracted from each frame of a walking sequence consists of features from each of the 7 regions. They are the relative height of the centroid of the whole body [12], for his or her body length which is taken as a single feature and then the four ellipse parameter consist of the x, y coordinates of the centroid, the orientation, and the elongation, which correspond to head region, chest, back, centroid of the head, orientation of the head, and finally, the mean and variance of the back calf. These six features constitute about the appearance of the gait and are shown in Table I.
In this proposed method, there are about six different image sequences of walking of both genders in an outdoor environment are taken from Weizmann dataset [16] for the classification of the gender using machine learning technique by analyzing their gait as shown in Figure 6.For this experimentation, there are about five subjects are walking in an outdoor environment consisting of three men and two women taken from weizmann dataset [15]. Totally six datasets have been taken for the classification of the gender by analyzing their gait as shown in fig (6). The detection of the moving objects from the foreground region, the background subtracted images are shown in fig (7). As shown in fig (3), the centroid has been taken for all the frames and the ellipse is fitted for each frame as shown in fig (4). Using the above information, the features which are used to classify the subject as a men or women are given as follows: front calf, back, head (centroid and orientation), mean and standard deviation of the back calf. Then these feature vectors are given for training and testing of the relevance vector machine.
Instead of using the twenty nine features, these primary six features are used for the gender classification.
II. CLASSIFICATION OF GENDER USING RELEVANCE VECTOR MACHINE In recent years, machine learning methods have become prevalent to examine the pattern of structural data. Kernel methods such as Support Vector Machines (SVM) are widely used to classify the poses of faces, hands, different human body parts and Robots. Relevance Vector Machine Yields a formulation similar to that of a Support Vector Machine and it also uses hyper parameters instead of Margin/Costs. The RVM is a Bayesian regression framework, in which the weights of each input example are governed by a set of hyper parameters [18]. These hyper parameters describe the posterior distribution of the weights and are estimated iteratively during training. Most hyper parameters approach infinity, causing the posterior distributions of the effectively setting the corresponding weights to zero. The remaining vectors with non-zero weights are called relevance vectors. It reduces the “inappropriate” parameters for the classification than SVM [[6], [17]]. In this framework, the attention is focused to the nature of the gait of a human to classify the subject as a man or woman.
Velammal College of Engineering and Technology, Madurai
Page 119
TABLE. I .TYPES OF FEATURES S.NO
REGION
1
front calf
FEATURES mean of orientation
2 3
Back Head
4 5
Head back calf
mean of orientation mean of x coordinate of centroid mean of orientation std of x of centroid
6
back calf
mean of x of centroid
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
x = Wk φ (z) +
ξk
……. (15)
Where, x – is the input for the system Wk - is the weight of the basis function ф(z) - is the vector of the basis function ξk - is the Gaussian noise vector It is used to minimize the cost function of the RVM regression using eqn (16)
( )
( )
LK = ∑ C K( n ) y k(n ) S k y k(n ) T
where, (n )
……. (16)
( )
y k = x (n ) − W K φ Z (n )
……… (17) Where, yk(n)-is the output with n sample points belongs to the mapping function k. Φ(z(n)) –is the design matrix of vector of the basis function. Sk – is the diagonal covariance matrix of the basis function. Ck(n) - is the probability that the sample point n belongs to the mapping function k.
III. RESULTS AND DISCUSSION OF RVM CLASSIFICATION The efficiency of the algorithm proposed has been evaluated by carrying out extensive works on the simulation of the algorithm .In this paper the video file used is Weizmann dataset [21]. The proposed method processes about 24 frames per second and the total number of frames is up to 460 at size of 240x320 each.
Frame 113
Frame125
Frame 145
features and are given as input for the RVM for the gender classification. The classification accuracy is measured with two types of kernels in this proposed method. The two types of feature vectors of each subject are used for testing, training and cross validation (CV) for both SVM and RVM. The experimental results are summarised in Table II. The accuracy is the average by the number of SVs, RVs, and classification rate. The test result of 6 selected features is a little higher than that of 29 original features. Also the numbers of relevant vectors are smaller when compared to the support vectors. The average accuracy of linear kernel was the best with around 100.0% for training, 94.0% for CV and 96.0% for testing in the 6 features. The error rate of the linear kernel was lower than that of other kernels in terms of classification rate and computational cost. The Fig.8 shows the initial stage of the training. The Fig. 9 show the result after the fifth iteration. It is to be noted that the 3 colors (Indigo, red and yellow) represents the gender as a male and the remaining 3 colors (Blue, pink and green) indicates the gender as a female by setting the threshold value. The color dots in the figure denote the relevance vectors of the corresponding dataset which is obtained after the fifth iteration. The maximum number of iterations used here is ten. Figure.10 shows the comparative representation of SVM and RVM vectors, for the gender classification. The vectors used for gender classification is reduced for RVM than in SVM. From the fig 10, the results for the benchmark datasets have been shown with the corresponding step by step procedure as explained by the proposed method. Finally the classified image is obtained as, the male sequences are identified through the green color rectangular bounding box and the female sequences a recognized through the red color bounding box. Output (Z+ weight)
These kinds of gestures can be described very well through the analysis of the gait. Moreover, to provide a detailed description of a shape, one has to take into account the gait. The walking pose of an articulated object, gait is represented by a parameter vector x and is given in eqn (15)
Input feature vector (Z) Fig.8.Initial Stage
Frame 124
Frame 132
Frame 129
Fig.7. background subtracted image of the weizmann dataset
The background subtracted images using GMM are shown in Fig 7. After background subtraction, the centroid is obtained and the silhouette is segmented into seven regions. Then ellipse is fitted for each segment and is shown in Figures (35). Then six primary features are extracted from the total 29
Velammal College of Engineering and Technology, Madurai
Page 120
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Output (Z+ weight)
Input feature vector (Z) Fig.9.Result for Classification
XI. CONCLUSION This paper has introduced the framework of relevance vector machine to classify the gender for the application of video
ACKNOWLEDGMENT We sincerely thank our management and our respectful principal for providing us with all the facilities we required for carrying on our research. We also thank them for their support and encouragement in all aspects for our research.
REFERENCES [1] M. Piccardi, “Background subtraction techniques: A review,” IEEE International Conference on Systems,Man and Cybernetics, vol.4, pp. 3099-3104, Oct. 2004. [2] C. Wren, A. Azarbayejani, T. Darrell, and A. Pentland, “Pfinder: Real-Time Tracking of the Human Body,” IEEE Trans.On Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp.780-785, July 1997. [3] A. Mittal, N. Paragios, “Motion-Based Background Subtraction using Adaptive Kernel Density Estimation,” In Proceedings of the Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2004), pp. 302-309, 2004 [4] J. Hu and T.Su “Robust background subtraction with shadow and Highlight removal for indoor surveillance,” In proceedings of the EURASIP Journal on Advances in signal processing, 14 pages,2007. [5] R. Cutler and L. Davis., “Robust real-time periodic motion detection, analysis, and applications”. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2000, 22(8):781 –796. [6] D.F. Llorca, F. Vilarino, Z. Zhou and G. Lacey, “A multi-class SVM classifier ensemble for automatic hand washing quality assessment”, British Machine Vision Conference proceedings 2007. [7] Cipolla and Philip et al “pose estimation and tracking using multivariate regression” pattern recognition Letters 2008. [8] Hui-Cheng Lian and Bao-Liang Lu “Multi-view Gender Classification Using Local Binary Patterns and Support Vector Machines”Verlag, pp. 202–209, 2006. [9] Yoo, D. Hwang and M. S. Nixon ,, “Gender Classification in Human Gait Using Support Vector Machine” ACIVS 2005, LNCS 3708, Springer-Verlag Berlin Heidelberg, pp. 138 – 145, 2005 [10] M. Nixon, J. Carter, J. Nash, P. Huang, D. Cunado, and S. Stevenage, “Automatic gait recognition”, IEE Colloquium on Motion Analysis and Tracking, pp. 3/1–3/6, 1999.
Velammal College of Engineering and Technology, Madurai
surveillance. The appearance features which are used to recognize the individual as a gender are extracted using elliptical view of seven segmented regions of a gait and fed into the RVM learning system. It reduces the number of vectors and improved the classification accuracy by choosing the appropriate kernel. To this end the gait sequence is classified as male or female. The method has been used as a component of a system for estimating the counts of gender for video surveillance. It is been noted that utilizing the six features instead of utilizing the twenty nine feature shows the good classification accuracy and these results were also compared with the SVM. Apart from this, the gender classification task can clearly handle a large number of subjects successfully. By this, these results show that people can be identified according to gender by their walking pattern. This accord with earlier psychological suggestions and buttressing other similar results.It is evident that RVM has good accuracy.
[11] J. Shutler, M. Nixon, and C. Harris. “Statistical gait recognition via velocity moments”, Visual Biometrics (Ref.No. 2000/018), IEE Colloquium, pp. 11/1–11/5, 2000. [12] A. Johnson and A. Bobick. “Gait recognition using static, activity specific parameters”, CVPR, 2001. [13] C. BenAbdelkader, R. Cutler, and L. Davis. “Motion-based recognition of people in eigengait space”, Automatic Face and Gesture Recognition Proceedings, Fifth IEEE International Conference, pp.267 – 272, May 2002. [14] Rong Zhang, Christian Vogler, Dimitris Metaxas , “Human gait recognition at sagittal plane” Image and Vision Computing, Vol 25, Issue 3, pp 321-330, March 2007. [15] Carl Edward Rasmussen, “The Infinite Gaussian Mixture Model” Advances in Neural Information Processing Systems, pp. 554–560, 2004. [16] Prahlad Kilambi, Evan Ribnick, Ajay J. Joshi, Osama Masoud, Nikolaos Papanikolopoulos, “Estimating pedestrian counts in groups”, Computer Vision and Image Understanding, pp. 43–59,2008. [17] Lily Lee, “Gait analysis for classification” – AI Technical Report Massachusetts Institute of Technology, Cambridge, USA. June 2003. [18] Tipping, M.E. “Sparse Bayesian learning and the relevance vector machine”, J. Mach. Learn. Res, pp. 211–244, 2001. [19] Tipping, M.E., Faul. A., “Fast marginal likelihood maximization for sparse bayesian models”, In: Proc. 9th Internat. Workshop on Artificial Intelligence and Statistics, 2003. [20] Agarwal, A., Triggs, B., “3D human pose from silhouettes by relevance vector regression”, In: Proc. Conf. on Computer Vision and Pattern Recognition, vol. II. Washington, DC, pp. 882–888, July 2004. [21] Weizmann dataset Downloaded from http://www.wisdom.weizmann.ac.il/~vision/SpaceTime Actions.html [22] S. Atev, O. Masoud, N.P. Papanikolopoulos, “Practical mixtures of Gaussians with brightness monitoring,”Proc. IEEE Int. Conf. Intel.Transport. Syst. pp.423–428, October, 2004.
C. Stauffer and W.E.L Grimson., “Adaptive background mixture models for real time tracking,” In proceedings of the IEEE Int’l conf.Computer Vision and Pattern recognition pp. 246 -252 1999
Page 121
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE II EXPERIMENTAL RESULTS FOR CLASSIFICATION Kernel
No. Of
SV
V
Classification Rate(%)
features
Training SVM
Guassian
29
78 54
8 19
29
64
6
57
6
94.4 93.4
Cross Validation RVM
SVM
Testing
RVM
SVM
RVM
94.6 96
92 91
92 92
93 94
94 96
34 100.0
100.0
95
96
94
95
28 100.0
100.0
95
96
95
97
Linear
Original image
Background subtracted image
Segmented image
Elliptical view of an image
Classified image
Denis
Eli walking
Ido
Daria
Lena_ walk1
Lena_ walk2
Velammal College of Engineering and Technology, Madurai
Page 122
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Survey on Gait Recognition Using HMM Model M.Siva Sangari M.Yuvaraju Lecturer / Dept of MCA Lecturer / Dept of CSE Sri Ramakrishna Engineering College Anna University Coimbatore Coimbatore [email protected] [email protected] serve as a useful filtering tool that allows us to narrow the Abstract Vision-based human identification at a distance in search down to a considerably smaller set of potential surveillance systems has attracted more attention recently, and its candidates. current focus is on face, gait or activity-specific recognition. The Gait of each person is identified to be identical which can be used for identification of a person. Potential sources for gait biometrics are two important aspects namely gait dynamics and gait shape. So set of individuals is being identified and their gaits are being stored in the form of databases. There are different types of databases available namely CMU Mobo Database set, UMD database set, SOTON Large data set, Human ID gait challenge dataset etc. This paper has concentrated on the aspects of various persons experimental results on these databases based on Hidden Markov Model.
Keywords: Gait Recognition, Hidden Markov Model (HMM), Human ID I
INTRODUCTION
Vision-based human identification at a distance in surveillance applications has recently gained more interests, e.g., the Human ID Program being supported by DARPA. GAIT refers to the walking style of an individual. Studies in psychophysics indicate that human have the capability of recognizing people from even impoverished displays of gait, indicating the presence of identity information of gait. A gait cycle corresponds to one complete cycle from rest position toright-foot-forward-to-rest-to-left-foot-forward-to-rest position. In comparison with other first-generation biometrics such as fingerprints and iris, gait has the advantage of being noninvasive and non-contact, and it is also the only perceivable biometric for personal identification at a distance. Given the video of an unknown individual, we wish to use gait as a cue to find who person is among the individuals in the database. For a normal walk, gait sequences are repetitive and exhibit nearly periodic behavior. As gait databases continue to grow in size, it is conceivable that identifying a person only by gait may be difficult. However, gait can still
Velammal College of Engineering and Technology, Madurai
Furthermore, unlike faces, gait is also difficult to conceal. Similar in principle to other biometrics, gait is affected by some physical factors such as drunkenness, pregnancy and injuries involving joints. Ideally, the recognition features extracted from images should be invariant to factors other than gait, such as color, texture, or type of clothing. In most gait recognition approaches recognition features are extracted from silhouette images. Approaches in computer vision to the gait recognition problem can be broadly classified as being either modelbased or model-free. Both methodologies follow the general framework of feature extraction, feature correspondence and high-level processing. The major difference is with regard to feature correspondence between two consecutive frames. Model based approach assumes a priori models that match the two- dimensional (2-D) image sequences to the model data. Feature correspondence is automatically achieved once matching between the images and the model data is established. Examples of this approach include the work of Lee et al., where several ellipses are fitted to different parts of the binarized silhouette of the person and the parameters of these ellipses such as location of its centroid, eccentricity, etc. are used as a feature to represent the gait of a person. Recognition is achieved by template matching. Model-free methods establish correspondence between successive frames based upon the prediction or estimation of features related to position, velocity, shape, texture, and color. Examples of this approach include the work of Huang et al., whose optical flow to derive a motion image sequence for a walk cycle. Principal Components Analysis (PCA) is then applied to the binarized silhouette to derive what are called Eigen gaits. Alternatively, they assume some implicit notion of what is being observed Little and Boyd describe the shape of the human motion with the scale-independent features from moments of the dense optical flow, recognized the individuals by phase vectors estimated from the feature sequences. Sundaresan et al proposed a Hidden Markov Model (HMM) based on individual recognition by gait. We considered two
Page 123
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The third aspect is the distance measure used, which can be Euclidean simple dot product-based based on probabilistic models. The second class of approaches opts for parameters that can be used to characterize gait dynamics, such as stride length, cadence, and stride speed. Sometimes static body parameters, such as the ratio of sizes of various body parts are considered in conjunction with these parameters. However, these approaches have not reported high performances on common databases, partly due to their need for 3D calibration information. The third class of approaches emphasizes the
silhouette shape similarity and disregards or underplays temporal information. While extracting the silhouette shape there are two options. 1) To use the entire silhouette 2) To use only the outer contour of the silhouette. The choice of using either of the above features depends upon the quality of the binarized silhouettes. If the silhouettes are of good quality, the outer contour retains all the information of the silhouette and allows a representation, the dimension of which is an order of magnitude lower than that of the binarized silhouette. However, for low quality, low resolution data, the extraction of the outer contour from the binarized silhouette may not be reliable. In such situations, direct use of the binarized silhouette may be more appropriate. When compared with other two approaches the silhouette shape-based approaches have been used by various researchers in determining the gait of a person. Sundaresan et al proposed a HMM on individual recognition by gait. II. COMPARISON OF USE OF HMM In this paper we compared how the HMM is used to determine the Gait of each person and to recognize them. Zongyi Liu and Sudeep Sarkar have used a generic walking model as captures by a population Hidden Markov Model (pHMM) defined for a set of individuals. The states of this pHMM represent gait stances over one gait cycle and the observations are the silhouettes of the corresponding gait stances. For each sequence, they first used Viterbi decoding of the gait dynamics to arrive at one dynamics-normalized, averaged, gait cycle of fixed length. The distance between two sequences is the distance between the two corresponding dynamics-normalized gait cycles, which they quantify by the sum of the distances between the corresponding gait stances. Distances between two silhouettes from the same generic gait stance are computed in the linear discriminant analysis space so as to maximize the discrimination between persons, while minimizing the variations of the same subject under different conditions. The distance computation is constructed so that it is invariant to dilations and erosions of the silhouettes. Thus they handled variations in silhouette shape that can occur with changing imaging conditions. The results were presented on three different, publicly available, data sets. First, the HumanID Gait Challenge data set, which is the largest gait benchmarking data set that is available (122 subjects), exercising five different factors, i.e., viewpoint, shoe, surface, carrying condition, and time. It was significantly improve the performance across the hard experiments involving surface change and briefcase carrying conditions. Second, improved performance on the UMD gait data set that exercises time variations for 55 subjects. Third, on the CMU Mobo data set, we show results for matching across different walking speeds. It is being concluded that worth noting that there was no separate training for the UMD and CMU data sets. Ju Han and Bir Bhanu, they have consider individual recognition by activity specific human motion, i.e., regular human walking, which is used in most current approaches of
Velammal College of Engineering and Technology, Madurai
Page 124
image features, one being the width of the outer contour of the binarized silhouette, and the other being the binary silhouette itself. A set of exemplars that occur during a gait cycle is derived for each individual. To obtain the observation vector from the image features we employ two different methods. In the indirect approach the high-dimensional image feature is transformed to a lower dimensional space by generating the frame to exemplar (FED) distance. The FED vector captures both structural and dynamic traits of each individual. For compact and effective gait representation and recognition, the gait information in the FED vector sequences is captured using a HMM model for each individual. In the direct approach, we work with the feature vector directly and train an HMM for gait representation. The difference between the direct and indirect methods is that in the former the feature vector is directly used as the observation vector for the HMM whereas in the latter, the FED is used as the observation vector. In the direct method, we estimate the observation probability by an alternative approach based on the distance between the exemplars and the image features. In this way, we avoid learning highdimensional probability density functions. The performance of the methods is tested on different databases. II. HIDDEN MARKOV MODEL (HMM) Gait recognition approaches are basically of three types: 1) temporal alignment-based, 2) static parameterbased, and 3) silhouette shape-based approaches. Here we have discussed the silhouette shape-based approach. The first stage is the extraction of features such as whole silhouettes, principal components of silhouette boundary vector variations, silhouette width vectors, silhouette parts or Fourier Descriptors. The gait research group at the University of Southampton (Nixon et al.) has probably experimented with the largest number of possible feature types for recognition. This step also involves some normalization of size to impart some invariance with respect to distance from camera. The second step involves the alignment of sequences of these features, corresponding to the given two sequences to be matched. The alignment process can be based on simple temporal correlation dynamic time warping hidden Markov models phase locked-loops or Fourier analysis.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of images on the stance set creating a vector (Frame-to-Stance Distance or FSD) representation for each frame and use these samples to train an HMM model using the Baum-Welch algorithm .The viterbi algorithm is used in the evaluation phase to compute the forward probabilities. The absolute values of the log probability values are recorded as the similarity scores. IV .CONCLUSIONS This paper gives an analysis of how Hidden Markov Model (HMM) used for gait identification by various researcher. This provides the effectiveness and computational efficiency for real world applications. It’s being proved that the HMM is mostly used to obtain the silhouette from the images. Though there are different models are available, HMM provides an efficient way to extract the silhouette.
individual recognition by gait. They have proposed a new spatio-temporal gait representation, called Gait Energy Image (GEI), to characterize human walking properties for individual recognition by gait. To address the problem of the lack of training templates, they also proposed a novel approach for human recognition by combining statistical gait features from real and synthetic templates. They directly compute the real templates from training silhouette sequences, while generating the synthetic templates from training sequences by simulating silhouette distortion. They have used a statistical approach for learning effective features from real and synthetic templates. We compare the proposed GEI-based gait recognition approach with other gait recognition approaches on USF HumanID Database. Experimental results show that the proposed GEI is an effective and efficient gait representation for individual recognition, and the proposed approach achieves highly competitive performance with respect to the published gait recognition approaches. The fundamental assumptions made here are: 1) the order of poses in human walking cycles is the same, i.e., limbs move forward and backward in a similar way among normal people, and 2) differences exist in the phase of poses in a walking cycle, the extend of limbs, and the shape of the torso, etc. Under these assumptions, it is possible to represent spatiotemporal information in a single 2D gait template instead of an ordered image sequence. Naresh Cuntoor, Amit Kale and Rama Chellappa has analyzed three different sets of features are extracted from the sequence of binarized images of the walking person. Firstly, we investigate the swing in the hands and legs. Since gait is not completely symmetric in that the extent of forward swing of hands and legs is not equal to the extent of the backward swing, we build the left and right projection vectors. To match these time varying signals, dynamic time warping is employed. Secondly, fusion of leg dynamics and height combines results from dynamic and static sources. A Hidden Markov model is used to represent the leg dynamics. While the above two components consider the side view, the third case explores frontal gait. We characterize the performance of the recognition system using the cumulative match scores computed using the aforesaid matrix of similarity scores. As in any recognition system, we would like to obtain the best possible performance in terms of recognition rates. Combination of evidences obtained is not only logical but also statistically meaningful. We show that combining evidence using simple strategies such as Sum, Product and MIN rules improves the overall performance. To compare the truncated width vectors that contain the information about leg dynamics, we use HMM, which is a generalization of the DTW framework. There exists Markovian dependence between frames since the way humans go through the motion of walking has limited degrees of freedom. K-means clustering is used to identify ’key frames’ or ’stances’ during a half-cycle. We found that a choice of is justified by the rate-distortion curve. We project the sequence
REFERENCES [1]. Zongyi Liu and Sudeep Sarkar,”Improved Gait Recognition by Gait Dynamics normalization” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 28, No. 6, June 2006. [2]. Murat EK_INC_I,” Human Identification Using Gait”, Turk J Elec Engin, VOL.14, NO.2 2006, c T¨UB_ITAK. [3].Naresh Cuntoor, Amit Kale and Rama Chellappa, Combining Multiple Evidences for Gait Recognition. [4]. Rong Zhang, Christian Vogler, Dimitris Metaxas, “Human Gait Recognition” [5]. S. Sarkar, P.J. Phillips, Z. Liu, I. Robledo-Vega, P. Grother, and K.W. Bowyer, “The Human ID Gait Challenge Problem: Data Sets, Performance, and Analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 27, no. 2, pp. 162-177, Feb. 2005 [6]. Z. Liu and S. Sarkar, Simplest Representation yet for Gait Recognition: Averaged Silhouette, Proc. Int’l Conf. Pattern Recognition, vol. 4, pp. 211-214, 2004. [7]. D. Tolliver and R. Collins, Gait Shape Estimation for Identification, Proc. Int’l Conf. Audio and Video-Based Biometric Person Authentication, pp. 734-742, 2003. [8]. L. Wang, T. Tan, H. Ning, and W. Hu, “Silhouette Analysis-Based Gait Recognition for Human Identification,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 25, no. 12, pp. 15051518, Dec. 2003 [9]. R. Collins, R. Gross, and J. Shi, “Silhouette-Based Human Identification from Body Shape and Gait,” Proc. Int’l Conf. Automatic Face and Gesture Recognition, pp. 366-371, 2002.
Velammal College of Engineering and Technology, Madurai
Page 125
V. FUTURE WORKS HMM can be compared with the other models namely Principal Component Analysis (PCA),Baseline Algorithm etc their efficiency can be observed with respect with the other two model which is being discussed here. Also efficiency of the algorithm can be checked with various factors like time, surface, carrying luckages, under various climate conditions can be analyzed with the same Model (HMM).
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[10].L. Lee and W. E. L. Grimson, “Gait analysis for recognition and classification,” in Proc. IEEE Conf. Face and Gesture Recognition, 2002, pp.155–161. [11]. J. Little and J. Boyd, “Recognizing people by their gait: The shape of motion,” Videre, vol. 1, no. 2, pp. 1–32, 1998. [12]. J. Cutting and L. Kozlowski, “Recognizing friends by their walk: gait perception without familiarity cues,” Bull. Psycho. Soc., vol. 9, pp. 353–356, 1977. [13]. M. P. Murray, “Gait as a total pattern of movement,” Amer. J. Phys. Med., vol. 46, pp. 290–332, June 1967. [14]. M. P. Murray, A. B. Drought, and R.C. Kory, “Walking patterns of normal men,” J. Bone and Joint Surgery, vol. 46A, no. 2, pp.335–360, 1964.
Velammal College of Engineering and Technology, Madurai
Page 126
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Congestion Management Routing Protocol In Mobile Adhoc Networks A. Valarmathi1 and RM. Chandrasekaran2 1Department of Computer Applications, Anna University Tiruchirappalli, Tiruchirappalli -620 024, India E-mail: [email protected] 2Department of Computer Science and Engineering, Anna University Tiruchirappalli, Tiruchirappalli-620 024, India :[email protected] Abstract Streaming multimedia applications over Mobile Ad hoc Networks (MANETs) require a guaranteed or controlled load service. Network congestion is the dominant reason for packet loss, longer delay and jitter in streaming multimedia applications. Most of the present routing protocols are not designed to adapt congestion. In the present paper, the original DSR protocol was modified to monitor the occurrence of congestion by using multiple resource utilization thresholds as QoS attributes and trigger multi-pathrouting during the periods of congestion to improve QoS in MANETs. This paper also deals with the effects of mobility on the modified routing protocol. Simulation experiments are conducted using NS-2 network simulator to evaluate the performance of the modified DSR in terms of jitter and percentage packet loss and compared with the original DSR. The results showed that a significant improvement in performance of modified DSR was achieved and decreases the network congestion with the use of multi-path routing. Keywords: Mobile ad hoc network, DSR, Congestion, Multi-path routing, NS-2.
1. INTRODUCTION Mobile ad-hoc networks (MANETs) are autonomous dynamic networks which provide high flexibility. These networks uses multi-hop radio relay that operate without the support of any fixed infrastructure [1]. Streaming multimedia or CBR applications over MANET require minimal delay and packet loss. To meet these critical requirements, a MANET inherently depends on the routing scheme employed. Routing protocols for Ad hoc networks can be broadly classified as proactive and reactive. Proactive (or) table-driven routing algorithms employ distance vector based or link state based routing strategies. However, the main drawback of these algorithms is that the need for frequent table updation consumes significant amount of memory, bandwidth and battery power [2]. Example of such protocols is Optimized Link State Routing (OLSR) [3] and Destination Sequenced Distance Vector routing (DSDV) [4]. In reactive routing
Velammal College of Engineering and Technology, Madurai
protocols, a node searches and maintains a route to another node only when it requires transmitting data to that particular node. This on-demand reactive approach minimizes bandwidth utilization compared to proactive routing protocols. Ad-hoc On-Demand Distance Vector (AODV) [5], Dynamic Source Routing (DSR) [6], Temporally Ordered Routing Algorithm (TORA) [7] and Associatively-Based Routing (ABR) [8] are examples of such protocols. Among these, DSR is a simple and efficient routing protocol which is widely used in MANET. It operates based on two mechanisms such as route discovery and route maintenance. All these protocols rely on single-path routing algorithms, where all traffic to a destination is required to be routed through a single successor and when a link become congested, the entire traffic has to be rerouted. Routing protocols for MANET also have another dimension as congestion-un-adaptive routing and congestion adaptive routing. Existing routing protocols belong to the first group where congestion nonadaptiveness in routing may result in longer delay, higher overhead and many packet losses for traffic intensive data like multimedia applications. DSR does not make dynamic routing adjustments based on memory, delay, number of service requests or battery utilization information which often indicates certain levels of usage and possible occurrences of congestion. Detection of undesired paths enables the ability to implement route admission control and to perform multi-path routing scheme which could further improve the performance of the network in a significant manner. In the present paper, a modified DSR protocol is proposed to monitor congestion or multiple resource utilization levels, notify the routing process of congestion and invoke multi-path routing during the periods of congestion for streaming multimedia applications. In addition, this paper also determine the effects of various mobility(low, medium, high) models on the modified routing protocol and compare the performance metrics of jitter and percentage packet loss of standard DSR protocol with modified version of DSR protocol.
Page 127
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2. RELATED WORK The research on congestion for MANET is still in the early stage and there is a need of new techniques. In this section, the research work related to congestion control and multipath routing is presented. Round trip time measurements are used to distribute load between paths in Multi-path Source Routing (MSR) [9]. A distributed multi-path DSR protocol (MP-DSR) was developed to improve QoS with respect to end-to-end reliability [10]. The protocol forwards outgoing packets along multiple paths that are subjected to an end-to-end reliability model. Split Multi-path Routing (SMR) utilized multiple routes of maximally disjoint paths which minimize route recovery process and control message overhead [11-12]. The protocol uses a per-packet allocation scheme to distribute data packets into multiple paths of active sessions which prevents nodes from being congested in heavily loaded traffic situations. Kui Wu and Janelle Harms [13] proposed a path selection criteria and an on-demand multi-path calculation method for DSR protocol. Peter Pham and Sylvie Perreau [14] proposed a multi-path DSR protocol with a load balancing policy which spreads the traffic equally into multiple paths that are available for each source-destination pair. A dynamic load-aware based load-balanced routing (DLBL) algorithm was developed which considers intermediate node routing load as the primary route selection metric [15]. This helps the protocol to discover a route with less network congestion and bottlenecks. When a link breaks because of the node mobility or power off, DLBL provides efficient path maintenance to patch up broken links to help to get a robust route from the source to the destination. A simple Loop-Free Multi-Path Routing (LFMPR) with QoS requirement was developed from AODV and DSR protocol [16]. In the route request phase, intermediate nodes record multi-reverse links which is applied to construct multiple-paths during the route reply phase. Each path is assigned unique flow identification in order to prevent routing loop problems. Rashida Hashim et al [17] proposed an adaptive multi-path QoS aware DSR Protocol. The protocol collects information about QoS metrics during the route discovery phase and uses them to choose a set of disjoint paths for data transmission. De Rango et al [18] proposed an energy aware multi-path routing protocol by considering minimum drain rate as a metric. An update mechanism and a simple data packet scheduling among the energy efficient paths have also been implemented to update the source route cache and for improving the traffic and energy load balancing. Raghavandra et al [19] proposed Congestion Adaptive Routing in Mobile Ad Hoc Networks (CRP). In CRP every node appearing on a route warns its previous node when prone to be congested. The previous node then uses a “bypass” route bypassing the potential congestion to the first non-congested node on the route. Traffic will be split probabilistically over these two routes, primary and bypass, thus effectively lessening the chance of congestion occurrence. Zhang XiangBo and Ki-Il Kim [20] proposed a multi-path routing protocol based on DSR which uses Multi-Probing and Round-Robin mechanisms (MRDSR) for updating alternate multiple paths.
Velammal College of Engineering and Technology, Madurai
The different multi-path schemes proposed earlier concentrates few of the performance aspects. For example, the measurement of jitter is very scarce, which is a critical performance metric for multimedia traffic. Moreover, the performance of a multi-path routing scheme should be quantitatively evaluated in different mobility scenarios to completely assess the applicability. 3. CONGESTION MANAGEMENT The existing DSR protocol was modified to perform multipath routing with congestion adaptive feature. In the proposed protocol, the battery level and queue length are used as the key resource utilization parameters. In the present paper, the congestion was assumed to be in existence when queue length was near capacity or when battery level fell below a predefined threshold. The proposed protocol works similar to normal DSR protocol if the current energy level is between 10 to 100 % of initial energy level and queue length is less than 50. The modified DSR algorithm invokes multi-path routing when it exceeds these threshold values. 4. SIMULATION SET-UP NS-2 version 2.30 [23] was used to simulate the performance of the modified DSR under different mobility scenarios. The 802.11 MAC protocol is defined to be the wireless channel. Each mobile node used a Two-Ray Ground radio propagation model with an Omni antenna. The initial battery capacity of each node assumed to be 100%. The battery capacity of a mobile node was decremented in a predefined manner by the txPower and rxPower levels which remains constant throughout the simulation. The CTS / RTS process used in IEEE 802.11b within NS2 normally generate huge volumes of data that made processing the trace files complicated. In order to restrict the processing of unnecessary data, the RTSThreshold_ variable was set to 3,000 bytes which restrict the processing of packets that have a size greater than 3,000 bytes. The random topology consists of 800 x 800 meter grid with 20 nodes. The packet sizes of CBR multimedia streams were fixed at 512 bytes with each node had a maximum queue size of 50 packets. The simulation time was set to be 400 seconds and the speed of the node is 10 m/s. Three different simulation scenarios of high, medium and low were created by varying pause times at 4, 10 and 20 seconds. The workload and data rate was kept constant at 5 CBR and 200 PPS, respectively. 5. PERFORMANCE METRICS In the present paper, the performance metrics such as average jitter and percentage packet loss was calculated and evaluated for both normal DSR and modified DSR using the following relationships. Average Jitter Jitter is a measure of variation in delay across multiple packets associated with a given traffic flow. In streaming multimedia applications, only a small and limited amount of jitter is tolerable.
Page 128
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Percentage Packet loss % Packet loss = Number of packets sent - Number of packets received / Number of packets sent *100
reduction in percentage packet loss in high mobility than in low and medium mobility scenario further confirms the effective usage of multi-path routing in high mobility scenario.
6. RESULTS AND DISCUSSION 6.1
AVERAGE JITTER
Durkin [22] stated that excessive jitter normally caused by unavailable data may lead to discontinuity in streaming multimedia applications. For example, an excessive jitter might result a video application to experience frozen screens. The computed average jitter values for different mobility scenarios are shown in Figure 1. In low and high mobility scenarios, the performance of modified DSR is only in comparable with original DSR. In medium mobility scenario, the measured jitter for modified DSR is higher than the original DSR.
Fig. 2. Percentage packet loss of original and modified DSR at different mobility scenarios. 7. CONCLUSION In this paper, Multiple Thresholds based Congestion-adaptive Multi-path Dynamic Source Routing for MANET is proposed which avoids the occurrence of congestion and effective distribution of data traffic in an optimal manner. This protocol exhibits substantial reduction in % packet loss and jitter than original DSR. For original DSR, the percentage packet loss lies between 18 20% at low and medium mobility scenarios. This increases to 36.3 % at high mobility. In modified DSR, the percentage packet loss for low and medium mobility scenarios are approximately 13 %, 5 - 7 % lower than original DSR. REFERENCES
The percentage packet loss of original and modified DSR at different mobility scenarios is shown in Figure 2. For original DSR, the percentage packet loss lies between 18 - 20% at low and medium mobility scenarios. This increases to 36.3 % at high mobility. This clearly indicates the effect of route failure at high mobility which causes excessive packet loss. In modified DSR, the percentage packet loss for low and medium mobility scenarios are approximately 13 %, 5 - 7 % lower than original DSR. In high mobility, the percentage packet loss is 20 % lower than original DSR. This drastic
[1] C. S. R. Murthy, B. Manoj, “Ad hoc wireless networks Architectures and protocols,” special edition, Printice Hall, 2004. [2] J. Macker, M. Corson and V. Park, “Mobile and wireless internet services: Putting the pieces together,” IEEE CommunicationsMagazine.36, 2001, pp. 146-155. [3] T. Clausen and P. Jacquet, “Optimized link state routing protocol,” IETF RFC 3626, Network Working Group, October 2003. [4] C. E. Parkens and P. Bhagwat, “Highly dynamic Desination-Sequenced Distance-Vector Routing (DSDV) for mobile computers,”Computer Communications Review 24 (1994) ,234-244. [5] C. Perkins, E. Belding-Royer and S. Das, “Ad hoc ondemand Distance Vector Routing,” RFC 3561, July 2003. [6] D. B. Johnson, D. A. Maltz and J. Broch, “DSR- The dynamic source routing protocol for multi hop wireless ad hoc networks,” in: C. E. Perkins (Eds.), Ad hoc Network, Chapter 5, Addison-Wesley, 2001.
Velammal College of Engineering and Technology, Madurai
Page 129
Fig. 1. Average Jitter of original and modified DSR at different mobility scenarios. 6.2
PERCENTAGE PACKET LOSS
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[7] S. Corson and V. Park, “Temporally-Ordered Routing Algorithm (TORA) Version 1 Functional Specification,” Mobile Ad Hoc Network (MANET) Working Group, IETF, Oct. 1999. [8] C. K. Toh, “A novel distributed routing protocol to support ad-hoc mobile computing,” Proceeding of IEEE 15th Annual International Phoenix Conference on Computers and Communication, 1996, pp. 480-486. [9] Lei Wang, Lianfang Zhang, Yantai Shu, Miao Dong, Multi-path source routing in wireless ad hoc networks, 2000 Canadian Conference on Electrical and Computer Engineering. 1 (2000) 479-483. [10] R. Leung, Jilei Liu, E. Poon, A.-L.C. Chan and Baochun Li, “MP-DSR: a QoS-aware multi-path dynamic source routing protocol for wireless ad-hoc networks”, Proc. 2001 Local Computer Networks, 26th Annual IEEE Conference, 2001, pp.132-141. [11] Sung- Ju Lee and Mario Gerla, “Dynamic load-aware routing in ad hoc networks‟, IEEE Conference on Communications. 10 ,2001,pp. 3206-3210. [12] Sung- Ju Lee and Mario Gerla, “Split Multi-path Routing with Maximally Disjoint Paths in Ad Hoc Networks,” Proceedings of IEEE International Conference on Communications, 2001, pp. 3201-3205. [13] K. Wu and J. Harms, “Performance Study of a Multi-path Routing Method for Wireless Mobile Ad Hoc Network,“ Proceedings of 9th IEEE International Symposium on modeling, Analysis, and Simulation of Computer and Telecommunication Systems MASCOTS'01),2001, PP. 1-7. [14] Peter Pham and Sylvie Perreau, “Multi-path routing protocol with load balancing policy in mobile ad hoc network,” 4th International Workshop on Mobile and Wireless Communications Network ,2002, pp. 48-52. [15] Xiangquan Zheng, Wei Guo, Renting Liu and Yongchun Tian, “A New Dynamic Loadaware Based Load-balanced Routing for Ad Hoc Networks,” IEEE, 2004, pp. 407-411. [16] Cheng-Ying Yang, Yi-Wei Ting and Chou-Chen Yang, “A Loop-Free Multi-Path Routing With QoS for Wireless Ad Hoc Network,” 20th International Conference on Advanced Information Networking and Applications. 2, 2006, pp. 179185. [17] Rashida Hashim, Qassim Nasir, Saad Harous, Adaptive Multi-path QoS Aware Dynamic Source Routing Protocol for Mobile Ad-Hoc Network, Innovations in Information Technology. 2006, pp. 1-5. [18] F. De Rango, P. Lonetti S. Marano, “Energy-aware metrics impact on Multi-path DSR in MANETs environment,“ International Symposium on Performance Evaluation of Computer and Telecommunication Systems, 2008,pp. 130137. [19] Duc A. Tran and Harish Raghavendra, “Congestion Adaptive Routing in Mobile Ad Hoc Networks, “ IEEE transactions on parallel and distributed systems. 17, 2006, pp. 1294-1305.
[20] Hang Xiang Bo and Ki-Il Kim, “Load-aware metric for efficient balancing on multi-path DSR protocol in Mobile Ad hoc Networks,” International Conference on Advanced Technologies for Communications. 2008, pp. 395-398. [21] K. Samdanis, A. Aghvami and V. Friderikos, “Quality of service routing (QoSR) interactions with the BRAIN candidate micro-mobility protocol (BCMP),“ Proceedings of 2nd International Workshop on Mobility Management and Wireless Access Protocols, 2004, pp. 35-42. [22] Durkin, Voice-enabling the data network: H.323, MGCP, SIP, QoS, SLAs, and security, first ed., Cisco Press, Indianapolis, IN, 2003. [23] The Network Simulator NS-2, http://www.isi.edu/nsnam/ns/ [24] N. Reide and R. Seide, 802.11 (Wi-Fi) Networking Handbook, first ed., McGraw-Hill, CA, 2003.
Velammal College of Engineering and Technology, Madurai
Page 130
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Human Motion Tracking And Behaviour Classification Using Multiple Cameras
M.P.Jancy#, B.Yogameena$ Research Scholar, $Assistant Professor Dept. of Electronics and Communication Engineering, Thiagarajar College of Engineering, Madurai.India [email protected], [email protected] #
Abstract- This paper presents a framework for human behaviour understanding using multi-camera system. The proposed methodology classifies human behaviour as normal or abnormal, by treating short-term behaviours such as walking, running, standing still, etc. The analysis includes foreground segmentation, tracking and feature extraction from the image sequences followed by the classification of short term behaviours. A set of calculated feature vectors are input to the Relevance Vector Machine which works on human behaviour classification. Experimental results demonstrate the proposed method to be highly accurate, robust and efficient in the classification of normal and abnormal human behaviours. Keywords - surveillance, feature extraction,, relevance vector machine, behaviour understanding.
I. INTRODUCTION Multicamera surveillance system aims at tracking people or object of interest. The fundamental problem in video surveillance is automatic detection of human motion, tracking human and classifying abnormal situations. Based on this motivation, this paper proposes an efficient approach for tracking people through multicamera system. It creates a model for normal human behaviour and any deviation from the basic model is analysed. Our methodology applies on short term behaviour referring those that can be localized in a spatio-temporal sense, i.e. brief and within a restricted space. Examples of such behaviours are walking, standing still, running, moving abruptly, waving hand etc. The algorithm provides clear discrimination of such anomalies. The paper is organized into five sections. Section II discusses the algorithms available in the present day. Section III describes the methodology Section IV describes background subtraction, extraction of feature and RVM learning system for classification are explained. Section V depicts the experimental results and the classification of human using the Relevance Vector Machine extraction method. Finally, the conclusion is presented.
Machines [4], Radial Basis Function Networks, Nearest Neighbor Algorithm, and Fisher Linear Discriminant and so on. Machine learning techniques are linear methods in a high dimensional feature space, nonlinearly related to the input space. Using appropriate kernel functions, it is possible to compute the hyper plane which classifies the two classes. An automated system based on SVM classifying the human activities as normal or abnormal, utilizes a set of feature vector data has been described by Scholkopf [5]. SVM needs the tuning of a regularization parameter during the training phase, requiring a technique which minimizes the number of active kernel functions to reduce computation time ([6]-[8]). Appropriate features for RVM are used to classify and thereby reduces the computational complexity by selecting an appropriate kernel to improve the classification accuracy. Experimental results demonstrated that the proposed approach is robust in classifying human behaviour. III. PROPOSED METHOD Surveillance system is divided into low level and high level process. The proposed methodology is shown in fig.1 based on the fusion of data collected from several cameras with overlapping fields of view. The low level addresses the problem of motion detection and blob analysis, whereas the upper level addresses the feature vectors. Background subtraction is applied for motion detection to obtain the foreground blob and its corresponding bounding box is extracted. Additional object information namely the object’s blob size, shape, centroid, edge, area, and skeleton are obtained. From the object information the feature vectors are calculated which are input to the Relevance Vector Machine Classifier.Motion-based techniques are mostly used for shortterm activity classification. Examples are walking, running, fighting. These techniques calculate features of the motion itself and perform recognition of behaviours based on these feature values.
II. EXISTING APPROACHES There have been a number of video surveillance systems, for multiple people detection and tracking [1], [2] in a crowded Environment [3]. All these systems can detect and track multiple people. There are many types of neural networks that can be used for binary classification problem, such as Support Vector
Velammal College of Engineering and Technology, Madurai
Page 131
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Video
Segmentation Frame 10
Foreground Blob
Frame 90
Frame 115
Fig.2 Background subtracted image for camera1
Low level High level
Feature Vectors
Frame 05
A. FEATURE CALCULATION Feature vectors are computed by taking into account both the background subtraction and the ground plane information. Object’s centroid is determined by using the following equations
Final Decision [Normal/Abnormal] Fig.1 System Overview
Motion-based techniques are mostly used for shortterm activity classification. Examples are walking, running, fighting. These techniques calculate features of the motion itself and perform recognition of IV. BACKGROUND SUBTRACTION The first stage of video surveillance system seeks background to automatically identify people, objects, or events of interest in various changing environments. Background substraction is accomplished in real-time using the adaptive mixture of Gaussians method proposed by Stauffer [9]. It describes K Gaussian distributions to model the surface reflectance value. k
P ( X t ) = ∑ ω i ,t *η ( X t , μ i ,t , ∑ i, t )
…(1)
where K is the number of distributions, ωi,t is an estimate of Gaussian, μ is the mean, th
∑ i, t
is the
covariance matrix of the i Gaussian. After ordering the Gaussians, the first B distributions are chosen as the background model as shown in eqn (2) b
B = arg min b (∑ ω k ⟩T )
xc = yc =
1 N 1 N
N
∑x j =1
N
∑
j =1
…(3)
i
yi
…(4) where (xc,
yc) represent the average contour pixel position, (xi, yi) represent the points on the human blob contour and there are a total of N number of points on the contour [16]. Boundary extraction is the minimal rectangle that surrounds the borderline. The rectangle’s height is expressed as
Y
max
−Y min
...(5)
And its width is expressed as
i =1
the weight of the i
Frame 15
Fig.3 Background subtracted image for camera2
Relevance Vector Machine
th
Frame 85
...(2)
k =1
where, T is a measure of minimum models for background. The background subtracted image obtained using adaptive GMM is shown in Fig.2 and Fig.3.
X
−X
...(6) Skeleton feature is obtained by calculating the distance max
min
di
from the centroid to contour points. 2 2 d i…(7) = ( xi − xFrom ( yi d− yplot c ) +the c ) the local maximum and i
minimum points are collected and star skeletonization is formed. These points are selected as features and fed into RVM learning system to classify normal and abnormal behaviour of human. B. CLASSIFICATION USING RELEVANCE VECTOR MACHINE The RVM is a Bayesian regression framework, in which the weights of each input example are governed by a set of hyper parameters [8]. These hyper parameters describe the posterior
Velammal College of Engineering and Technology, Madurai
Page 132
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
distribution of the weights and are estimated iteratively during training. Most hyper parameters approach infinity, causing the posterior distributions of the effectively setting the corresponding weights to zero. The remaining vectors with non-zero weights are called relevance vectors. It reduces the “inappropriate” parameters for the classification than SVM [[11], [12]]. In this framework, the attention is focused to the nature of the behaviour of a human to classify his activities as normal or abnormal. The walking pose of a person is considered as an articulated object, and is represented by a parameter vector x, given in eqn (15) x = Wk φ (z) + ξ k …. (15) where, x is the input for the system, Wk is the weight of the basis function, ф(z) is the vector of the basis function and ξk is the Gaussian noise vector. It is used to minimize the cost function of the RVM regression using eqn (16)
Fig.4 Edge image for camera1 and camera 2
Fig.5 Centroid and Bounding Box for camera1 and camera2.
LK = ∑ C K( n ) ( y k(n ) ) S k ( y k(n ) ) ……. (16) T
( )
y k(n ) = x (n ) − W K φ Z (n )
……… (17) where, yk is the output with n sample points belonging to the mapping function k. Φ(z(n)) is the design matrix vector of the basis function, Sk is the diagonal covariance matrix of the basis function and Ck(n) is the probability that the sample point n belongs to the mapping function k. (n)
Fig.6 skeleton feature for human
V.EXPERIMENTAL RESULTS FOR CLASSIFICATION In this work, the video is taken in Indoor and Outdoor . The proposed method processes about 24 frames per second for RGB images and the total number of frames is equal to 3600, and the frame size is 152 × 172 on PC with a 2.5Ghz in a Pentium IV CPU to demonstrate the performance of the proposed method. It classifies the behaviour as normal or abnormal efficiently, by considering the human postures or actions that are explicitly visible over a larger field of view with the multi-camera system. Thus it overcomes the constraint of a single camera system where the human postures or actions may be misinterpreted due to a single field of view which leads to incorrect classification. In this proposed method, the input video file is first converted into frames. Then the Gaussian Mixture Model background subtraction method is used to separate the foreground objects from the background scenes. After extracting the background image, the edges are extracted using the canny filter and its shown in fig.4. The features such as boundary, centroid, area, edge and skeleton are calculated and are shown in fig.5 and fig.6. These features are input to the classifier. Finally the RVM classifier distinguishes the normal and abnormal behaviour.
Velammal College of Engineering and Technology, Madurai
Fig.7 Training stage
Fig.8 Classified RVM output
Page 133
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The Fig.7 shows the initial stage of the training. Then the Fig 8 shows the result after the iteration. The circles in the figure denote the relevance vector obtained after the iteration. The maximum number of iterations used here is Twenty Five. Thus the resultant vector obtained here is closer to the original template and there is no probability of misclassification of vectors. This is the main advantage in this proposed method. Normal behaviour classification is shown in Fig 9. and the datasets were taken from indoor and from caviar benchmark dataset. The algorithm has been tested using 2100 samples. Abnormal behaviour classification of a person bending down and a person waving his hand are shown in Fig 10. . VIII. CONCLUSION This paper has introduced the framework of relevance vector machine for the classification of different kind of human activities for the application of video surveillance. To this end we have classified the given image sequence as standing, running and so on. Results demonstrate the proposed method to be highly accurate, robust and efficient in the classification of normal and abnormal human behaviours. REFERENCES 1. M. A. Ali, S. Indupalli and B. Boufama,” Tracking Multiple People for Video Surveillance”, School of Computer Science, University of Windsor,Windsor, ON N9B 3P4, 2004 2. Mykhaylo Andriluka Stefan Roth Bernt Schiele,”People-Trackingby-Detection and People-Detection-by-Tracking” Computer Science Department Technische Universit¨at Darmstadt,2007. 3. Tao Zhao Ram Nevatia,“Tracking Multiple Humans in Crowded Environment”, Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’04)1063-6919/04 $20.00 © 2004 IEEE. 4. Xinyu WU, Yongsheng OU, Huihuan QIAN, and Yangsheng XU, “Detection System for Human Abnormal Behavior”,2004. 5. B. Scholkopf, J.C. Platt, J. Shawe-Taylor, A.J. Smola, and R.C. Williamson.Estimating the Support of a High-Dimensional Distribution. Neural Computation, 13(7):1443–1471, 2001. 6. Rong Zhang, Christian Vogler, Dimitris Metaxas , “Human gait recognition at sagittal plane” Image and Vision Computing, Vol 25, Issue 3, pp 321-330, March 2007. 7. Carl Edward Rasmussen, “The Infinite Gaussian Mixture Model” Advances in Neural Information Processing Systems, pp. 554–560, 2004. 8. Prahlad Kilambi, Evan Ribnick, Ajay J. Joshi, Osama Masoud, Nikolaos Papanikolopoulos, “Estimating pedestrian counts in groups”, Computer Vision and Image Understanding, pp. 43–59,2008. 9. C. Stauffer, W. Eric L. Grimson. Learning patterns of activity using real-time tracking. IEEE Trans. on Pattern Analysis and Machine Intelligence, Volume 22, Issue 8, pp.747-757, 2000. 10. C. BenAbdelkader, R. Cutler, and L. Davis. “Motion-based recognition of people in eigengait space”, Automatic Face and Gesture Recognition Proceedings, Fifth IEEE International Conference, pp.267 – 272, May 2002 11. C.Chang, C.Lin, LIBSVM: a library for support vector machines, Software available at: http://www.csie,ntu.edu.tw/cjlin/libsvmi, vol. 80, 2001, pp.604–611. 12. D. Kosmopoulos, P. Antonakaki, K. Valasoulis, D. Katsoulas, “Monitoring human behavior in an assistive environment using multiple views”, in: 1st International Conference on Pervasive Technologies Related to Assistive Environments PETRA’ 08, Athens, Greece, 2008.
Velammal College of Engineering and Technology, Madurai
Page 134
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hierarchical Zone Based Intrusion Detection System for Mobile Adhoc Networks. D G Jyothi #1, S.N Chandra shekara*2 #Department of Computer Science and Engineering, Bangalore Institute of Technology, Visveswaraiah technological University, Bangalore, India. 1
[email protected]
*Department of Computer Science and Engineering, SJC Institute of Technology , Visveswaraiah technological University, Chickballapur, India 2
[email protected]
ABSTRACT. Intrusion Detection Systems (IDSs) for Mobile Ad hoc Networks (MANETs) are indispensable since traditional intrusion prevention based techniques are not strong enough to protect MANETs. However, the dynamic environment of MANETs makes the design and implementation of IDSs a very challenging task. In this paper, we present a hierarchical ZoneBased Intrusion Detection System (HZBIDS) model that fits the requirement of MANETs. The model utilizes intelligent light weight mobile agent which collects data from the different mobile nodes(Audit data), preprocess the data, alert and alarm messages among HZBIDS local mobile agents and gateway nodes. With alert information from a wider area, gateway nodes’ IDS can effectively suppress many falsified alerts and provide more diagnostic information about the occurring attacks. The model can adjust itself dynamically to adapt to the change of the external environmental. Also the model is robust and Scalable. Keywords – MANET, ADHOC network, Mobile agents,
MANET requirements which is efficient and flexible and ensures scalability of the IDS. In our proposed model, we use a Local IDA which are attached to every node in the zone and global IDA to the gateway nodes of the zone. The rest of this paper is organized as follows: In the next section, we review some related work in intrusion detection for mobile Ad Hoc networks. In Section 3, we present and explain our intrusion detection scheme.
INTRODUCTION The unique characteristics of Mobile Ad hoc Networks (MANETs), such as arbitrary node movement and lack of centralized control, make them vulnerable to a wide variety of outside and inside attacks [1]. How to provide effective security protection for MANETs has become one of the main challenges in deploying MANET in reality. Intrusion prevention techniques, such as encryption and authentication, can determine attackers from malicious behavior. But prevention based techniques alone cannot totally eliminate intrusions. The security research in the Internet demonstrates that sooner or later a smart and determined attacker can exploit some security holes to break into a system no matter how many intrusion prevention measures are deployed. Therefore, intrusion detection systems (IDSs), serving as the second line of defense, are indispensable for a reliable system. IDSs for MANETs can complement and integrate with existing MANET intrusion prevention methods to provide highly survivable networks [1]. Nevertheless, it is very difficult to design a once-for-all detection model. Intrusion detection techniques used in wired networks cannot be directly applied to mobile Ad Hoc networks due to special characteristics of the networks. In this paper, we are concerned with the design of intrusion detection system for MANET. Our goal is to design a new Hierarchical ZBIDS model, derived from
2.RELATED WORK. An IDS for MANET The Intrusion Detection systems architecture for a wireless ad-hoc network may depend on the network infrastructure itself. Wireless ad-hoc networks may be configured in either a flat or multi-layered network infrastructure. In a flat network infrastructure, all nodes are equal; therefore it may be suitable to civilian applications such as virtual classrooms or conferences [5]. On the contrary, in the multi-layered network some nodes considered different. Nodes can be divided into clusters, with one cluster head in each cluster. In traditional wired networks, many intrusion detection systems have been implemented where switches, routers or gateways play key role to make IDS implemented in these devices easily [3]. But on the other hand these devices do not exist in MANET. For that reason several intrusion detection architectures proposed for ad hoc network which include stand-alone IDS architecture, distributed and cooperative IDS architecture, and hierarchical IDS architecture [4]. In a stand-alone Intrusion Detection System architecture, each node runs an IDS independently to detects malicious attacks and determine intrusions, Since stand-alone IDS do not cooperate or share information with other detection systems running on other nodes , all intrusion detection decisions are based on information available to the individual node. Furthermore nodes do not know anything about the situation on other nodes in the same network [5]. Zhang and Lee [1], proposed a distributed and cooperative intrusion detection architecture that is shown in figure 1, in this architecture each node runs an IDS agent and makes local detection decision independently, by monitoring activities such as user and system activities and the communication activities within the radio range. At the same time, all the nodes cooperate in a global detection
Velammal College of Engineering and Technology, Madurai
Page 135
IDS, Gateway node.
1.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
decision making. [7] In this architecture, if IDS agent on any node detects an intrusion with weak or inconclusive evidence, it can initiate a cooperative global intrusion detection procedure or if a node detects locally an intrusion with strong evidence, it can independently determine an attack on the network. Similarly to stand-alone IDS architecture, distributed and cooperative intrusion detection system architecture is more suitable for flat network infrastructure and not multilayered network. Although distributed and cooperative IDS architecture can solve some limitations that exist in the stand-alone IDS architecture, it has the following problems. First, cooperative intrusion detection may lead to unnecessary communication and calculation between nodes in MANET, causing decrease in network performance. Second, the sharing data between trusted nodes is not in general true since there are a lot of possible threats in a wireless network environment [8].
Figure 1: Distributed and Cooperative IDS in MANETs [1].
Hierarchical Intrusion Detection Systems architectures have been designed for multi-layered ad hoc network infrastructures where the network is divided into clusters. Cluster heads (CH) of each cluster usually have more functionality than other node in the clusters, for example cluster-head nodes centralized routing for all nodes in the cluster and routing packets across clusters. Therefore, these cluster heads, act as manage points which are similar to switches, routers, or gateways in traditional wired networks. Each IDS agent runs on every node. Also it is responsible for detecting intrusion locally by monitoring local activities. A cluster head is responsible locally for its node as well as globally for its cluster, e.g. monitoring network packets and initiating a global response when network intrusion is detected [2,6]. The two-level non overlapping Zone-Based Intrusion Detection System (ZBIDS) is proposed to meet the unique requirement of MANETs. With ZBIDS, the network is divided into non overlapping zones and each IDS agent broadcasts the locally generated alerts inside the zone. Gateway nodes (also called inter zone nodes, which have physical connections to nodes in different zones) are responsible for the aggregation and correlation of locally generated alerts. Furthermore, in MANET intrusion detection system there are two types of decision making including
Velammal College of Engineering and Technology, Madurai
collaborative decision making and independent decision making. In collaborative decision making, each node participates actively in the intrusion detection procedure. Once one node detects an intrusion with strong confidence, this node can start a response to the intrusion by starting a majority voting to determine whether attack happens, like in [1]. On the other hand, in the independent decision making framework, certain nodes are assigned for intrusion detection .These nodes collect intrusion alerts from other nodes and decide whether any node in the network is under attack [7, 5]. These nodes do not need other nodes’ participation in decision making. 2.1 ZBIDS Framework It is obvious that local detection alone cannot guarantee satisfactory performance because of limited security information obtained by each IDS agent. What’s more, we may experience alert flooding problems given the distributed nature of MANETs. Therefore, a suitable framework is needed to integrate the alert information from a wider area. Moreover, attacks are likely to generate multiple related alerts. For example, because of the broadcast nature of radio channel, there may exist many victims suffering from same falsified routing control packets. The triggered alerts should have high correlations correspondingly. Therefore, it is desirable to treat them together. Based on the above considerations, we adopt a non overlapping zone based framework. The whole network is divided into non overlapping zones. We assume the existence of such a framework. This could be done easily based on techniques like geographic partitioning [8]. As illustrated in Fig. 2, there are two categories of nodes in ZBIDS: intra zone nodes and gateway nodes (also called inter zone nodes). If one node has a physical connection to a node in a different zone, this node is called a gateway node, for example, node 4, 7, 8 in Fig. 2. Otherwise, it is called an intra zone node. An agent is attached to each mobile node, and all agents collaboratively perform the intrusion detection task. Each IDS agent runs independently to monitor local node’s activities, such as the user behavior, system behavior, radio communication activities, etc. and performs intrusion detection tasks locally. Intra zone nodes will report their locally generated alerts to the gateway nodes in the same zone, and the gateway nodes will aggregate and correlate the received alerts. Gateway nodes in neighboring zones can further collaborate in order to perform intrusion detection tasks in a wider area. There may Fig. 2. The Zone-Based Intrusion Detection System (ZBIDS) for MANETs.
Page 136
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
exist many gateway nodes in a zone, thus avoiding the issue of single point of failure. 2.2 ) Mobile agents for IDS: Mobile agents have been deployed in many techniques for IDSs in MANETs. Due to its ability of moving in network, each mobile agent is considered for performing just one special task and then one or more mobile agents are distributed amongst network nodes. This operation allows the distributed intrusion detection in the system. Some responsibilities are not delegated to every node, and so it helps in reducing the energy consumption, which is also an important factor in MANET network. It also provides for fault tolerance in such a way that if the network is segmented or some of the agents break down, they can still continue to function. In addition, they can work in big and different environments because mobile agents can work irrespective of their architecture, but these systems require a secure module that enables mobile agents to settle down. Moreover, Mobile agents must be able to protect themselves from secure modules on remote hosts. 3.Proposed Intrusion Detection System In this section we describe the intrusion detection system and procedure for intrusion detection. Fig3. IDS for HZBIDS
To other GIDS
LID S
Response Agent
GIDS
Global Association Agent
Data Association Agent Global Detection Agent Detection Agent
To illustrate the system structure in our design of the HZBIDS, there are two major components for each IDS. Local IDS (LIDS) and Global IDS (GIDS). The LIDS Data collection module collects local data from the different mobile nodes. Local Detection Agent (LDA), will preprocess the audit data received, performs the intrusion detection locally. When it detects the intrusion, it initiates the local alarm. Local Data Association Agent (LDAA), aggregates the detection result from different LDA.LRA (Local Response Agent) sends aggregated results to the GIDS of that zone. Only gateway nodes run the GIDS and they are organized in multiple layers iDs GAA, GDA and GRA perform communication between the LIDS and GIDS. GAA maintains the configuration and status of each agent in the framework. .1 Procedure for Intrusion Detection 1. The local alarm is received by the GIDS as an indication of attack which are recognized and sent by the local agents. 2. GDAA aggregates the detection results received from all other mobile node in the same zone and send the combined result to GDA. 3. GDA will invoke some intrusion detection procedures to know the type of intrusion and determine the attacking node and sends intrusion detection result to GRA and LRA. 4. The GRA will keep the attacking node away from the network with the identity based Cryptosystem scheme [9] and broadcast it to all other nodes in the same zone and with the neighboring gateway nodes. 4. Conclusion We have proposed Hierarchical Zone Based Intrusion Detection System for MANETS. Since it uses hierarchical approach, it can be extended to deal with MANETS of any sizes. This fits distributed feature of MANET. As this model uses gateway nodes, it is supposed to make the adhoc network more efficient limiting the resources. The model represented in this paper represents the effective effort for the development of a sophisticated intrusion detection system for MANETS. As a future work, we plan to develop and improve the model and implement it using NS2 simulator to identify and eliminate the attacks which are vulnerable in MANETS due to open medium. 5. References
Data Collection
Global Response Agent
[1] Y. Zhang and W. Lee, “Intrusion Detection in Wireless Ad Hoc Networks,” the 6th Annual International Conf. on Mobile Computing andNetworking (ACM MobiCom’00), Boston, MA, Aug., 2000, pp. 275-283. [2] Fu, Yingfang; He, Jingsha; Li, Guorui,”A Distributed Intrusion Detection Scheme for Mobile Ad Hoc Networks”, IEEE Computer Software and Applications Conference, 2007.COMPSAC 2007 - Vol. 2. 31st Annual International Volume 2, Issue, 24-27 July 2007 Page(s):75 – 80 [3] Y. Hu and A. Perrig, “A Survey of Secure Wireless Ad Hoc routing”, IEEE Security and Privacy, IEEE press, pp: 28-39, 2004
Audit data
Velammal College of Engineering and Technology, Madurai
Page 137
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[4]Li Y.,Wei J., “Guidelines on Selecting Intrusion Detection Methods in MANET” Proc of ISECON 2004, v 21 [5]H.Deng, R. Xu, J. Li, F. Zhang, R. Levy, W. Lee, "Agentbased Cooperative Anomaly Detection for Wireless Ad Hoc Networks," in Proc. the IEEE Twelfth International Conference on Parallel and Distributed Systems (ICPDS'06), 2006 [6] Y. Huang and W. Lee, “A Cooperative Intrusion Detection System for Ad Hoc Networks”, Proc. 1st ACM Workshop Security of Ad Hoc and Sensor Networks,Virginia, ACM press, pp:135-147, 2003. [7] Anantvalee, Tiranuch and Wu, Jie, A survey on “Intrusion detection in mobile ad hoc networks”. In: Xiao, Y., Shen, X., Du, D.-Z. (Eds.), Wireless/Mobile Network Security, Springer. pp. 170-196. [8] Fu, Yingfang; He, Jingsha; Li, Guorui,”A Distributed Intrusion Detection Scheme for Mobile Ad Hoc Networks”, IEEE Computer Software and Applications Conference, 2007. COMPSAC 2007 - Vol. 2. 31st Annual International Volume 2, Issue, 24-27 July 2007 Page(s):75 – 80 [9] X. Lin and X. Zen, “Mobile Ad Hoc Network -Self-Organization Wireless Network Technology”,Publishing House of Electronics Industry, Beijing, China,2006.
Velammal College of Engineering and Technology, Madurai
Page 138
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Implementing High Performance Hybrid Search Using CELL Processor Mrs.Umarani Srikanth Assistant Professor, Computer Science and Engineering Department, S.A.Engineering College, Affiliated to Anna University Thiruverkadu Post, Chennai, Tamilnadu State, India [email protected]
Abstract— The slowing pace of commodity microprocessor performance improvements combined with ever-increasing chip power demands has become of utmost concern to computational scientists. As a result, the high performance computing community is examining alternative architectures that address the limitations of modern cache-based designs. Multicore processors are an architectural paradigm shift that promises a dramatic increase in performance. But they also bring of complexity in algorithmic design and software development. In this paper the challenges involved in designing the Hybrid Search which is the combination of both Breadth-First-Search algorithm (BFS) and Depth-First-Search (DFS) algorithm for the Cell/BE processor is discussed. The proposed methodology combines a high-level algorithmic design that captures the machineindependent aspects to guarantee portability with performance to future processors. The Hybrid Search helps searching terminology to be made simpler. There are a lot of searching terminologies in practice, which have its own advantages and disadvantages. The performance factor plays a key role in searching algorithms. The Cell broadband engine uses one PPE and eight SPEs thereby making the performance issues to be resolved and helps in faster computations. The Hybrid-Searching algorithm performs well on the multicore processor. This work combines the logics of both breadth first search and depth first search algorithms. It is proved that the parallelization improves the performance of the hybrid searching. Keywords— Multicore Processor, Parallel Computing, Cell Broadband Engine processor, Parallelization techniques, Depth First Search algorithm(DFS) and Breadth-First Search Algorithm(BFS).
I. INTRODUCTION Over the last decade, high-performance computing has ridden the wave of commodity technology, building clusters that could leverage the tremendous growth in processor performance fuelled by the commercial world. As this pace slows down, processor designers are facing complex problems when increasing gate density, reducing power consumption and designing efficient memory hierarchies. Traditionally, performance gains in commodity processors have come through higher clock frequencies, an exponential increase in number of devices integrated on a chip, and other architectural improvements. Power consumption is increasingly becoming the driving constraint in processor design. Processors are much more power limited rather than area limited. Current general purpose processors are optimized for single-threaded
Velammal College of Engineering and Technology, Madurai
workloads and spend a significant amount of resources to extract parallelism. Common techniques are out-of-order execution, register renaming, dynamic schedulers, branch prediction, and reorder buffers. Experiencing diminishing returns, processor designers are turning their attention to thread level, VLIW and SIMD parallelism [1]. Explicitly parallel techniques, where a user or a compiler expresses the available parallelism as a set of cooperating threads, offer a more efficient means of converting power into performance than techniques that speculatively discover the implicit and often limited instruction-level parallelism hidden in a single thread. Another important trend in computer architectures is the implementation of highly integrated chips. Several design avenues have been explored such as AMD Opteron, IBM Power5, SunNiagara, Intel Montecito, Itanium etc[2],[3],[4][5]. The Cell Broadband Engine (Cell/BE) processor, jointly developed by IBM, Sony, and Toshiba is a heterogeneous chip with nine cores capable of massive floating-point processing and is optimized for computationally intensive workloads and broadband rich media applications. The processing power of the single-precision Gflops/second has not passed unobserved.[30],[36]. To fully exploit the potential of multicore processors, we need a significant paradigm shift in the software development process. Unfortunately, for some classes of applications, this implies the radical redesign of algorithms. Together with the initial excitement of early evaluations [6], several concerns have emerged. More specifically, there is an interest in understanding what the complexity is of developing new applications and parallelizing compilers [9], whether there is a clear migration path for existing legacy software [7],[8], and what fraction of the peak performance can actually be achieved by real applications. Several recent works provide insight into these problems [10],[11],[12],[13]. In fact, to develop efficient multicore algorithms, one must understand in depth multiple algorithmic and architectural aspects. The list includes the following: • identifying the available dimensions of parallelism, • mapping parallel threads of activity to a potentially large pool of processing and functional units,
Page 139
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
• using simple processing cores with limited functionalities, • coping with the limited on-chip memory per core, • coping with the limited off-chip memory bandwidth The programming landscape of these advanced multicore processors is rather sophisticated. Many similarities appear with cluster computing: in both fields, we need to extract explicit parallelism, deal with communication, and take care of how threads of computation are mapped onto the physical machine [14],[15],[16]. There are also differences, mostly in the data orchestration between processing cores and main memory, which demand a fresh look at the problem and the design of new algorithms. 1.2 .Graph Exploration Algorithms By a graph search we mean some process that visits each node of a graph (arcs are merely the highways by which we perform the search). These algorithms known as breadth first search (BFS) and depth first search (DFS) are polar opposites. BFS is timid, it does not go ten yards from the den until it has explored everything within five yards. DFS is adventurous, as it goes as far as it can in any one direction before it tries a different direction. Many areas of science demand techniques for exploring large-scale data sets that are represented by graphs. Among graph search algorithms, Breadth-First Search (BFS) is probably the most common and the building block for a wide range of graph applications. For example, in the analysis of semantic graphs, the relationship between two vertices is expressed by the properties of the shortest path between them, given by a BFS search [17], [18], and [19]. BFS is also the basic building block for best-first search, uniform-cost search, greedy search, and A*, which are commonly used in motion planning for robotics [20], [21]. A good amount of literature deals with the design of BFS or dedicated hardware [22], [24]. But no studies have investigated how effectively the Cell/BE can be employed to perform BFS on large graphs and how it compares with other commodity processors and supercomputers. The Cell/BE, with its impressive amount of parallelism [36], [37] promises interesting speedups in the BFS exploration of the many graph topologies that can benefit from it. Searching large graphs poses difficult challenges, because the vast search space is combined with the lack of spatial and temporal locality in the data access pattern. In fact, few parallel algorithms outperform their best sequential implementations on clusters due to long memory latencies and high synchronization costs. These issues call for even more attention on multicores like the Cell/BE because of the memory hierarchy, which must be managed explicitly [38], [39], [40]. Breadth-first search is complete. Breadth-first Search (BFS) [25] is a fundamental graph theory algorithm that is extensively used to abstract various challenging computational problems. Due to the fine-grained irregular memory accesses, parallelization of BFS [41], [42] can exhibit limited performance on cache-based systems. If the goal node is at
Velammal College of Engineering and Technology, Madurai
some finite depth say d, BFS will eventually find it after expanding all shallower nodes. However the time taken to find out a solution is large. Whereas, DFS progresses by expanding the first child node of the search tree that appears and thus going deeper and deeper until a goal node is found, or until it hits a node that has no children. Then the search backtracks, returning to the most recent node it hasn't finished exploring. If the key to be found seems to reside at very high depths, the DFS algorithm may run into infinite looping whereby the searching process becomes incomplete. Breadthfirst search is ideal in situations where the answer is near the top of the tree and Depth-first search works well when the goal node is near the bottom of the tree. Depth-first search has much lower memory requirements. The combination of these two algorithms makes searching in an efficient manner in the sense that it takes the only the merits of both the algorithms and new parallelized algorithm is formulated called Hybrid Search and it is implemented which works very fast in searching the elements using the Cell Broadband Engine (Cell BE) Processor. As the modern processors moving more towards improving parallelization and multithreading, it has become impossible for performance gains in older compilers as technology advances. Any multicore architecture relies on improving parallelism than on improving single core performance. The main advantage of Hybrid Searching is that combining both of these traditional search strategies, one overcomes the disadvantage faced by the other. Experiments show an almost linear scaling over the number of used synergistic processing elements in the Cell/BE platform and compares favourably against other systems. II. MULTICORE PROCESSORS In computing, multi-core processor is a processing system composed of two or more independent cores. One can describe it as an integrated circuit to which two or more individual processors (cores) have been attached. The cores are typically integrated onto a single integrated circuit die [26] (known as a chip multiprocessor or CMP), or they may be integrated onto multiple dies in a single chip package. A many-core processor is one in which the number of cores is large enough that traditional multi-processor techniques are no longer efficient. With so many different designs it is nearly impossible to set any standard for cache coherence, interconnections, and layout. The greatest difficulty remains in teaching parallel programming techniques and in redesigning current applications to run optimally on a multicore system. Multicore processors are an important innovation in the microprocessor timeline. With skilled programmers capable of writing parallelized applications multicore efficiency could be increased dramatically. In years to come we will see much in the way of improvements to these systems. These improvements will provide faster programs and a better computing experience.
Page 140
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
As personal computers have become more prevalent and more applications have been designed for them, the end-user has seen the need for a faster, more capable system to keep up. Speedup has been achieved by increasing clock speeds and, more recently, adding multiple processing cores to the same chip. Although chip speed has increased exponentially over the years, that time is ending and manufacturers have shifted toward multicore processing. However, by increasing the number of cores on a single chip challenges arise with memory and cache coherence as well as communication between the cores. Coherence protocols and interconnection networks have resolved some issues, but until programmers learn to write parallel applications, the full benefit and efficiency of multicore processors will not be attained[28],[29]. The trend of increasing a processor’s speed to get a boost in performance is a way of the past. Multicore processors are the new direction manufacturers are focusing on. Using multiple cores on a single chip is advantageous in raw processing power, but nothing comes for free. With additional cores, power consumption and heat dissipation become a concern and must be simulated before lay-out to determine the best floor plan which distributes heat across the chip, while being careful not to form any hot spots. Distributed and shared caches on the chip must adhere to coherence protocols to make sure that when a core reads from memory it is reading the current piece of data and not a value that has been updated by a different core. 2.1.Moore’s Law One of the guiding principles of computer architecture is known as Moore’s Law. In 1965 Gordon Moore stated that the number of transistors on a chip wills roughly double each year and computer performance will double every 18 months [28]. The graph in Figure 1 plots many of the early microprocessors briefly discussed against the number of transistors per chip.
instructions are pre-fetched, split into sub-components and executed out-of-order. A major focus of computer architects is the branch instruction. Branch instructions are the equivalent of a fork in the road; the processor has to gather all necessary information before making a decision. In order to speed up this process, the processor predicts which path will be taken; if the wrong path is chosen the processor must throw out any data computed while taking the wrong path and backtrack to take the correct path. Often even when an incorrect branch is taken the effect is equivalent to having waited to take the correct path. Branches are also removed using loop unrolling and sophisticated neural network-based predictors are used to minimize the misprediction rate. Other techniques used for performance enhancement include register renaming, trace caches, reorder buffers, dynamic/software scheduling, and data value prediction 2.2 The Need for Multicore Speeding up processor frequency had run its course in the earlier part of this decade; computer architects needed a new approach to improve performance. Adding an additional processing core to the same chip would, in theory, result in twice the performance and dissipate less heat, though in practice the actual speed of each core is slower than the fastest single core processor. Multicore is not a new concept, as the idea has been used in embedded systems and for specialized applications for some time, but recently the technology has become mainstream with Intel and Advanced Micro Devices (AMD) introducing many commercially available multicore chips. In contrast to commercially available two and four core machines in 2008, some experts believe that by 2017 embedded processors could sport 4,096 cores, server CPUs might have 512 cores and desktop chips could use 128 cores[29]. This rate of growth is astounding considering that current desktop chips are on the cusp of using four cores and a single core has been used for the past 30 years[43]. 2.3 . MultiCore Framework The MultiCore Framework (MCF) is an API for programming a heterogeneous multicore chip oriented toward n-dimensional matrix computation. An MCF application consists of a manager program (that runs on the Cell’s PPE) and one or more workers (SPEs on the Cell). MCF’s most important features include the ability to:
To gain performance within a single core many techniques are used. Superscalar processors with the ability to issue multiple instructions concurrently are the standard. In these pipelines,
• define teams of worker processes, • assign tasks to a team of workers, • load and unload functions (called plugins) by worker programs, • synchronize between the manager and worker teams, • define the organization and distribution of n-dimensional data sets through data distribution objects, and • Move blocks of data via channels.
Velammal College of Engineering and Technology, Madurai
Page 141
Fig(1) Moore’s law
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
This last feature is perhaps the most important. It facilitates multi-buffered strip mining of data between a large memory (XDR memory on the Cell) and small worker memories (SPE local store on the Cell) in such a way as to insulate worker code from the details of the data organization in main memory. III.CELL BROADBAND ENGINE ARCHITECTURE The Cell architecture was patented by Ken Kuturagi (Sony) in 1999[30], [31], [32], [33], [34]. He differentiated between software and hardware cells. A software cell is a program with the associated data, and the hardware cell is an execution unit with the capability to execute a software cell. A cell processor consists of a so called PPE (PowerPC Processing Element) which acts as the main processor to distribute tasks (software cells), a MFC (Memory Flow Controller) which interfaces between the computing and memory units, many SPE’s (Synergistic Processing Elements) which are the hardware cells with their own memory. The Cell CPU is essentially a combination of a Power Processor with 8 small Vector Processors (cmp. Co-Processors). All these Units are connected via an EIB (Element Interconnect Bus) and communicate with peripheral devices or other CPU’s via the FlexIO Interface. Unfortunately, the Cell design incorporates hardware DRM features, but the positive aspects outbalance this easily. A single Cell, essentially a Network on Chip, offers up to 256 GFlop single precision floating point performance. A block-diagram of the processor is shown in figure 2.
transistors where 2/3 of them are dedicated to the SRAM(memory). The processor has no branch prediction or scheduling logic, and relies on the programmer/compiler to find parallelism in the code. As the PPE, it uses two independent pipelines and issues two instructions per cycle, one SIMD computation operation and one memory access operation. All instructions are processed strictly in-order and each instruction works with 128 Bit compound data items. 4 single precision floating point units and 4 integer units offer up to 32GOps each. The single precision floating point units are not IEEE754 compliant in terms of rounding and special values. The single precision units can also be used to compute double precision floating point numbers which are compliant to the IEEE754 standard. But their computation is rather slow (3-4GFlops). The memory layout of the SPE is also quite special, each SPE has its own 256kB RAM which is called Local Storage (LS). This SRAM storage can be accessed extremely fast in 128 bit lines. Additionally, each SPE has a large register file of 128, 128 bit registers which store all available data types. There is no cache, virtual memory support or coherency for the Local Storage, and the data can be moved with DMA from/to main memory via the EIB. The Memory Flow Controller (MFC) acts like a MMU in this context and provides virtual memory translations for main memory access. 3.3 .The Element Interconnect Bus
The SPE is essentially a full blown vector CPU with own RAM. Its ISA is not compatible to VMX and has a fixed length of 32 Bit. Current SPEs have about 21 million
The EIB is the central communication channel inside a Cell processor, it consists of four 128 bit wide concentric rings. The ring uses buffered point to point communication to transfer the data and is therewith scalable. It can move 96 bytes per cycle and is optimized for 1024 bit data blocks. Additional nodes (e.g. SPEs) can be added easily and increase only the maximal latency of the ring. Each device has a hardware guaranteed bandwidth of 1/numDevices to enhance the suitability for real time computing. The I/O Interconnect connects the Cell processor to other cell processors. It offers 12 unidirectional byte-lanes which are 96 wires. Each lane may transport up to 6.4GB/s, which make 76.8 GB accumulated bandwidth. 7 lanes are outgoing (44.8 GB/s) and 5 lanes incoming (32 GB/s). There are cache coherent, non coherent links and two cell processors can be connected glueless. The Memory Interface Controller (MIC) connects the EIB to the main DRAM memory, which is in this case Rambus XDR memory which offers a bandwidth of 25.2 GB/s. IV CELL PROGRAMMING The programming of the cell will be as special as the architecture. The big advantage is that there is no abstraction layer between an external ISA and the internal core (cmp. x86) [8]. But the strict RISC design moves the effort to generate optimal code up, to the programmer or compiler. And the general problems of parallel or streaming application development stay the same as for multi-core or multi processor machines. The SPEs are programmed in a direct
Velammal College of Engineering and Technology, Madurai
Page 142
Fig 2. Cell Processor Architecture 3.1. The Power Processing Element
The Power Processing Element (PPE) [36] offers the normal PowerPC (PPC) ISA. It is a dual threaded 64 bit power processor which includes VMX. Its architecture is very simple to guarantee high clock rates. Thus, it uses only in order execution with a deep super scalar 2-way pipeline with more than 20 stages. It offers a 2x32kB L1 split cache, a 512kB L2 cache and virtualization. Altogether the PPE can be seen as a simplified Power processor 3.2 .The Synergistic Processing Element
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
manner, as own autonomous processors with their 256kB Local Storage and 128 Registers. An Assembly language specification is available from IBM but higher level languages such as C/C++ are also supported for application development. The task distribution and allocation of SPEs is fully done in software. The operating system could use them as a shared resource and virtualize them for many tasks (each task sees their own SPEs, in the whole more than available). The PPE is programmed like a standard PPC970 and Linux is running as is (without SPE support, but a patch is available from IBM). The SPEs can be used in different scenarios. A job queue can be used to processes a fixed amount of independent jobs as fast as possible, the SPEs could also be used in a self multitasking manner, as if the cell were a multicore or multi CPU system. Stream processing (Pipelining) is also supported and especially very reasonable for media data (e.g. HDTV). The Local Storage could be used as a cache, but has to be managed by the software for this task. Additionally, MPI support is thinkable, where each SPE is a single node. All these different programming models are just ideas, the future will show which models will be used on the new Cell processors. V.SEQUENTIAL BREADTH-FIRST SEARCH ALGORITHM First the notation employed throughout the rest of this paper and sequential version of BFS is introduced. Then a simplified parallel algorithm as a collection of cooperating shared memory threads is described. Finally the level of detail with a second parallel algorithm that explicitly manages a hierarchy of working sets is described. Given a graph G(V ,E) and a root vertex r Є V , the BFS algorithm explores the edges of G to discover all the vertices reachable from r, and it produces a breadth-first tree rooted at r, containing all the vertices reachable from r. Vertices are visited in levels: when a vertex is visited at level l, it is also said to be at distance l from the root. A graph G = (V, E) is composed by a set of vertices V and a set of edges E. Define the size of a graph as the number of vertices |V|. Given a vertex vЄV, we indicate with Ev the set of neighboring vertices of v such that Ev = {wЄV:(v,w) Є E}, and with dv the vertex degree, i.e. the number of elements |Ev|. Denote D as the average degree of the vertices in a graph, D = vЄV |Ev|/|V|. Given a graph G(V,E) and a root vertex r ЄV, the BFS algorithm explores the edges of G to discover all the vertices reachable from r, and it produces a breadth-first tree, rooted at r, containing all the vertices reachable from r. Vertices are visited in levels: when a vertex is visited at level l it is also said to be at distance l from the root. The pseudo-code description of a sequential BFS algorithm is given below. Breadth-first search (BFS) is a graph search algorithm that begins at the root node and explores all the neighboring nodes. Then for each of those nearest nodes, it explores their unexplored neighbour nodes, and so on, until it finds the goal node[35].
Velammal College of Engineering and Technology, Madurai
Breadth First Search (G, S) Input : A graph G and a vertex. Output: Edges labelled as discovery and cross edges in the connected component. Create a Queue Q. ENQUEUE (Q, S) // Insert S into Q. While Q is not empty do for each vertex v in Q do for all edges e incident on v do if edge e is unexplored then let w be the other endpoint of e. if vertex w is unexpected then mark e as a discovery edge insert w into Q else mark e as a cross edge VI. SEQUENTIAL DEPTH-FIRST SEARCH ALGORITHM The algorithm starts at a specific vertex S in G, which becomes current vertex. Then algorithm traverse graph by any edge (u, v) incident to the current vertex u. If the edge (u, v) leads to an already visited vertex v, then backtrack to current vertex u. If, on other hand, edge (u, v) leads to an unvisited vertex v, then go to v and v becomes the current vertex. Proceed in this manner until reach to "deadend". At this point start back tracking. The process terminates when backtracking leads back to the start vertex. Edges leads to new vertex are called discovery or tree edges and edges lead to already visited are called back edges. Depth-first search (DFS) is an algorithm for traversing or searching a tree, tree structure, or graph. One starts at the root (selecting some node as the root in the graph case) and explores as far as possible along each branch before backtracking. DFS is an uninformed search that progresses by expanding the first child node of the search tree that appears and thus going deeper and deeper until a goal node is found, or until it hits a node that has no children. Then the search backtracks, returning to the most recent node it hasn't finished exploring. In a non-recursive implementation, all freshly expanded nodes are added to a stack for exploration. We will separate the DFS algorithm into four distinct procedures: Get Root, Go Forward, Go Backward and DFS. Note that once the algorithm is underway, there is a current node, C. If a node N has label (x,y) we write V(N) = x and L(N) = y. At any given time the algorithm is designed to go forward (that is further from the root). When it can’t go further it backs up to the prior node and tries to go further down another path. When the algorithm cannot go further and cannot back up then the algorithm looks for another root. When the algorithm looks for another root and fails to find one, it terminates. The DFS algorithm is as follows: Get Root {This procedure is only run when it is called from DFS} If there is a node labeled (0,0) with no antecedent
Page 143
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Then make this node C {it is current} and set V(C)←T; {T for time} Else If there is a node labeled (0,0) {and all such nodes have antecedents} Make it C and set V(C) ←T; Else Stop {algorithm is finished}; T←T+1. {end Get Root} Go Forward {This procedure is only run when it is called from DFS} Iterate2 ← True; Iterate←False While Iterate2 If an arc goes from C to a node labeled (0,0) Then make it C and set V(C)←T; Iterate←True; T←T+1; Else Iterate2 = False; {end Go Forward}. Go Backward {This procedure is only run when it is called from DFS} Iterate←False; L(C)←T If there is an antecedent node Then Find an antecedent node, N, with V(N) = V(C)-1; Set L(C)←T; C←N; {make N current} Else Iterate←False; T←T+1.{End Go Backward}.
for(i=1;i<spe count;i++) { Start:=spe[i]->thread[no of ele , key ,spe id] //invoke other SPEs to search in DFS } Found:=spu_read_out_mbox(); // Read status from outbound mailbox of SPEs. If(Found) Print “Key found ” Else Print “Key not found ” Terminate(spe[i]->thread) exit(0) end
DFS {This is the main procedure} Mark each node (0,0); Set T←0 Repeat Get Root {termination occurs in Get Root} Iterate←True While Iterate Go Forward Go Backward {end DFS}.
algorithm for the SPEs doing DFS create Tree(no of ele/7) dfs search tree(node* root , int key) if key found Found =1 Spu_write_out_mbox(spe id, found); // write the result to the ppe
algorithm for SPE doing BFS create Tree(no of ele); bfs search tree(node*root, int key) If key found // if key found , mark found as 1 Found=1 Spu_write_out_mbox(spe id, found); // write the result to the ppe Else Found =0 Spu_write_out_mbox(spe id, found);
VII. PARALLELIZED HYBRID SEARCH ALGORITHM Here it is proposed to combine two of the most popular searching algorithms used. The disadvantage of one algorithm is overcome by the other and thus making the searching algorithm to be developed in a better way. Parallel algorithm used on these searching techniques make them to run on the different cores thereby making effective utilization of the resources. And also the drawbacks of each of the algorithm are overcome by the other thereby reducing the time taken for the overall throughput. Algorithm for PPU No_of_ele:= read no of elements Spe_count:=read no of spes to be used Key:=key to be searched Start: spe[0]->thread(no of file , key ,spe id) //invoke first SPE to search in BFS
Velammal College of Engineering and Technology, Madurai
Fig3 pictorial representation of hybrid search
Page 144
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VIII. IMPLEMENTATION OF PARALLEL HYBRID SEARCH FOR CELL PROCESSOR To implement the hybrid search The IBM Full System simulator CELL SDK3.1 is used. The Cell Broadband Engine is a generalized simulator that can be configured to simulate a broad range of full-system configurations. It supports full functional simulation, including the PPE, SPEs, MFCs, PPE caches, bus, and memory controller. It can simulate and capture many levels of operational details on instruction execution, cache and memory subsystems, interrupt subsystems, communications, and other important system functions. It also supports some cycle-accurate simulation (performance or timing simulation). Where KEY is the element to be searched against the tree, spe_count is the number of SPEs to be utilized in the searching process, TREE SIZE is the total number of elements to be searched. The PPE determines the tree size for the SPEs performing DFS, depending on two factors namely, the input tree size and the number of SPEs to be utilized. PPE initiates the first SPE which is supposed to search its tree in BFS fashion by sending the key element to be searched. The entire tree is given as input to this SPE as this SPE will search the entire set of elements in BFS fashion. The PPE then initiates all other seven SPEs and it fixes the tree size for each SPE which is the total number of elements to be processed divided by the number of SPEs to be utilized. It then sends these trees to the respective SPE’s and it also sends the key element to be searched to each SPE. The number of SPEs invoked depends on the spe_count entered by the user. The processing of the SPEs starts now. Each SPE creates a binary tree with its own prescribed set of elements. The creation of nodes in the binary tree takes place in breadth first fashion. For the last SPE, the tree size differs slightly as the relation mentioned above doesn’t stand perfectly divisible at all times. So for the last SPE, the quotient obtained from the above relation is summed up with the remainder left out after the division is carried out, so that no element is left out within the tree_size prescribed. The nodes are created and added to the tree in level order fashion. Trees get constructed and stored in each of the SPEs local store. As the SPEs local store memory is very limited, this poses a limit on the number of elements that can be processed by this cell BE architecture. The actual searching process starts after the binary trees are created by all of the SPEs. The first SPE makes use of BFS strategy to search the elements of its tree. The rest of the SPEs search their own trees in Depth first search fashion. This is pictorially represented in the fig 3. Once a SPE finds the key, it immediately writes the key status to its outbound mailbox. The PPE keeps listening to the outbound mailbox of all of the SPEs continuously, The PPE
Velammal College of Engineering and Technology, Madurai
stalls until the outbound mailbox channel remains empty. When this channel gets filled up with the key status reported by any of the SPEs, the PPE immediately reads it and displays the result to the user. Once the key is reported, the PPE immediately terminates all SPE threads, to prevent them from searching further and all resources held are freed up. When the key element is not found by any of the SPEs, the SPEs performing DFS terminates with no signal of response. The responsibility of reporting that the key isn’t found in the tree prescribed is allotted to the SPE performing BFS. This is because BFS is always complete by which there is 100% guarantee of searching all the elements where DFS doesn’t stand complete, there is a possibility of this search running into infinite depths. So if the key to be searched is not found by the SPE performing BFS, then it writes the “KEY NOT FOUND” status to its outbound mailbox, which will be read by the PPE and reported to the user. The main advantage of combining both of these traditional search strategies is that, one overcomes the disadvantage faced by the other. If the key to be searched is found very deep in the tree, then it will be searched and reported by the SPE performing DFS. Where as if the key to be searched is found at very shallow depths, then the SPE performing BFS reports it first, by which time taken for searching the key element is greatly minimized IX.PERFORMANCE ANALYSIS The overall process and control flow of the parallelized and synchronized hybrid search is implemented by using cell sdk
.
3.1 simulator Performance in parallel algorithms is measured in terms of accuracy of the results and speed of execution of the algorithm. In this case the key element is searched parallely by all SPEs using both the graph searching algorithms. Since each SPE is working on its own set of nodes the result is obtained hardly within few seconds as soon as the process is started. The performance graph shown in figure 4 is meant for the key value 500, where the number of elements to be searched is 7000. Here we could see SPE7 doing the BFS, while all other SPEs doing DFS as shown in the figure. X. CONCLUSION Together with an unprecedented level of performance, multicore processors are also bringing an unprecedented level of complexity in terms of software development. A clear shift of paradigm from classical parallel computing, where parallelism is typically expressed in a single dimension, to the complex multidimensional parallelization space of multicore processors, where several levels of control and data parallelism must be exploited in order to gain the expected performance is seen. With this work, it is proved that for the specific case of the Hybrid Search graph exploration, it is possible to tame the algorithmic and software development process and achieve, at the same time, an impressive level of Performance. Thus the performance issues encountered using the existing searching algorithms are overcome by the hybrid search algorithm. The hybrid algorithms work fine and good
Page 145
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
with all the cores of the cell broadband engine. The accuracy of the results is also greatly improved. We can select an application for which this algorithm can be used to find optimal solution and enhance the performance by reducing the running time of the application. Also we can find paths from the root to the solution. We can also add weight corresponding to each edge.
[5]String Searching with Multicore Processors, Oreste Villa, Politecnico di Milano/Pacific Northwest National Laboratory
Daniele
Paolo Scarpazza and Fabrizio Petrini, IBM T.J. Watson Research Center,2009 Published by the IEEE Computer Society [6] S. Williams, J. Shalf, L. Oliker, S. Kamil, P. Husbands, and K. Yelick, “The Potential of the Cell Processor for Scientific Computing,” Proc. ACM Int’l Conf. Computing
Frontiers (CF ’06), May 2006. Fig.4 graph shows functioning of all SPEs . XI.REFERENCES [1] Optimizing a Fast Stream Cipher for VLIW, SIMD, and Superscalar Processors, Craig S.K. Clapp Proceedings of Fast Software Encryption Workshop 1997 [2] R. Kota and R. Oehler, “Horus: Large-Scale Symmetric Multiprocessing for Opteron Systems,” IEEE Micro, vol. 25, no. 2, pp. 30-40, Mar./Apr. 2005. [3]A Brief History of Microprocessors, Microelectronics Industrial Centre, Northumbria University, 2002, http://mic.unn.ac.uk/miclearning/modules/micros/ch1/micro01 hist.html [2] B. Brey, “The Intel Microprocessors”, Sixth Edition, Prentice Hall, 2003 [4]Video Transcript, “Excerpts from a Conversation with Gordon Moore: Moore’s Law”, Intel Corporation, 2005
Velammal College of Engineering and Technology, Madurai
[7]J.A. Kahle, M.N. Day, H.P. Hofstee, C.R. Johns, T.R. Maeurer, and D. Shippy, “Introduction to the Cell Multiprocessor,” IBM J. Research and Development, pp. 589604, July-Sept. 2005. [8]M. Ohara, H. Inoue, Y. Sohda, H. Komatsu, and T. Nakatani, “MPI Microtask for Programming the Cell Broadband Engine Processor,” IBM Systems J., vol. 45, no. 1, pp. 85-102, Jan. 2006. [9] K. Fatahalian, T.J. Knight, M. Houston, M. Erez, D.R. Horn, L. Leem, J.Y. Park, M. Ren, A. Aiken, W.J. Dally, and P. Hanrahan, “Sequoia: Programming the Memory Hierarchy,” Proc. Int’l Conf. High-Performance Computing, Networking, Storage and Analysis (SuperComputing ’06), Nov. 2006. [10] M.C. Smith, J.S. Vetter, and X. Liang, “Accelerating Scientific Applications with the SRC-6 Reconfigurable Computer: Methodologies and Analysis,” Proc. 19th IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’05), vol. 4, Apr. 2005.
Page 146
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[11] M. Guo, “Automatic Parallelization and Optimization for Irregular Scientific Applications,” Proc. 18th IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’04), Apr. 2004. [12] P. Bellens, J.M. Perez, R.M. Badia, and J. Labarta, “CellSs: A Programming Model for the Cell BE Architecture,” Proc. Int’l Conf. High-Performance Computing, Networking, Storage and Analysis (SuperComputing ’06), Nov. 2006. [13] D. Kunzman, G. Zheng, E. Bohm, and L.V. Kale`, “Charm++, Offload API, and the Cell Processor,” Proc. Workshop Programming Models for Ubiquitous Parallelism (PMUP ’06), Sept. 2006. [14]R. Drost, C. Forrest, B. Guenin, R. Ho, A. Krishnamoorty, D. Cohen, J. Cunningham, B. Tourancheau, A. Zingher, A. Chow, G. Lauterbach, and I. Sutherland, “Challenges in Building a Flat-Bandwidth Memory Hierarchy for a LargeScale Computer with Proximity Communication,” Proc. 13th IEEE Symp. High-Performance Interconnects (Hot Interconnects ’05), Aug. 2005. [15] J. Duato, “A New Theory of Deadlock-Free Adaptive Routing in Wormhole Networks,” IEEE Trans. Parallel and Distributed Systems, vol. 4, no. 12, pp. 1320-1331, Dec. 1993. [16]G. Bell, J. Gray, and A. Szalay, “Petascale Computational Systems,” Computer, vol. 39, no. 1, pp. 110-112, Jan. 2006 [17]M.E.J. Newman, “Detecting Community Structure in Networks,” European Physical J. B, vol. 38, pp. 321-330, May 2004. [18]M.E.J. Newman, “Fast Algorithm for Detecting Community Structure in Networks,” Physical Rev. E, vol. 69, no. 6, p. 066133, June 2004. [19] M.E.J. Newman and M. Girvan, “Finding and Evaluating Community Structure in Networks,” Physical Rev. E, vol. 69, no. 2, p. 026113, Feb. 2004 [20]A. Sud, E. Andersen, S. Curtis, M.C. Lin, and D. Manocha, “Real-Time Path Planning for Virtual Agents in Dynamic Environments,”Proc. IEEE Virtual Reality Conf. (VR ’07), Mar. 2007. [21]L. Zhang, Y.J. Kim, and D. Manocha, “A Simple Path Non-Existence Algorithm Using C-Obstacle Query,” Proc. Int’l Workshop Algorithmic Foundations of Robotics (WAFR ’06), July 2006. [22] M. deLorimier, N. Kapre, N. Mehta, D. Rizzo, I. Eslick, R. Rubin, T.E. Uribe, T.F.J. Knight, and A. DeHon, “GraphStep: A System Architecture for Sparse-Graph Algorithms,” Proc. 14th IEEE Symp. Field-Programmable Custom Computing Machines (FCCM), 2006. [23] D. Bader, V. Agarwal, and K. Madduri, “On the Design and Analysis of Irregular Algorithms on the Cell Processor: A Case Study on List Ranking,” Proc. 21st IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’07), Mar. 2007. [24] F. Blagojevic, A. Stamatakis, C. Antonopoulos, and D. Nikolopoulos, “RAxML-Cell: Parallel Phylogenetic Tree Inference on the Cell Broadband Engine,” Proc. 21st IEEE Int’l Parallel and Distributed Processing Symp. (IPDPS ’07), Mar. 2007.
[25] Efficient Breadth-First Search on the Cell/BE Processor Daniele Paolo Scarpazza, Oreste Villa, and Fabrizio Petrini, IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 19, NO. 10, OCTOBER 2008 [26] W. Knight, “Two Heads Are Better Than One”, IEEE Review, September 2005 [27]R.Merritt, “CPU Designers Debate Multi-core Future”, EETimes Online, February 2008, http://www.eetimes.com/showArticle.jhtml?articleID=206105 179 [28] P. Frost Gorder, “Multicore Processors for Science and Engineering”, IEEE CS, March/April 2007 [29] D.Geer, “Chip Makers Turn to Multicore Processors”, Computer, IEEE Computer Society, May 2005 [30]D. Pham et al, “The Design and Implementation of a First-Generation CELL Processor”, ISSCC [31]P.Hofstee and M. Day, “Hardware and Software Architecture for the CELL Processor”, CODES+ISSS ‟05, September 2005 [32]J. Kahle, “The Cell Processor Architecture”, MICRO-38 Keynote, 2005 [33]D.Stasiak et al, “Cell Processor Low-Power Design Methodology”, IEEE MICRO, 2005 [34]D.Pham et al, “Overview of the Architecture, Circuit Design, and Physical Implementation of a First-Generation Cell Processor”, IEEE Journal of Solid-State Circuits, Vol. 41, No. 1, January 2006 Ma [35]J.B. Orlin, K. Madduri, K. Subramani, and M. Williamson. A faster algorithm for the single source shortest path problem with few distinct positive lengths. Journal of Discrete Algorithms, 2009. [36]Programming the Cell Broadband Engine Architecture Examples and Best Practices: www.redbooks.ibm.com/redbooks/pdfs/sg247575.pdf [37]D.Bader,V.Agarwal and K.Madduri, “On the Design and Analysis of Irregular Algorithms on the Cell Processor: A Case Study on List Ranking,”Proc. 21st IEEE Int’l Parallel and Distributed Processing Symp(IPDPS ’07), Mar. 2007. [38]P.Bellens,J.M.Perez, R.M.Badia and J. Labarta, “CellSs: A Programming Model for the Cell BE Architecture,” Proc. Int’l Conf. High-Performance Computing, Networking, Storage and Analysis (SuperComputing ’06), Nov. 2006. [39]Parallel exact inference on the cell broadband engine processor, Proceedings of the 2008 ACM/IEEE
Velammal College of Engineering and Technology, Madurai
Page 147
conference on Supercomputing,2008 [40]K. Subramani and K. Madduri. A Randomized Queueless Algorithm for Breadth-First Search. Int'l. Journal of Computers and their Applications, 15(3):177{186, 2008.} [41]String Searching with Multicore Processors, Oreste Villa, Politecnico di Milano/Pacific Northwest National Laboratory
Daniele
Paolo Scarpazza and Fabrizio Petrini, IBM T.J. Watson Research Center,2009 Published by the IEEE Computer Society
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[42]Approximating Decision Trees with Multiway Branches, Venkatesan T. Chakaravarthy, Vinayaka Pandit and Sambuddha International Colloquium on Automata, Languages and Programming (ICALP) 2009 [43]B.Bouzas, R. Cooper, J. Greene, M. Pepe, and M. J. Prelle. Multicore framework: An API for programming heterogeneous multicore processors. In STMCS. Mercury Computer Systems, 2006.
Velammal College of Engineering and Technology, Madurai
Page 148
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Phase Based Shape Matching For Trademarks Retrieval B.Sathya BamaLecturer, M.AnithaStudent, Dr.S.RajuProfessor Electronics and Communication Engineering, Thiagarajar College of Engineering Thiruparankundram, Tamil Nadu, India-625015 [email protected],[email protected],[email protected] Abstract: This paper proposes a phase based method for shape matching for retrieving trademarks in digital databases. The proposed approach utilizes phase as a key to develop shape descriptors. The phase combined with magnitude information improves the recognition rate in terms of discrimination power. Experiments were carried out with 30 samples of trademarks and results were examined. Comparative study with existing algorithm shows that the proposed method is highly efficient in elimination of irrelevant matches. Keywords—Shape matching, Phase discrimination, Chord distribution, Radon transform, Chamfer distance transform.
II. INTRODUCTION SHAPE matching is a challenging problem particularly by huge proliferation of images on the Internet, mainly in computer databases containing thousands or millions of images. Applications of shape recognition can be found in Computer Aided Design/Computer Aided Manufacturing (CAD/CAM), virtual reality (VR), medicine, molecular biology, security, and entertainment [1]. Existing approaches can be divided into [1]: Statistical descriptors, like for example geometric 3D moments employed by [2] and the shape distribution [3]. Extensionbased descriptors, which are calculated from features sampled along certain directions from a position within the object [4]. Volume-based descriptors use the volumetric representation of a 3D object to extract features (examples are Shape histograms [5], Model Voxelization [6] and point set methods [7]). Descriptors using the surface geometry compute curvature measures the distribution of surface normal vectors [8]. Image based descriptors reduce the problem of 3D shape matching to an image similarity problem by comparing2D projections of the 3D objects [9]. Methods matching the topology of the two objects (for example Reeb graphs, where the topology of the 3D object is described by a graph structure [10]). Skeletons are intuitive object descriptions and can be obtained from a 3D object by applying a thinning algorithm on the voxelization of a solid object like in [11]. Despite such a long history, interest in the problem has been recently reignited [12] by the fantastic proliferation of
Velammal College of Engineering and Technology, Madurai
electronic imaging and digital color images Although many typical image database algorithms, such as finding two matching or cropped photographs, may rely on a variety of shading, texture, and color attributes, there are significant opportunities for the use of shape as a discriminator. The problem of trademark infringement is a particularly fitting example: most trademarks are high-contrast images from which a shape is readily acquired, and where shading and texture, play only a secondary role, or none at all. Fourier shape descriptors have been studied extensively; however the descriptor phases have been mostly neglected. The phase has either been ignored entirely, leading to highly ambiguous matches, or treated in a simple manner which is reasonable for complex, natural shapes, but which fails for simply-structured shapes common among trademarks [13]. We emphasize that the purpose of this paper is a theoretical analysis of phase in the context of shape matching. We will derive highly-robust phase tests, invariant to shift, rotation, scale, and orientation. However this leaves us with a class of tests, of which the optimum, found through detailed experiments with other existing algorithms, is the subject of ongoing research. The rest of this paper is organized as follows: Section 2 discusses about the background in the context of shape matching. Section 3 describes the implementation of the proposed work. Section 4 gives the experimental results. Finally, Section 5 concludes this paper. III. BACKGROUND IN SHAPE MATCHING A. Phase in Shape Descriptors Given a huge number of binary images, which are to be tested on the basis of shape, there are two, obvious criteria: a) Comparison Speed, and b) Compactness of Representation. Considering two-dimensional shapes bounded by simple, closed curve may be a much more compact description than the huge two-dimensional array of pixels making up the shape.
Page 149
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Indeed, chain codes been developed for such a purpose, in which x(t), y(t) clockwise trace the outline of a shape, where parameter t measures the boundary length from some arbitrary starting point t = 0. Since x(t) and y(t) are clearly periodic, Fourier approaches have long been used in compacting and analyzing shapes. In particular, let
z ( t ) = x ( t ) + iy ( t )
(1) represent the shape outline in the two-dimensional, complex domain where t is not a continuous variable. Using the FFT, z(n) is easily converted to the Fourier domain
f ( k ) = F { z ( n )}
(2) where the complex f(k) are known as the Fourier shape descriptors. The entire challenge, then, of shape matching is the interpretation and comparison of the Fourier descriptors fm(k) from shape m with those of some other shape n.
Chamfer distance transform: The similarity between two shapes can be measured using their chamfer distance [15].
U = {u }n andV = {v }m
i i =1 j j =1 .the Given the two point sets chamfer distance function is the mean of the distances between each point, ui€U and its closest point in V:
d cham (U , V ) =
1 n
∑ min
u i ∈U
|| u i − v j ||
v j ∈V
(4) The symmetric chamfer distance is obtained by adding
d cham (U , V )
.
IV. PHASE BASED FOURIER SHAPE MATCHING Figure 3.1 shows the complete Phase based shape matching for trademark-matching problem. Given some model shape outline z(n), we can consider four perturbations which do not affect the inherent shape: _
z ( n ) = Δ + e i φ rz (( n − δ ) mod N ) (5) where r is scaling,δ is orientation, φ is rotation and Δ is shift and N is the length of the boundary in pixels. Applying the Fourier transform, we find that _
− i 2 πkδ / N
iφ
f (k ) = e {Δ .( k = 0 ) + e rf ( k )} (6) Since z1(n), z2(n) are equivalent shapes, the goal is to determine how to find descriptors invariant to r, φ, Δ, and δ, and thus variant only to inherent changes in shape. Fig 2.1. Sensitivity of shape to the phase information
Because of the sensitivity of the phase (the complex angle component) of f (k) to image rotations and changes in the shape origin (where we define our “start” point for tracing the outline), often only the magnitude components are examined. However the pitfalls of ignoring phase are long established in image processing; indeed, each column in Figure 2.1 has constant magnitude characteristics but with varying phase. The situation becomes progressively worse for more complex shapes, which would include many trademark images, where wildly differing shapes may be naively matched by magnitude comparators. B.
Input trademark Fourier spectrum using FFT Magnitude Normalization Phase Discrimination
ℜ-Transform
The ℜ-Transform to represent a shape is based on the Radon transform. The Radon transform is the projection of the image intensity along a radial line oriented at a specific angle. We assume that the function f is the domain of a shape and its Radon transform is defined by:
TR ( ρ , θ ) =
−∞ −∞
∫ ∫ f ( x, y )δ ( x cosθ + y sin θ − ρ )dxdy
− ∞− ∞
(3)
Selecting common reference
Phase Comparison Fig 3.1. Flow Chart for proposed work
First, the term f(0) is the only one sensitive to Δ, so we ignore it. Next, we normalize to f(1), =
f (k ) =
_
f (k ) _
f (1)
= e − i 2 π ( k −1) δ / N
f (k )
(6)
f (1)
where δ(.) is the Dirac delta-function.
Velammal College of Engineering and Technology, Madurai
Page 150
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
which eliminates the effects of r and φ, leaving us with magnitude independent of the perturbations r, φ, Δ, δ leading to the usual temptation to ignore the phase terms. =
f (k )
| f ( k ) |=
Step 2: Separation of magnitude and phase is performed and magnitude normalization is done.
(7)
f (1 )
But Phase component contributes more for shape information and from Figure 2.1 with varying phase for complex shapes results may be naively matched by using magnitude comparators. If we consider two shapes, z1(n), z2(n) which are identical except for the phase in the qth Fourier coefficient (8) f ( q ) = a∠ φ f ( q ) = a∠ φ 1
1,
2
2
We have to work with phase differences other than those from adjacent Fourier coefficients. So we generalize to normalize out δ for all coefficients. Given a reference index j ≠1, we can create _ =
=
| f (k ) |=| f (k ) | _
=
=
=
∠ f (k ) = ∠ f (k ) − ∠ f ( j). = ∠f (k ) −
k− j j −1
(9)
k− j k −1 ∠f ( j) + ∠f (1) j −1 j −1
where j is selected to be a meaningful reference that is, =
| f ( j ) | | is not small. Thus we have constructed a δ-invariant phase sequence. Given two shapes f1(k), f2(k), the phase comparison thus involves selecting a good common reference j, typically by _
_
=
=
f maximizing | f1 ( j ). f 2 ( j ) | , computing
1
. f
2
and discriminating as _
_
=
∑| f
=
1
_
=
Fig 4.1. Input datasets
Step 3: Phase discrimination is done by three steps which involves selection of common reference j is by • Taking maximum value as Common reference. Step 4: Phase Comparison with the query image by minimizing the discrimination value representing relevant match. In Phase based Fourier Descriptors, a pixel with maximum value in the image is taken as common reference. Results obtained are compared with existing approaches of chord distribution and randon transform. A sample of database and query image with perturbations is shown in Figure 4.2.
_
=
( k ). f 2 ( k ) |( 2 − 2 cos( ∠ f 1 ( k ) − ∠ f 2 ( k ))) (10)
k
Obtaining minimum values in objective function for discrimination represent effective matching of two shapes. 4. EXPERIMENTAL RESULTS Fig 4.2.Database image and perturbed query image
In the experimental activity, a database with 30 samples of different trademarks from internet has been taken. Nine samples in the database are shown in Figure 4.1. Matlab software is used for simulation. The size of the trademark images in database are of 50 X 50 pixels. A.
Fig 4.3 shows the randon transformation of the two input trademarks shown above. It is clearly seen that when the input rotates there is an equivalent amount of difference in frequency domain.
Algorithm
Step 1: Input trademark is read and its Fourier spectrum is obtained using FFT.
Velammal College of Engineering and Technology, Madurai
Page 151
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
B.
Precision, Recall, Efficiency graph
Many different methods for measuring the performance of shape matching algorithm have been created. This paper uses the most common evaluation methods namely, Precision and Recall usually presented as a Precision vs Recall graph. We can always make recall value 1 just by retrieving all images. In a similar way precision value can be kept in a higher value by retrieving only few images or precision and recall should either be used together or the number of images retrieved should be specified. Thus we have plotted two graphs: Precision vs No of retrievals; Recall vs No of retrievals.
Fig 4.3.Randon transforms for above images
Fig 4.6(a).Precision Fig 4.4.Fourier transform. Magnitude normalization and phase comparison
Magnitude comparison shows similar results such that we cannot have complete discrimination. But when we consider the phase information there is a significant change while discriminating inputs based on phase comparison shown in fig 4.4. The discrimination values for three approaches for six sample trademarks have been compared in the Figure 4.5. From the Figure we obtain minimum phase discrimination value using Phase based shape matching compared to other methods. The matching accuracy using phase has been improved twofold times and outperforms existing method in eliminating irrelevant matches.
7
Phase- F F T
5
R and o n t r ansf o r m
4 3 2 1 0 l ogo1
l ogo2
l o go 3
l ogo4
l o g o5
Phase-FFT
0.8
Randon transform 0.6 0.4 0.2 0 6
12
18
24
30
N o . o f R e t r i e v a l s
Fig 4.6(b).Recall
Comparision
6
Recall 1
l o go 6
l ogo7
l o g o8
l og o 9
I np u t t r a d e m a r k
Fig 4.5.Comparision of proposed method with existing algorithm
Velammal College of Engineering and Technology, Madurai
As can be seen in the Fig 4.6 (a), (b) the precision and recall values by Phase based shape matching for different number of retrievals for all the database trademarks are greater than the existing method. Thus it automatically improves the retrieval efficiency and reduces the error rate. V. CONCLUSION A novel phase based shape descriptor for trademark-matching problem has been proposed in this paper. The phase-based discrimination has been implemented and proved to be effective in recognition of trademarks from large databases. The performance has been improved greatly, particularly in the elimination of highly irrelevant matches which appeared in earlier
Page 152
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
ACKNOWLEDGMENT The authors would like to thank the Management and Department of Electronics and Communication Engineering of Thiagarajar College of Engineering for their support and assistance to carry out this work. REFERENCES [23] B. Bustos, D. A. Keim, D. Saupe, T. Schreck, and D. V. Vranic. “Feature-based similarity search in 3d object databases”, ACM Computer Survey, 37(4):345–387, 2005. [24] M. Elad, A. Tal, and S. Ar. “Content based retrieval of vrml objects: an iterative and interactive approach. In Euro graphics Workshop on Multimedia”, . Springer, pages 107–118, New York, 2001. [25] R. Osada, T. Funkhouser, B. Chazelle, and D. Dobkin. Shape distributions. ACM Trans.Graph., 21(4):807–832, 2002. [26] D. Vranic and D. Saupe. “Description of 3d-shape using a complex function on the sphere”, In International conference on Multimedia and Expo., pgs 177–180. IEEE, 2002. [27] M. Ankerst, G. Kastenmller, H.-P. Kriegel, and T. Seidl. “3d shape histograms for similarity search and classification in spatial databases”. In 6th Int. Symp. on Advances in Spatial Databases, Springer, pages 201–226, London, UK, 1999. [28] D. Vranic and D. Saupe. “3d shape descriptor based on 3d fourier transforms”. In EURASIP Conference on Digital Signal Processing for for Multimedia Com munications and Services, pages 271–274. Comenius University, 2001 [29] J. W. H. Tangelder and R. C. Veltkamp.” Polyhedral model retrieval using weighted point sets”. In Shape Modeling International, pages 119–129, IEEE. Seoul, Korea, 2003. [30] E. Paquet and M. Rioux. Nefertiti: A tool for 3-dshape databases management. SAE transactions, 108:387–393, 1999. [31] C. M. Cyr and B. B. Kimia. A similarity based aspect graph approach to 3d object recognition. International Journal of Computer Vision, 57(1):5–22, 2004. [32] Y. Shinagawa, T. Kunii, and Y. Kergosien. Surface coding based on morse theory. Computer Graphics and Applications, IEEE, 11(5):66–78, September 1991. [33] H. Sundar, D. Silver, N. Gagvani, and S. Dickinson. Skeleton based shape matching and retrieval. Shape Modeling International, pages 130– 139, 12-15 May 2003. [34] T. Bui, G. Chen, “Invariant Fourier-wavelet descriptor for pattern recognition,” Pattern Recognition, 1999. [35] Paul Fieguth, Paul Bloore Adrian Domsa,” Phase-Based Methods For Fourier Shape Matching”, ICASSP,2004 [36] S. P. Smith and A. K. Jain, “Chord distribution for shape matching,” Computer Graphics and Image Processing, vol. 20, pp. 259-271, 1982. [37] Thayananthan, B. Stenger, P. H. S. Torr, and R. Cipolla, “Shape context and chamfer matching in cluttered scenes,” in Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. [38] S. Berretti, A. D. Bimbo, and P. Pala, “Retrieval by shape similarity with perceptual distance and effective indexing,” IEEE Trans. on Multimedia, vol. 2(4), pp. 225-239, 2000. [39] D. Guru and H. Nagendraswam, “Symbolic representation of twodimensional shapes,”Pattern Recognition Letters, vol. 28, pp. 144-155, 2007 [40] N. Alajlan, M. S. Kamel, and G. Freeman, “Multi-object image retrieval based on shape and topology,” Signal Processing: Image Communication, vol. 21, pp. 904-918, 2006. [41] N. Alajlan, I. E. Rube, M. S. Kamel, and G. Freeman, “Shape retrieval using triangle-area representation and dynamic space warping,” Pattern Recognition, vol. 40(7), pp. 1911-1920, 2007. [42] S. Han and S. Yang, “An invariant feature representation for shape retrieval,” in Proc. Sixth International Conference on Parallel and Distributed Computing, Applications and Technologies, 2005.
Velammal College of Engineering and Technology, Madurai
Page 153
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Enhanced Knowledge Base Representation Technique for Intelligent Storage and Efficient Retrieval Using Knowledge Based Markup Language A. Meenakshi1, V.Thirunageswaran2, M.G. Avenash3, Dept of CSE, KLN College of Information Technology, Sivagangai Dist., Tamil Nadu. [email protected],[email protected]
Abstract— Knowledge Engineering is an engineering discipline that involves integrating knowledge into computer systems in order to solve complex problems normally requiring a high level of human expertise. Knowledge Engineering is the technique applied by knowledge engineers to build intelligent systems: Expert Systems, Knowledge Based Systems, Knowledge based Decision Support Systems, Expert Database Systems etc. This system builds a Knowledge Base using Knowledge Base Markup Language (KBML) which is derived from XML architecture. All the Meta information is stored in a KBML file whereas the actual data may be available in any data source. Knowledge sharing is achieved, as this system can retrieve information from heterogeneous data source through this KBML file. This system also provides facilities to search/add the contents to and from the Knowledge Base though mobile phones and Windows Mobile phones without using GPRS. Keywords— Knowledge – Knowledge Base Markup Language – Meta information – data source.
I. INTRODUCTION IVING in a fast moving world, it is natural to expect things faster. Similarly in our quest for data search we need fast and efficient retrieving methodologies. With the evolution of new technology and new products in various domains, researchers are focusing more on exploring new techniques of storage, managing and retrieval of data and knowledge from a repository which has been acquired from various sources. Only having a repository of data or efficiently organizing the data cannot guide decision makers or management to make accurate decisions as humans do. The best approach is to integrate and manage the data in the form of knowledge. Retrieving of exact knowledge through online is increasing and it requires more amount of time to retrieve from different data sources and creating knowledge from information available. Knowledge searching through mobile phones does not exist through SMS. To defeat these anomalies we have built a Knowledge Base using Knowledge Base Markup Language (KBML) which is derived from XML architecture. This system also provides facilities to search/add the contents to and from the Knowledge Base though mobile phones and Windows Mobile phones without using GPRS. The aim of our project is to build a secured intelligent storage mechanism, which can store the information in the form of knowledge using the knowledge
L
Velammal College of Engineering and Technology, Madurai
based representation technique with the help of KBML tags, also the retrieval process is also simplified as it can just refer the KBML file which contains the Meta information about the Knowledge which is going to be stored in distributed data sources. 1.1. EXISTING SYSTEM AND ITS LIMITATIONS In any knowledge based system, the first step is to model the domain knowledge collected from experts, so as to enable effective retrieval of knowledge. Some of the existing knowledge based systems have employed the data structure termed as K-graphs, tree data structure for representing the expert knowledge in their domain of interest. The K-graphs was able to represent the expert knowledge about domains in problem-solving, minimizing the semantic loss that would occur if production rules were used instead. A tree data structure very much resembles a graph, and it can also be defined as an acyclic connected graph where each node has a set of zero or more children nodes, and at most one parent node. Existing Data Structure suffers from the following limitations: y Ambiguous design in storage of data. y Slower performance. y Insecure data storage. y Complexity in retrieving the appropriate data. y Increased Space and Time Complexity. 1 .2 CONVERSION OF META INFORMATION IN KBML This system performs the searching of knowledge across several data sources. The searching facilitates the user to select particular data sources from a list and to get the search result. During the search process, the existence of the search string is first looked up into a KBML file. This KBML file has the Meta information about all the knowledge in a particular data source, the KBML file which is a feature, enabled with derived XML architecture so that our goal of efficient retrieval of exact data is achieved. Information is stored in form of knowledge (KBML Tags). Security is a factor of major concern in any mode of data storage, so we take immense care of this stored information. Here, each knowledgebase is provided with a unique ID, which is stored in an encrypted format. When a search is
Page 154
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
made, the title is first picked and the corresponding IDs are used to navigate for retrieving the description. If the search keyword is not found in the KBML file, the control is passed to search the knowledge in the database for retrieval. End-users may wish to create knowledge when the particular search is not present. In such cases, users should create an account and enter the knowledge along with the title and description. At this stage, a KBML file is created for this newly created knowledge with all its constraints.
1.
ENHANCED KNOWLEDGE BASE REPRESENTATION Our vision of efficient retrieval comes true by the fast fetching of information stored in the form of knowledge. This system builds a Knowledge Base using Knowledge Base Markup Language (KBML) which is derived from XML architecture. All the Meta information is stored in a KBML file whereas the actual data may be available in any data source. Knowledge sharing is achieved, as this system can retrieve information from heterogeneous data source through this KBML file. This system also provides facilities to search/add the contents to and from the Knowledge Base though mobile phones and Windows Mobile phones without using GPRS. 2.1. DERIVATION OF KBML USING XML Extensible markup language (XML) is a fast emerging technical tool. It has several advantages to its name. The limelight feature of “user defined tags” makes this technology worth its salt. The sample structure of KBML is represented below. We use the structure of XML and derive our KBML file which is purely knowledge based. XML has a predefined structure whereas KBML has exclusively user defined structure, this is the major reason why we prefer KBML to XML. Apart from this KBML has a hierarchical structure which superimposes the characteristics of its predecessor XML. The exciting feature that KBML possesses is user can add his own information and by giving it with a unique id, which serves as a Meta information. This forms a part of the knowledge store that can be modified accordingly. Search of information which is our ultimate aim is also made easy.
Fig.1 Structure of KBML
2.2. KBML IN EDAPHOLOGY The KBML file is used to store the Meta information about the soil corresponding to each plant. The edaphology deals with the plant and its classification, here edaphology is a best suited case where we can impart the Knowledge base to represent the details of plants and its description as knowledge and this process is enhanced by the use of KBML file which contains the Meta information of the respective data stored in it. In order to avoid the complexities in using the tree structure we go for KBML. 2. STORAGE 3.1. ADDING THE KNOWLEDGE It is a key feature of the project that makes our database a dynamic one. Knowledge can be represented in any form such as text, documents etc. When a particular search is out of reach of the data sources, users may create their own knowledge to the data source. To facilitate the users, a wizard is designed which contains the simple steps to add the knowledge to the specified Knowledge base and the corresponding data source, provided the users are already registered. The system is also extended to add the user’s knowledge to the knowledge base through Mobile phones. Through the use of this feature our project seems to be “userfriendly”.
Fig. 2 Plant Data Source
3.2. UPDATING THE KNOWLEDGE Day in and day out we have new information. Any knowledge created is not constant, changes are mandatory so as to maintain the efficiency. Changes or updating are done when new techniques are introduced. To carry out these updates, users are allowed to update the existing knowledge that helps
Velammal College of Engineering and Technology, Madurai
Page 155
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
other searchers to gain knowledge. Also we can comment/ question on the particular knowledge of the user.
4.
EXPERIMENTAL RESULTS Users have the option of adding data in the already existing data base. While adding data about a plant the fields like name of the plant, taxonomy, geology should be filled in the appropriate text boxes as given in the form.
Fig. 3 Adding Plant Data Source Fig. 5Adding Description Source
3. RETRIEVAL 4.1 SEARCHING THE KNOWLEDGE it is the main part of the project, by which the data that a user needs is retrieved unto him. It is easier for us to search the relevant information which is available from the selected knowledge base through windows/mobile applications. When the user searches a particular knowledge, they are allowed to select knowledge base and data sources from which they need to search the knowledge from a varied variety of options. All the relevant results are displayed as a list and the knowledge is obtained by navigating to the specified data source.
6.1
CONCLUSION
Development of the Knowledge Base Markup Language (KBML) is used to improve the performance of the search process for a particular knowledge by selecting data sources. Each data source contains a set of knowledge descriptions to serve the requests. Using this, the search can be made efficient through both Windows/ Mobile applications. With world moving too fast all that a user expects is an efficient time constraint and we have strived to impose the same and have achieved a reasonable success. 6.2 FUTURE WORKS Collaborate through mobile: Depending upon the number of search results required, the user will get in their mobile phones. And through mobile system, new users can be connected to the knowledge base. Reception of updates performed (for e.g., daily – at the end of the day, the subscribed user will get the updates in their mobile phones) in the knowledge base. The boom of technology demands such enhancements which if done would be a great success.
REFERENCES [1] “Knowledge Engineering” from http://pages.cpsc.ucalgary.ca/~kremer/courses/CG/. [2] Jay Whittaker, Michelle Burns, John Van Beveren, "Understanding and measuring the effect of social capital on knowledge transfer within clusters of small-medium enterprises" in proceedings of the 16th Annual Conference of Small EnterpriseAssociation of Australia and New Zealand, 28 September – 1 October 2003. Fig. 4 Retrieval
Velammal College of Engineering and Technology, Madurai
Page 156
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[3] Brian D. Newman, Kurt W. Conrad,” A Framework for Characterizing Knowledge Management Methods, Practices, and Technologies" n Proc. of the Third Int. Conf. on Practical Aspects of Knowledge Management (PAKM2000) Basel, Switzerland, 30-31 Oct. 2000. [4] Venkatesh Mahadevan, Robin Braun and Zenon Chaczko,” A Holistic approach of Knowledge management Initiative for a Telecollaboration Business system” in Proceedings of International Business Research Conference, Victoria University of Technology, Level 12, 300 Flinders Street,Melbourne, November 15-16, 2004. [5] Kurt Conrad, Brian Newman,” A Framework for characterizing knowledge management methods, Practices & Technologies”, in proceedings of white paper published in TDAN.com, April 1, 2000. [6] Minsoo shin, “A framework for evaluating economics of knowledge management systems” in proceedings of Information and management, vol.42, no.1, pp: 179-196,2004. [7] Thi Bich Ngoc Pham, “Intra organizational knowledge Transfer process In vietnam’s Information technology companies”. [8] Gail-Joon Ahna, Badrinath Mohana, Seng-Phil Hongb, “Towards secure information sharing using role-based delegation”, Journal of Network and Computer Applications archive, vol 30 , pp: 42-59, 2007. [9]National Treatment Agency,” Confidentiality and information sharing”, September 2003. [10] Ravi Sandhu, Kumar Ranganathan, Xinwen Zhang, “Secure Information Sharing Enabled by Trusted Computing and PEI Models”, Ravi Sandhu and Kumar Ranganathan and Xinwen Zhang, in the Proceedings of ACM Symposium on Information, Computer, and Communication Security (ASIACCS), Taipei, Taiwan, pp: 2-12, 2006. [11] Ganesh Godavari, Edward Chow, “Secure Information Sharing Using Attribute Certificates and Role Based Access Control”, Security and Management, pp: 209-276,2005.
Velammal College of Engineering and Technology, Madurai
Page 157
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Semantic Web Based Personalization Of E-Learning Courseware Using Concept Maps And Clustering Anitha D Department of Computer Science, Lady Doak College, Madurai, Tamilnadu, India 625002 [email protected] Abstract: E-learning has now been regarded as a global learning system with its potential to reach the unreachable. Knowledge based e-learning tools are the key areas of inventions in the present educational scenario to improve the higher education. The recent research works have realized the web services to be invincible in providing the best services to e-learning. These web services are now made more intelligent as Semantic web and have proved themselves to be the best practices that can be adopted in e-learning system and the development of the learning materials. This paper proposes a systematic construction of e-learning courseware based on the semantic web in three phases, domain knowledge base representation, course delivery and evaluation of learners. The approach recommends the conceptual representation of domain knowledge and confirms the effectiveness of concept maps in accessing and delivering it. The paper finds the application of k means clustering and rule based inference to evaluate the learner performance and redesign the course delivery to the learners’ pace. Keywords -- E-learning, Semantic Web, Ontology, Concept Maps, Clustering
I. INTRODUCTION E-Learning, as a package of technology enhanced education tends to replace the standard practices of learning of board teaching and boring lectures. The concept of “Study anywhere, anytime and at anyone’s own pace” has made it a generally acceptable learning system. During the last decade, the introduction of information technology in the educational domain has resulted in a tremendous change in reshaping the method of delivery of academic knowledge. The World Wide Web has increased the intensity of technological play in the development of higher education especially through e-learning by making very fast access to relevant educational resource at any time and place. The web enabled e-learning system enables socially excluded communities to access higher education and connects different societies, communities, resources and learners. The very important resource of e-learning systems is the learning materials and so the development of knowledge based educational resources is the major area of research in the present educational scenario. The e-learning courseware is always expected to satisfy the knowledge needs of different types of learners and provide personalized delivery of academic knowledge. Recent research works in the field of e-learning shows that the main focus is on developing intelligent e-learning systems i.e. Web based systems that are more understandable by machines. The intelligent web based services are now rendered through Semantic web [1][2][5], which has been hottest topic of research now.
Velammal College of Engineering and Technology, Madurai
This paper proposes a method for intelligent semantic web based e-learning courseware development using concept maps and personalization based on clustering. The paper is organized as follows: Section II presents an overview of Semantic web and concept maps and their extensibility to build knowledge base systems , Section III introduces the semantic web model of e-learning courseware design using concept maps and its performance evaluation and Section IV pertains to the future issues regarding the work. II. SEMANTIC WEB AND CONCEPT MAPS - OVERVIEW A. Semantic Web Semantic web is a web architecture i.e. (common-sharedmeaning and machine-process able metadata), enabled by a set of suitable agents. The Semantic web is an extension of the World Wide Web, where content is expressed in a language with enough semantics so software tools can locate exchange and interpret information more efficiently. In the Semantic Web, machines will be able to navigate, integrate and process information in a meaningful way. The semantic retrieval of knowledge will focus searches in a much more concise manner. The Semantic web application will thrive through two key technologies, ontology (formal, explicit specification of a shared conceptualization) as well as reasoning and inference services. Ontologies are introduced to construct, share and reuse knowledge bases contextually which is the crucial application of the Semantic web [1][2][6][7]. They provide a common reference frame for communication. The ontologies are defined and developed using some specific languages among which OWL (Ontology Web Language) [11] being the latest and standardized one. OWL is a language for defining and instantiating Web ontologies. It makes the Web resources more readily accessible to automated processes by adding information about the resources that describe or provide Web content. It provides representation of classes and relationships with its constituent elements “class, instance, datatype property, object property, value and type of values”. OWL represents a domain clearly by defining classes and properties of those classes, defines instances and asserts properties about them, and reasons about these classes and instances to the degree permitted by the formal semantics of OWL. Data property is the relation between instances of classes and the data types of the individual members of the classes. Object property is the specification of relations between instances of two classes. It can assign value to a data property or
Page 158
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
represent an instance property with its syntactic structures. OWL is suggested for defining the domain knowledge as it preserves and enhances the conceptual modeling environment demanded by . They also serve as an educational tool by demonstrating the Semantic Web. the subject concepts with a diagrammatic representation thereby improving the learner’s ability to comprehend. B. Concept Maps Concept map is a flexible and informal form of knowledge Learner assessment could be made more efficient by representation. It is a tool that helps to organize knowledge for mapping the learners’ acquired knowledge into concepts meaningful learning. The concepts are represented in nodes on the and comparing them with the stored concept maps built by map, and the relation between the concepts is represented with experts of the domain. III. PROPOSED MODEL OF E-LEARNING COURSEWARE linking words. The linking words that are used to connect concepts provide meanings for the proposition and hence the DESIGN domain of knowledge. The fundamental components of concept The model is based on semantic construction of learning maps are: nodes (internal and external concepts), categories with materials and meaningful navigation through them, enabled which concepts are qualified (concept types), the graphic aspect with an ontological background. Semantic Web is a very of these concept types, the relations between nodes, the relation suitable platform for implementing an e-learning system, types that qualify those relations. because it provides all means for (e-learning): ontology Earlier studies on concept maps confirm their support for process development, ontology-based personalized and contextual oriented learning and their suitability for context management of course delivery, and rule based learner evaluation. The learning domain [3][7][8]. The reason is that the concept maps proposed e-learning system architecture for personalization reconstruct the knowledge context and span their applicability in of e-learning component is illustrated in fig. 1. representing the domain ontology. They also provide progressive exploration of knowledge and structured placement of documentary resources. Thus, concept maps support a systematic assessment which leads to improving teaching-learning strategies
Fig. 1 Personalized e-learning system architecture
Velammal College of Engineering and Technology, Madurai
Page 159
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
This architecture identifies a user profile database and ontology based courseware database. The implementation of Semantic web based e-learning necessitates the construction of ontology based design of the database. The Course delivery agent is responsible for providing learner interface and contextual course delivery. It is realized with the concept maps to represent the academic knowledge conceptually. Test Agent is responsible for launching intermediary assessment sequence and final testing of the learners. It updates the user profile with their performance grades. It also tests the correctness of the topic sequence in a course and reconstructs the courseware if necessary. The Learner evaluation agent with its rules to evaluate and classify the learners provides the competence level of the learner and tends to update the courseware design and the course delivery in tune with the interest of the learner. Courseware modeling agent is responsible for creating initial model of the course, analyzing the content and selecting the course structure. The modeling agent is accountable for structuring a course according to the current needs of learners. The management agent defines the ontology and constructs the knowledge base of the domain. A. Courseware Ontology construction The e-learning domain can be regarded as a knowledge management system since it deals with academic knowledge and requires its effective management [8]. Ontology construction is always perceived as knowledge acquisition in any knowledge management system. An important aspect of knowledge acquisition is the use of knowledge modeling, as a way of structuring of the acquired knowledge and storing knowledge for future use. A model is often a formalized visual representation of the acquired knowledge. Knowledge is perceived as a concept, and it is strongly argued and agreed that a knowledge unit could be represented in the notion of concepts, because knowledge itself is a conceptual understanding of the human brain. Knowledge models are structured representations of the identified concepts and relationship between them. Recent research works in ontology construction have proved the concept model of domain ontology could be efficiently built up by OWL (Web Ontology Language)[5][6][10][11]. OWL is the recent update of the semantic web world. The representation elements of OWL such as “class, instance, datatype property, object property, value and type of values” provide suitable classes and relationship for the identified domain concepts. Data property is the relation between instances of classes and the data types of the individual members of the classes while the object property is the specification of relations between instances of two classes. The rich syntactic structures of OWL represent a domain clearly and enable the semantic web agents to explore the domain in all its aspects. Protégé, an ontology editor developed by Standford University is chosen as a solution to enter the domain ontology due to its user friendliness and free access. In this regard, an initial attempt of ontology construction has been made for a small topic in the computer science subject, Data Structures: Stack Operation – PUSH. The following concepts and relationships are identified and listed out (Concepts are represented in bold font and the relationships are represented in italics style). The assumed ontology is realized with Protégé and
Velammal College of Engineering and Technology, Madurai
the following OWL classes are created with RDF support. Table I shows the preliminary Data Structures ontology of stack push operation and the relevant OWL-RDF format of the designed ontology. B. Personalized Course delivery using Concept Maps The very important feature that any learning environment demands is that the coherence between the learners’ needs and the learning content. An increase of coherence is achieved through the use of concept maps as educational aids that aim at interpreting, integrating and relating new ideas to existing ones – that are presented by the system upon the request of the learner [7]. Concept maps are proved to be the best implementation model of the semantic web based e-learning systems [4][7][8][9]. Earlier studies on concept maps confirm their support for process oriented learning [8]. The teachers will be constructing concept maps from the domain ontology and establishing relationship among concept maps. Fig. 2 shows the initial mapping of the ontology description to the concept maps. The class, instance, data type, value attributes of the OWL representation get mapped into concepts and the object property gets mapped into links of the concept maps. For example, “Stack has top pointer with initial value 0” identifies the concepts “Stack” as class, “Top pointer” as Datatype property, “Initial value” as Value property, “has toppointer” and “withinitialvalue” as object properties.. So, in the concept maps, “Stack, Toppointer, initial value of 0” are mapped as concepts and “has, with” is mapped as a link.
Fig. 2 OWL - Concept map Mapping
Page 160
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE I A PARTIAL LOOK OF THE PRELIMINARY STACK ONTOLOGY AND OWL REPRESENTATION
GG. HH.
JJ. Sample Ontology representation
Ontology Specification
KK.
(OWL – RDF)
II. LL.
WWW.
MM.
XXX.
NN. Stack ->has: Definition -> is: Data structure in which items are inserted and deleted at the same end(Definition type)
YYY. <owl:versionInfo>$Id: datastructures.owl $ </owl:versionInfo>
OO.
Stack ->Characteristic : Homogenouselements
ZZZ.
PP.
->Characteristic: Last-in-First-out
AAAA.
QQ.
BBBB.
RR. Homogenouselements -> characteristics : Elements ->has same : Datatype
CCCC. DDDD.
SS. Elements->are : operational data TT. Datatype -> types: integer, character, float,
EEEE.
<owl:Ontology rdf:about="">
<rdfs:comment>OWL ontology for Data Structures </rdfs:comment> <rdfs:seeAlso rdf:resource="resource location" />
<owl:Class rdf:ID="Stack"> <rdfs:label>Stack</rdfs:label>
string, records
FFFF.
UU. VV.
Stack -> has : MAXSIZE
GGGG. <rdfs:comment> Data structure in which items are inserted and deleted at the same end
WW.
Stack -> has : Top pointer
HHHH.
XX.
Toppointer -> denotes: ArrayLocation ->of: Insertion
IIII.
YY.
->of: Deletion
</rdfs:comment> </owl:Class>
JJJJ.
ZZ. Toppointer->initialvalue ->zero
KKKK.
AAA.
LLLL.
<owl:ObjecProperrty rdf:ID="hasMAXSIZE"> <rdfs:domain rdf:resource="#Stack" />
BBB.
Stack -> has: Operations
MMMM. </owl:ObjectProperty>
CCC.
Operations -> consists of : Push
NNNN.
DDD.
Operations -> consists of : Pop
OOOO.
EEE. FFF. GGG.
Stack: Push -> does : Insertion Push->has : PushArgument
Velammal College of Engineering and Technology, Madurai
<owl:DatatypeProperty rdf:ID="MAXSIZE">
PPPP.
<rdfs:domain rdf:resource="#Stack" />
QQQQ.
</owl:DatatypeProperty>
RRRR.
Page 161
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
HHH.
PushArgument -> consists of : Elements
III. JJJ.
Stack: Insertion -> adds : Elements
SSSS.
<owl:DatatypeProperty rdf:ID="toppointer">
TTTT.
<rdfs:domain rdf:resource="#Stack" />
UUUU.
</owl:DatatypeProperty>
KKK.
Insertion -> has: PushSteps
VVVV.
LLL.
PushSteps -> startwith : CheckOverflow
WWWW.<owl:FunctionalProperty rdf:ID="Operations">
MMM.
CheckOverflow ->check: overflow -> is : toppointer
XXXX.
<rdfs:domain rdf:resource="#Stack" />
YYYY.
<rdfs:range rdf:resource="#Operations" />
ZZZZ.
</owl:FunctionalProperty>
>= MAXSIZE
NNN.
CheckOverflow -> true : StackOverflowReport
OOO.
CheckOverflow->false: PushStep2
PPP.
PushStep2 -> increments: Top pointer
QQQ.
PushStep2 ->nextstep: PushStep3
RRR.
PushStep3 -> assigns : Pushargument ->
SSS.
AAAAA. BBBBB. <owl:Class rdf:about="#Operations"> CCCCC. arraylocationof :Toppointer
TTT. UUU.
<rdfs:comment>Push operation inserts element into
stack
DDDDD.
</rdfs:comment>
EEEEE. Stack : Push -> implemented by : Reference to C Program
FFFFF.
<rdfs:subClassOf>
code
VVV.
GGGGG.
<owl:type rdf:resource="#Push" />
HHHHH.
</rdfs:subClassOf>
IIIII.
</owl:Class>
JJJJJ. KKKKK.<owl:DatatypeProperty rdf:ID="push argument"> LLLLL.
<rdfs:domain rdf:resource="#Push" />
MMMMM.
</owl:DatatypeProperty>
NNNNN.<owl:Class rdf:about="#Operations"> OOOOO.
<rdfs:comment>Stack has two operations
PPPPP. </rdfs:comment> QQQQQ.
Velammal College of Engineering and Technology, Madurai
<rdfs:subClassOf>
RRRRR.
<owl:type rdf:resource="#Push" />
SSSSS.
</rdfs:subClassOf>
TTTTT.
</owl:Class>
Page 162
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig 3. Partial Concept map for Stack Push operation
Fig. 3 shows a partial concept map for the Push operation in the stack. The learners are listed with the concepts and their relationships to other concepts by means of concept maps and they are shown the parts of the concept map in which they are interested in. They are made to extend their understandability by showing indirectly related concepts on demand. For e.g. a concept map which deals with the basic sorting algorithms may also have a reference to the current applications of sorting along with their efficiency considerations and also may have a reference to an URL or a textbook reference related to Sorting. If the learner is interested in moving to the advanced level of learning, he/she can move through the advanced links of concept maps. Those advanced links of concept maps may be programmed to display only upon the demands of the learner i.e. based on learners’ interest. Similarly, the students are given option to specify the keywords of their concern to extract specific parts of the concept maps. They are also encouraged to build their own concept maps for any topic of their interest and explore the relationships between unrelated concepts. The developed concept maps are verified and added into the content knowledge base thus
transforming the learner capability into a reusable knowledge entry.. The web-enabled courseware design also demands interactive presentations and intermediary assessments of the learner on completion of each topic. C. Rule based learner evaluation and dynamic course delivery using k means clustering The learner assessment is based on his/her response for intermediate test questions in each topic of the learning content. The Test Agent of the proposed method prepares the multi choice objective type test questionnaire with a knowledge evaluation pattern which constitutes three levels of knowledge evaluation: 1. Knowledge (Knowing and defining the concepts (i.e.) what is what?), 2. Understanding (Identifying, classifying, interpreting concepts) and 3. Application (Analyzing and applying existing concepts, arriving and justifying at new concepts).
TABLE II THREE LEVEL CLUSTERING FOR A LEARNER PERFORMANCE
UUUUU.
BBBBBB.
Level
Knowledge
VVVVV. Clusters WWWWW. ry High
Ve
XXXXX. Hig h
YYYYY. Avera ge
ZZZZZ. Lo w
AAAAAA. ry low
CCCCCC. pic numbers
To
DDDDDD.
EEEEEE.
FFFFFF.
GGGGGG.
Velammal College of Engineering and Technology, Madurai
Ve
Page 163
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
HHHHHH. ing
Understand
IIIIII.
JJJJJJ.
KKKKKK.
LLLLLL.
MMMMMM.
NNNNNN.
Application
OOOOOO.
PPPPPP.
QQQQQQ.
RRRRRR.
SSSSSS.
At the end of each topic, the learner is assessed with such a set of questionnaire comprising a minimum of 20 multi choice objective questions supporting all the levels of the given knowledge evaluation pattern. If he/she answers correctly one mark is given, otherwise zero mark is given for each question in a topic. The learner’s response of correct answers is recorded as fraction of 1 in each of these three levels. Based on the fraction of the correct answers scored in each of these 3 levels, the performance of the learner in all the topics is clustered into five major clusters – Very high, High, Average, Low, Very low. The clustering is applied to all the three levels of knowledge evaluation pattern. Table II shows the initial clustering of the learner performance. The learners will definitely fall into any one of these categories for each topic and so the k means clustering algorithm is used to determine the performance rate of a learner where k = 5 indicating the five clusters. Given an initial set of k means m1(1),…,m5(1), which may be specified randomly or by some heuristic, and in this case, the means chosen initially are 0.9, 0.7, 0.5, 0.3, 0.1 with an heuristic approach based on the nature of the clusters respectively. The algorithm proceeds by alternating between two steps: 1. 2.
Assignment step: Assign each observation to the cluster with the closest mean Si(t) = { xj : || xj – mi(t) || ≤ ||xj – mi*(t) || for all i* = 1,…,k } (1) Update step: Calculate the new means to be the centroid of the observations in the cluster.
(2) The algorithm is deemed to have converged when the assignments no longer change. The clusters thus formed are then able to represent a single learner’s behavior in each topic. The three level clusters make the personalization of e-learning more powerful by automatically classifying the learning capabilities of learner in each topic. A set of rules are framed with these cluster information and the rule based inference enables perfect assessment of learner capabilities. Table III shows some of the rules and the inferences made. The entries in the table denote the following: K – Knowledge, U- Understanding, A- Application, VH – Very High, H- High, M – Average, L – Low, VL – Very Low.
Assignment Update
Velammal College of Engineering and Technology, Madurai
Page 164
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE III FEW RULE BASED INFERENCE ON CLUSTERS AND THE ASSESSMENT OF LEARNER BEHAVIOR
TTTTTT.Rule
WWWWWW. ∩ U x L)
Ti in{ K x VH
ZZZZZZ.Ti in{ U x VH ∩ A x L)
UUUUUU.
Inference
XXXXXX. Good in knowledge but low in Understanding level AAAAAAA. very high
Understanding–
BBBBBBB.
Application - Low
VVVVVV.
Assessment
YYYYYY.Needs more presentations and classification works
DDDDDDD. Needs more realtime examples and problem solving exercises
CCCCCCC.
EEEEEEE. ∩ U x VL)
IIIIIII.
Ti in{ K x VL
Ti in{ K x M ∩ U x M)
FFFFFFF. low
Knowledge – Very
GGGGGGG. Very low
Understanding–
LLLLLLL.
i => j
MMMMMMM. performance
Ti affects Tj
RRRRRRR.
i => j
JJJJJJJ. and KKKKKKK. U x L)
Tj in{ K x M ∩
OOOOOOO. U x L)
Ti in{ K x M ∩
PPPPPPP.
and
QQQQQQQ. U x H)
Tj in{ K x H ∩
UUUUUUU.
T1..n in (K x H)
SSSSSSS. Inspite of the average performance in Tj , good performance in Tj
VVVVVVV. High performance in all the topics for the knowledge level
Velammal College of Engineering and Technology, Madurai
HHHHHHH. A very dull learner. Needs more attention. Need More visualized presentations in Ti. Need more worked examples.
NNNNNNN. Learner understanding of Ti must be improved to improve his performance in Tj
TTTTTTT. Swapping of the two related topics suggested for the learner
WWWWWWW.
A consistent learner
Page 165
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
These rules are to be well defined with the collaborative effort of teachers and experts of the domain. Knowledge sharing is needed between them to predict a learner behavior based on the rules. An experimental study was made with 50 students of computer science undergraduate discipline by conducting intermediary tests in 10 basic topics of the data structure subject. Their test results are analyzed and clustered with the k means clustering technique and grouped into different categories. Fig. 4 shows a sample cluster distribution of a student performance in the subject. The cluster distribution statistics of the entire set of 50 students for the knowledge level is shown in fig. 5. From the graph shown, it is made clear that the students could easily make up with definition and introduction of new concepts but it is not easy to show the same response when it comes to the next higher level of applications based on the concepts.
Cluster Distribution 4 No. of 3 topics in 2 the cluster 1 0
VH
H
M
L
VL
Knowledge
4
2
3
1
0
Understanding
3
3
2
1
1
Application
1
2
2
2
3
Identified level of learner Fig. 3 Cluster distributions for a student test data
N o . o f s tu d e n ts o f th e s a m e c l u s te r
Distribution of topics in the student clusters 20 Very high
15
High Medium
10
Low 5
Very low
0 1
2
3
4
5
6
7
8
9 10
Topic No. Fig. 5 Students performance organized in clusters
After finding out the clusters, the rule based approach of assessing them was applied and the results are appreciable fulfilling the intended purpose of this proposed method. For
Velammal College of Engineering and Technology, Madurai
example, topics 4 and 7 which are designed to be higher level of application based on the defined concepts have low performance of students. More detailed presentations and solved problems are needed on such topics. This intended method is sure to open a way to find out the learner capabilities based on the test results and improve their performance. IV. IMPLICATIONS & FUTURE ISSUES The present e-learning scenario demands an extensive research to derive innovative practices in analyzing, constructing and restructuring the e-learning management systems. The system becomes more reliable when it is designed to be intelligent and so the proposed solution emphasizes the conceptual representation and access of the domain knowledge. The learner evaluation, being an inevitable and inseparable part of any courseware design, has to be performed deliberately. Before the implementation of the proposed solution, the rule based inference must be carefully designed by the experts of learning domain. The current study of the e-learning courseware design has marked the courseware content sequencing to play an important role in affecting learner behavior. A case study of content sequencing [6] has shown the incorrect content sequencing to be a reason for the decrease in performance of learners during the intermediary assessment of the misplaced topics. Recent research works in learner evaluation demands the inclusion of the time spent by the learner in each topic, the participation of the learner in discussion forums, the number of suggestions or clarifications raised by the learner and the assignment performance. The proposed solution may be extended to study the impact of lesson sequencing on the learners’ performance and the learner evaluation including the identified factors and the future issues pertain to the specified findings. V. CONCLUSION Effective technology is very important for the sustainable development of education. This paper finds a machine intelligent approach for constructing e-learning courseware. It identifies three phases related to the courseware design and discusses the need of conceptual representation to simulate the human brain. It also explains the technical implementation of the three phases along with a case study on the third phase of learner evaluation. The knowledge gained during learning is the most important factor for measuring the quality of education and it can be measured with the intermediary tests. This paper may be an opening for the realization of web based intelligent educational tools for effective utilization of academic knowledge. REFERENCES [1] Alsultanny Y, E-Learning System Overview based on Semantic Web , The Electronic Journal of e-Learning, Volume 4 Issue 2, 2006, pp 111 - 118, Available online: www.ejel.org [2] Fayed Ghaleb, Sameh Daoud, Ahmad Hasna, Jihad M. ALJa’am, Samir A. El-Seoud, and Hosam El-Sofany, E-Learning Model Based On Semantic Web Technology, International Journal of Computing & Information Sciences Vol. 4, No. 2, Pages 63-71, August 2006 . [3] Hsinchun Chen, Ann Lally, Byron Marshall, Yiwen Zhang, Convergence of Knowledge Management and E-Learning: the GetSmart
Page 166
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Experience, Paper presented in the 3rd ACM/IEEE-CS joint conference on Digital libraries, 2003, ISBN:0-7695-1939-3. [4] F. P. Rokou et al.,Modeling web-based educational systems: process design teaching model, Educational Technology and Society, Vol. 7, pp. 4250, 2004 [5] Ungkyu Park, Rafael A. Calvo, Automatic Concept Map Scoring Framework Using the Semantic WebTechnologies, Proceedings of Eighth IEEE International Conference on Advanced Learning Technologies, DOI 10.1109/ICALT.2008.125, July 1-- July 5, 2008,, Spain. [6] Chih-Ming Chen1 and Chi-Jui Peng2, Personalized E-learning System based on Ontology-based Concept MapGeneration Scheme, Seventh IEEE International Conference on Advanced Learning Technologies (ICALT 2007), 0-7695-2916-X/07, July 18-20, 2007, Japan. [7] Thanasis Giouvanakis,Garyfallos Fragidis,Eyaggelos Kehris,Haido Samaras, Exploiting Concept Mapping in a Semantic Web Environment, Proceedings of the Fifth IEEE International Conference on Advanced Learning Technologies (ICALT’05), 0-7695-2338-2/0, July 5-8, 2005, Taiwan. [8] Marco Pedroni, E-learning and Knowledge Management: Context Structuration, Paper Presented at the Informing Science and IT Education Joint Conference, 2007, 22-25 June 2007, Ljubljana, Slovenia. [9] Juha Puustjärvi, Leena Puustjärvi, Using Semantic Web Technologies in Visualizing Medicinal Vocabularies, IEEE 8th International Conference on Computer and Information Technology Workshops, 978-0-7695-3242-4/08, DOI 10.1109/CIT.2008. [10] Srimathi H, Knowledge Representation in Personalized ELearning, Academic open Internet Journal, Volume 23, ISSN 1311-4360 , 2008. [11] W3C Recommendation 10 , OWL Web Ontology Language Guide”, February 2004.
Velammal College of Engineering and Technology, Madurai
Page 167
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Adaptive visible watermarking and copy protection of reverted multimedia data S.T.Veena#1, Dr.K.Muneeswaran*2 # PG Graduate, * Professor & Head Computer Science and Engineering Department, Mepco Schlenk Engineering College Sivakasi, Tamilnadu, India 1 [email protected] [email protected]
2
Abstract —Visible watermark is traditionally used to authenticate the ownership of the media unambiguously and is irreversible. However it is desirable in some applications like medical imagery, remote sensing, to have a visible watermark, to authenticate the media, at the same time hide the important details from being accessible to outsiders. It is also known that even small alternation to the media in such applications may be disaster. So a reversible visible watermarking scheme which combats copyright piracy, at the same time losslessly recovers the originals back by authorized person is proposed. It is not possible in such case to have original watermark at the retrieval end. And thus it has to be blind. In case of applications like artwork preserving and press, the same can be used against piracy of the retrieved media by embedding a buyer authentication code in the retrieved media with negligible changes. In case of copy protection violation, it can be used to track to the buyer involved in piracy. The digital media may be gray scale image, color image or a video. The watermark is a binary image. The watermarking application uses visible embedding method for embedding watermark and reversible data embedding (invisible) method for recovering the digital media (original and watermark) losslessly and blindly. The visible watermark is embedded in the spatial domain such that it is not so obtrusive of the beneath host signal, while clearly authenticates the signal. So an adaptive (host dependent) scaling factor found by exploiting the HVS properties is used for embedding the watermark in user-specified region of the host media. In order to achieve reversibility, a reconstruction/ recovery packet, is constructed, compressed and embedded in the nonwatermarked region of the host media using a reversible data embedding technique. Buyer authentication is done by generating the hash value and embedding it reversibly in the watermark area. Retrieval and verifying against buyer authentication code helps to track to that buyer. Keywords — Blind watermarking, Human visual system HVS, reversible data embedding, visible watermarking, data compression, Authentication code, Copyright piracy.
I. INTRODUCTION Conventional visible image watermarking schemes [1],[2] impose strict irreversibility. However
Velammal College of Engineering and Technology, Madurai
certain applications like medical imagery, prepress industry, image archival systems, precious artworks, remotely sensed images require them to be removable and lossless. Various reversible
schemes have been developed using the techniques of modulo arithmetic, the circular interpretation of the bijective transform, wavelets and sorting or difference expansion [3]-[5].They are unsuitable to visible watermarking, because the embedding distortion inflicted by a visible watermark is often far greater. Achieving lossless recovery of the original host signal from a visibly watermarked signal is an acute challenge. The prior works in this area [6]-[8], need the original watermark for original image recovery. Moreover, all the existing methods do not consider Human Visual System (HVS) characteristics in the visible watermark embedding process. As a result, they are less visually satisfactory and more intrusive. In paper [9], the binary watermark is embedded spatially in the host image in user specified region using adaptive scaling factor. This factor utilizes HVS characteristics of the host image and varies with region. Reversibility is achieved by hiding a recovery packet in the non-watermarked region of the host using LSB hiding of pixels after applying simple integer transform (Reversible contrast mapping). The recovery packet is the difference between the watermarked image and the approximated image in the watermarked region. To reduce the payload data to be hidden, the packet is encoded and compressed. In this paper we extend the ideology of the prior work to color images, video and authenticate it to buyer in case of piracy. The rest of the paper is organized as follows. In Section II, the embedding process of the proposed reversible visible watermarking algorithm for color image is presented in detail. Section III briefly introduces the watermark removal and lossless media recovery. Section IV discuss the process of identifying the buyer to piracy. Section V provides the experimental results for evaluating
Page 168
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the performance of the algorithm. The conclusion is given in Section VI.
II. WATERMARKING PROCESS Here the embedding process involves separation of the luminance part of the media and data embedding. This is done because humans are more sensitive to the black-and-white information than to other color components (more cones than rods in our vision system) A. Separating Luminance Part To separate the luminance (luma) part of the image which is a grayscale version of the image, it is converted from RGB color space to YCbCr color space. The Red Green and Blue components are converted to its respective Luminance/luma (Y/Y’), blue difference chroma and red difference chroma components. Once luminance part is separated, the watermark is translucently overlaid in the spatial positions defined by the user in the ROI of it. The watermarked luma component is again combined with Cb and Cr components. Then the image from this color space is converted back to RGB color space. B. Data Embedding The data embedding process involves two main procedures: visible watermarking and invisible reversible data hiding. The details of the processes are given in the following section. 1) Visible Watermarking: Watermarking is either additive or multiplicative. This is a multiplicative watermarking scheme where the change is done in the host where the corresponding overlaying watermark has a pixel value zero. This is chosen because the binary mark is chosen such that white represents 1 and black zero and the watermark is in white backdrop. Let N = no of 8*8 blocks of Y component of image I S ⊂ {1, 2, . . . , N} be the set of block ID numbers corresponding to the 8×8 blocks in ROI. ⎧ ⎣α n × Y n (i , j )⎦, if W n ' = 0 ⎫ Y nw (i , j ) = ⎨ ⎬ if W n ' = 1 ⎭ ⎩ Y n (i , j ), 1 ≤ i, j ≤ 8 and n ∈ S (1)
…
and
Ynw(i, j) = Yn (i, j) , 1 ≤ i , j ≤ 8 and n∈{1,2 ,K, N }− S …(2) where Ynw and Yn is the (i,j) th spatial pixel value in the nth 8 × 8 block of Yth component of Image I and watermarked Y component of Iw. ⎣•⎦ is the represents the mathematical floor function, In of the watermarked and αn is the adaptive scaling factor for the nth block of I. Wn′ is the n′th position in
Velammal College of Engineering and Technology, Madurai
the watermark. It is given n′ for watermark since the absolute spatial positions of both watermark and image are different. 2) Finding Adaptive Scaling Factor: According to HVS, the human eye is sensitive to mid luminance area while less sensitive to textured region. Exploiting this fact, instead of having a constant scaling factor for embedding, it can be chosen in such a way that the scaling factor depends on the underlying host image. Higher the texture, lower the scaling factor and more of luminance, high the scaling factor is selected. It is that discrete cosine transform actively captures this information through its ac and dc components. The dc coefficients of the DCT domain are best represented by Gaussian distribution [11] and ac coefficients by [12]. The adaptive scaling factor is therefore determined as per equation 3 and values are adjusted to be in the range [0.84 0.89]. This is done because values >0.89 tend to move towards invisibility and < 0.84 tend to be obtrusive.
⎧− [Y (1 , 1) − μ ]2 ⎫ ˆ exp⎨ n 2 ⎬ +ν n 1 ≤ n ≤ N 2 σ 2 ⎩ ⎭ 2πσ 1
αn = μ=
1 N
σ2 =
vˆn =
L(3)
N
∑Y (1 , 1) n
is the mean of
dc coef
and
n =1
1 N
2
N
∑[Yn (1, 1) − μ] is
the variance
n =1
vn − min n (vn ) max n (vn ) − min n (vn )
is the normalised log arithm of vn
vn = ln(vn ), 1 ≤ n ≤ N vn =
1 [Yn (i , j ) − η n ]2 × 63 (i , j )≠(1,1)
ηn =
1 × Yn (i, j ), 63 (i , j )≠(1,1)
∑
∑
1≤ i, j ≤ 8 and 1 ≤ n ≤ N
2) .
3) Facilitating Recovery:The watermark and the Image must be recovered blindly at the detection/extraction end. To facilitate this, the difference information must be provided along with the watermarked image. This will have a great payload. Hence an approximate image is generated of the original. This is done by first finding the approximate adaptive scaling factor. The factor is determined by generating an approximate original image using prediction of pixels 4) Pixel Prediction : This prediction method uses neighboring non-watermarked pixels to predict the original
Page 169
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
value of watermark pixel. It is an iterative process. It involves choosing overlapping windows of size 3×3. The window is chosen such that the pixel to be recovered (watermarked pixel) is at the center. The new pixel intensity of that center pixel of the window is the average of non-watermarked pixels and recovered watermarked pixels. The window is moved across the image in raster scan and the recovered pixels are used in subsequent windows for further recovery in that window. 5) Approximate Scaling factor and Image Generation : With this approximated version of the image P, find the adaptive scaling factor αˆ using equation 3 as described above with the original image Y. Then use this to get an approximate version of the original Ya by reversing the embedding process. ⎫ ⎪⎪ ⎬ ⎪ ⎪⎭
⎧ ⎢ P w (i, j ) ⎥ ⎪⎪ ⎢ n ⎥, if Wn (i, j ) = 0 Yna (i, j ) = ⎨ ⎢⎣ αˆ n ⎥⎦ ⎪ w ⎪⎩Pn (i, j ), if Wn (i, j ) = 1 1 ≤ i, j ≤ 8 and n ∈ S
L(4)
L (5 )
6) Construction of Recovery Packet : A recovery packet is constructed as a difference between the watermarked image and the approximated image of the watermarked area as in (4)
D = (Y − Ya)ROI
C. Invisible reversible data embedding The invisible hiding process uses reverse contrast mapping [10], an Integer transform, to embed the data in the LSB using the transforms. It inverts exactly back even when LSB is lost/ changed, except in case of odd pixels, which is also taken care. This algorithm even though has low embedding capacity (yet satisfies requirement) than other reversible data embedding mechanisms has less arithmetic complexity and is fast. The transform processes are given as Forward transform is … x′ = 2x − y; y′ = 2 y − x (8) Reverse transform is
Yna (i , j ) = Pnw (i , j ),
1 ≤ i , j ≤ 8 and n ∈ {1, 2, K , N } − S
After encoding, the packet is compressed using simple runlength encoding. For the purpose of security, the encoded and compressed (De) is exclusively ored (XOR) with a pseudo random numbers of a secret seed before inserted into the non-ROI region.
… (6)
7) Compression and encoding: Once the recovery packet D is constructed, the values are compressed to reduce the payload. The payload is then invisibly hidden in the nonwatermarked region. To facilitate compression, the values of D are encoded in L bit 2’s complement binary form where L-1 is maximum number of bits needed to represent D as per equation (7). For example we have the difference in the range -4 to 2 pixel intensity. The maximum of the absolute value of pixel intensity is 4. Therefore maximum bits needed for coding this is 3 bits (100). Assign 4 bits(L) one for sign (ie) MSB or Lth bit is 1 if the number is negative else 0. Thus in example -4 is assigned as 1100 and +4 as 0100. This helps in recovering signed number after decompression. Similarly zeroes are specially coded as those belonging to watermarked pixel (i , j) ∈ Γ and those that do not (i , j) ∈ Γ ′. The former zero is represented with MSB 1 while the latter with 0. ⎧⎪ ( L−1) − D(i, j ) where D (i, j )≤0 and (i, j ) ∈ Γ⎫⎪ = 2 ⎬ … (7) De(i, j ) ⎨⎪D (i, j ) ⎪⎭ otherwise ⎩ (i, j ) ∈ Γ∪Γ′
Velammal College of Engineering and Technology, Madurai
2 ⎤ ⎡ 2 (9) 1 ⎤ ⎡1 x =… y = ⎢ x′ + y′⎥ ⎢ 3 x′ + 3 y′⎥ 3 ⎥ ⎢ ⎥ ⎢3 where (x,y), (x′,y′) are the chosen pixel values in the range [0-255] to prevent overflow and underflow conditions. III.
RECOVERY AND DECODING PROCESS The process of recovery of original and watermark extraction to prove ownership is the reverse of the operations involved in embedding. The process is blind and the keys used in reconstruction are the top-left position of ROI, size of watermark, size of payload, auxiliary information of compression, size of encoding length L, secret key – seed of pseudo-random generator. First, transform the RGB to YCbCr color space. Second, extract the invisibly embedded bits from LSB by using inverse transform. Third, using secret key (seed) reproduce the compressed payload back (recovery packet). Next, decompress the payload to recover the encoded reconstruction packet. The packet is decoded by taking the 2’s complement of L-bit number. ⎧2L−1 − D(i, j) if MSB is 1⎫ D′(i, j) = ⎨ ⎬ where (i, j) ∈ Γ ∪ Γ′ otherwise ⎩D(i, j) ⎭ … (10)
⎧1 if MSB is 0 and D′ (i, j)=0⎫ W(i, j) = ⎨ ⎬ where (i, j) ∈Γ∪Γ′ ⎩0 otherwise ⎭ …
(11) Apply pixel prediction method to the watermarked image after payload is removed, as done previously to estimate a scaling factor and use it to get an approximated version of original image Y. Now to recover original pixel values in
Page 170
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
watermarked region add recovery packet to the approximated version to get back the originals. Transform back to RGB color space from YCbCr color space. In case of video, the above process is repeated for each frame with its correspondent keys to recover back the original video. IV. BUYER AUTHENTICATION The copy protection does not stop with recovery of original and selling the media to an authorized buyer. But it is in some applications necessary to keep track of piracy even after sales. The following ideology helps to track down to the buyer in case of piracy. A. Embedding hash code To accomplish this, first a hash value is generated for say identifier code. Save the length of the digest, it serves as the secret key. Secondly, the hash value is reversibly hidden into the region of the media (decided by the seller) using the RCM integer transform, a LSB technique, with the modification that instead of using all pixels in the selected region, embedding is done only on the odd pixels belonging to the transform domain (Dc).This helps in reverting both the pixel and the digest. The reverted digest can be used to track to the buyer responsible for the piracy. Let the identifier code be concatenation of sales id, date and some other details. Apply hash can be SHA-1 or md5 as application demands and this can be kept as a secret. The hash bits are then embedded into a region of the recovered media by utilizing modified RCM technique. B. Hash checking Regenerate the hash code with the identifier code stored in database with its key. Check it with the one retrieved. If it matches then the corresponding buyer is involved in copyright infringement else proceed with other identifier code until it can be tracked to the buyer. IMPLEMENTATION AND RESULTS The proposed method has been implemented and tested on various media (images (webber database) & video) different formats & sizes with different watermarks in different ROIs. The perceptually inspired metrics PSNR, WPSNR, Structural Similarity (SSIM) are used to access the scheme. Figures 3 and 4 shows the watermarking results for image and video sequences. Table I and II show the performance of watermarking of various images. PSNR-1 is watermarked with original except ROI. PSNR-2 is for watermarked with original. PSNR-3 is for original with retrieved.WPSNR-1 is weighted PSNR using CSF function for watermarked with original.WPSNR-2 is that with retrieved and original.WPSNR-3 is between original and retrieved in ROI.SSIM-1,-2 are structural similarity between original and watermarked & original and retrieved.SSIM-3 is between original and retrieved in ROI In embedding the visible watermark in different ROIs, it was noted that there is a difference in watermark visibility. The watermark is more visible in smooth image
areas. To access the visual quality of the visible watermark, perceptually insipired merics WPSNR and SSIM are used and results are tabulated in Table I. From PSNR-2 it can be seen that image with large smooth regions (LENA & F-16) has the lowest PSNR. This is because of the large the embedding distortion caused by visible watermarking. The same can be concluded from WPSNR-3 & SSIM-3. Taking into consideration the size of the hidden payload, which is invisibly embedded into the non-ROI region, it can be inferred from PSNR-1 that the image with smaller payload has high PSNR values. Difference (in F-16 & splash ) may occurs due to encoding length (L=4 or 3). The prediction is more correct for smooth regions. In such case the approximate image generated is very close to the original image. This can be inferred from PSNR-3,SSIM2,WPSNR-2. The value of SSIM that is closer to 1 infers that the image quality is retained due to HVS exploitation. Table-II shows the result of embedding watermark mark-3 of 4 sizes at upper-left of the image lena. It can be clearly seen that larger the size of watermark implies larger payload. As a result the PSNR of Non-ROI decreases (PSNR -1). Also relatively low PSNR-2 demonstrates the distortion in embedding in smooth region. In addition it was noted that watermark complexity also added to payload. Thus of all watermarks of same size considered mark-3 was the one with larger payload. In summary the watermarking process depends on embedding area in host, watermark size and complexity and encoding length. Table III shows the results of embedding buyer digest which suggests that the digest length contributes to PSNR value. Higher the length, lower the PSNR value. However it was authenticable in all cases. Images of test:
V.
Velammal College of Engineering and Technology, Madurai
Fig.1a Lena
Fig.1b Peppers
Fig.1c Splash
Fig.1d Airplane
Watermarks used:
Fig.2a mark-1
Fig.2b mark-2
Fig.2c mark-3
Fig.2d mark-4
Page 171
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Watermarked Images:
Fig.3a
Fig.3b
Fig.3c
Fig 4(a) 11thframeFig. 4(b) 61thframe Fig. 4(c) 200thframe Fig.4(d) 400thframe
Fig.3d
Watermarked Video Frames:
Sample Buyer id : abcd Its Length : 107 Its Digest : 10011111000100000101000010100000110110000111111011011101110 100001111000111100110001101111011000110110111001
Fig.5a. Watermarked lena with mark-3
Fig.5c.Digest embedded lena
Fig.5b.Retrieved lena
Fig.5d.Extracted Lena
TABLE I: PERFORMANCE EVALUATION: WATERMARKING VARIOUS IMAGES WITH MARK-3 OF SIZE 128 × 128 AT VARIOUS ROI.
Images
Lena F-16 Peppers Splash
Payload (in bits) 25372 25701 26143 25188 24515 25207 23473 25099 24832 24828 25246 24322
Position
PSNR-1
PSNR-2
PSNR-3
(0,0) (160,160) (352,272) (0,0) (160,160) (352,272) (0,0) (160,160) (352,272) (0,0) (160,160) (352,272)
55.3490 42.8205 42.5805 66.8455 45.8896 45.7781 85.7012 50.6652 53.2093 65.2598 44.0934 44.0734
37.8900 39.0298 38.0277 39.1492 39.2714 39.5457 44.1700 40.1900 44.3139 42.2465 40.3683 38.6686
59.8601 65.0116 65.0359 65.2507 65.1645 64.9491 62.8522 65.0808 65.0948 65.2507 64.5645 64.6513
WPSNR -1 30.2225 28.7887 28.7301 25.0714 24.9415 25.1664 37.6478 46.8588 50.4606 34.2956 33.6492 33.4906
WPSNR -2 49.4495 49.2789 52.6813 62.4066 58.8382 57.0555 Inf 54.9316 Inf Inf Inf 47.0625
WPSNR -3 33.2196 49.0391 32.6273 29.2657 46.0943 47.1568 53.9220 47.7934 52.3948 54.0074 50.4374 47.0983
SSIM-1
SSIM-2
0.9631 0.9735 0.9634 0.9681 0.9738 0.9799 0.9857 0.9807 0.9852 0.9770 0.9705 0.9643
0.9995 0.9998 0.9996 0.9996 0.9996 0.9998 0.9992 0.9997 0.9995 0.9996 0.9995 0.9995
SSIM3 0.9642 0.9873 0.9788 0.9686 0.9829 0.9895 0.9858 0.9850 0.9892 0.9777 0.9835 0.9773
TABLE II : PERFORMANCE EVALUATION : WATERMARKING LENA IMAGE WITH MARK-3 OF VARIOUS SIZE AT (0,0)
Watermark size 32 × 32 64 × 64 128 × 128 256 × 256
Payload (in bits) 2831 9319 25372 52700
PSNR-1
81.0282 81.0093 55.3490 40.2160
PSNR-2
PSNR-3
49.2997 43.5129 37.8900 33.2593
66.1034 61.2387 59.8601 61.1828
WPSNR1 58.0810 43.7084 30.2225 30.2871
WPSNR2 Inf 55.6605 49.4495 Inf
WPSNR3 58.9150 49.4605 33.2196 31.0454
SSIM-1
SSIM-2
SSIM-3
0.9973 0.9877 0.9631 0.9277
0.9998 0.9994 0.9995 0.9992
0.9973 0.9877 0.9642 0.9366
TABLE III : PERFORMANCE EVALUATION : BUYER AUTHENTICATION OF RETRIEVED IMAGE AT ROI STARTING FROM (0,0).
Velammal College of Engineering and Technology, Madurai
Page 172
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Image
Authenticable
Digest Length
Lena F-16 Peppers Splash
yes yes yes yes
110 107 117 120
VI. CONCLUSION A reversible visible watermarking technique is presented in this paper which can be applied to any visual media. The paper proposes a method which considers HVS of the host to watermark, to achieve the desired features of visible watermarking. And at same time the image quality is retained as shown by our results. Further the application being blind makes it suitable for extraction of original at any place & time As a key dependent method the scheme allows only authentic users with correct key to retrieve the original. The buyer authentication when embraced in the process, though a simple technique, helps a lot in not only preventing piracy but also to find the buyer involved in piracy.
PSNR of retrieved image 59.8601 65.2507 62.8522 65.2507
PSNR of authentic image 59.9861 65.2663 63.4003 66.0321
[12] F.Muller, “Distribution shape of two dimensional DCT coefficients of natural Images” Electronics letters vol 29, no 22, pp 1935-1936, Oct 1993.
REFERENCES [1] G. Braudaway, K. A. Magerlein, and F. Mintzer, “Protecting publicly available images with a visible image watermark,” Proc. SPIE, International Conference on Electronic Imaging, vol. 2659, pp. 126–133, Feb. 1–2, 1996. [2] M. S. Kankanhalli, Rajmohan, and K. R. Ramakrishnan, “Adaptive visible watermarking of images,” in Proc. IEEE Int. Conf. Multimedia Comput. Syst., vol. 1. Florence, SC, Jul. 1999, pp. 568–573. [3] A. M. Alattar, “A Novel Difference Expansion Transform for Reversible Data Embedding,” IEEE Trans . on Information Forensics and Security,, vol. 3, no. 3, pp. 456–465, Sep. 2008. [4] C. De Vleeschouwer, J.-F. Delaigle, and B. Macq, “Circular interpretation of bijective transformations in lossless watermarking for media asset management,” IEEE Trans. Multimedia, vol. 5, no. 1, pp. 97–105, Mar. 2003. [5] Kamstra, L., Heijmans, H.J.A.M. “Reversible data embedding into images using wavelet techniques and sorting,” IEEE Trans. Image Process., vol. 14, no. 12, pp. 2082–2090, Dec. 2005. [6] S. C. Pei and Y. C. Zeng, “A novel image recovery algorithm for visible watermarked images,” IEEE Trans. Inf. Forens. Security, vol. 1, no. 4,pp. 543–550, Dec. 2006. [7] Y. Yang, X. Sun, H. Yang, and C.-T. Li, “Removable visible image watermarking algorithm in the discrete cosine transform domain,” J.Electron. Imaging, vol. 17, no. 3, pp. 033008-1–033008-11 Jul.–Sep. 2008. [8] Y. J. Hu and B. Jeon, “Reversible visible watermarking and lossless recovery of original images,” IEEE Trans. Circuits Syst. Video Technol., vol. 16, no. 11, pp. 1423–1429, Nov. 2006. [9] Ying Yang, Xingming Sun, Hengfu Yang, Chang-Tsun Li, and Rong Xiao “A Contrast-Sensitive Reversible Visible Image Watermarking Technique,” IEEE Transactions On Circuits And Systems For Video Technology, vol. 19, no. 5, pp. 656-677, May 2009 [10] D. Coltuc and J. M. Chassery, “Very fast watermarking by reversible contrast mapping,” IEEE Signal Process. Lett., vol. 14, no. 4, pp. 255–258, Apr. 2007. [11] Randall C. Reiningek and Jerry D. Gibson “Distribution of two dimensional DCT coefficients for Images” IEEE transactions on communications, vol. com -31, no 6 Jun 1983.
Velammal College of Engineering and Technology, Madurai
Page 173
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Web Personalization System for evolving user profiles in Dynamic Web Sites based on Web Usage Mining Techniques and Agent Technology G.Karthik)#1 B.E.,M.E.,(Ph.D), R.Vivekanandam*2 M.Sc., M.S.,(Ph.D), P.Rupa Ezhil Arasi#3 M.Sc., M.Phil.,M.E 1 Department of CSE,Vinayaka Mission Kirupananda Variyar Engineering College Salem, Tamilnadu,India 2 Department of Applied Science, Muthayammal Engineering College Raipuram, Namakkal Dt, TamilNadu,India 3 Department of CSE, Vinayaka Mission Kirupananda Variyar Engineering College Salem, Tamilnadu,India 1
[email protected] 2 [email protected] 3 [email protected] Abstract - Customer Relationship Management use data from within and outside an organization to allow an understanding of its customers on an individual basis or on a group basis such as by forming customer profiles. These profiles can be discovered using web usage mining techniques and can be later personalized. Web personalization system captures and models behavior and profiles of users interacting with the web sites. Web personalization is the process of customizing a web site to the needs of specific users taking advantage of the knowledge acquired from the analysis of user’s navigational behavior in correlation with the information collected namely
I.
structure, content and user profile data. The output of the web usage mining process can be improved by using agent technology. Agent technology provides a dynamic personalized guidance to the visitors of the web. This paper describes the design of a web usage mining architecture for web personalization system implemented using a multi-agent platform. Keywords-Customer Relationship Management, Web usage Mining, Web personalization, Agent technology, Customer profiles
INTRODUCTION
The output of the web usage mining process can be improved by using agent technology. This creates a friendly relationship between the web and its users. In this paper, we are interested in the web usagemining domain, which is described as the process of customizing the content and the structure of the web sites in order to provide users with the information they are interested in. Various personalization schemes have been suggested in the literature. Lelizia [9] is perhaps the first system, which takes into account the user’s navigation through the first system. Yan et al. [2] propose a methodology for the automatic classification of web users according to their access patterns, using cluster analysis on the web logs. In [3], Joachims et al. describe Web Watcher, and similarly the Personal Web Watcher in [4], an intelligent agent system that provides navigation hints to the user, on the basis of a knowledge of the user’s interests, the location and relevance of the many items in the site, and the way in which other users interacted with the collection in the past. We can finally conclude that most of the existing works try to classify a user i) while she is browsing the web site or ii)
Velammal College of Engineering and Technology, Madurai
using registration information. The main criticism stands in the fact that in some applications it is not possible to perform an “on line” classification if the number of visited pages is not sufficiently great. By the way using the registration form alone may result inaccurate if the interests of user change over time. The novelty of our approach is that we perform clustering of the user sessions extracted from the web logs to partition the users into several homogeneous groups with similar activities and then extract user profiles from each cluster as a set of relevant URLs. This procedure is repeated in subsequent new periods of web logging then the Previously discovered user profiles are tracked, and their evolution pattern is categorized. In order to eliminate most sessions from further analysis and to focus the mining on truly new sessions FM model is used. The FM model is suitable for real time matching of session to pre-generated cluster and it offers scalable models of clusters. The rest of this paper is organized as follows: In Section 2, we present an overview of web usage mining. In Section 3, we describe our approach to profile discovery using web
Page 174
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
usage mining. In Section 4, we discuss our approach for tracking evolving user profiles. In Section 5, we present the overview of web personalization. In Section 6 we present the personalization solution. In Section 7, we present the main agents of our architecture along with their roles and interactions among them. In Section 8, we present our results. Finally in Section 9, we present our conclusions. II. AN OVERVIEW OF WEB USAGE MINING Web usage mining is a tool for web personalization, since it captures and models behaviors and profiles of users interacting with a web site. These models can be used by the personalization systems to better understand the behavioral characteristics of visitors, the content and structure of the web sites and provide dynamic recommendation to visitors. Web mining is performed in several stages [5], [6] to achieve its goals 1) Collection of Web data such as activities / click streams recorded in web server logs. 2) Preprocessing of Web data such as filtering, crawlers requests, requests to graphics and identifying user sessions. 3) Analysis of web data, also known as web usage mining [7], to discover interesting usage patterns of profiles and 4) Interpretation / evaluation of the discovered profiles 5) Tracking the evolution of discovered profiles.Web usage mining can use various data mining or machine learning techniques to model and understand web user activity. A. Handling Profile Evolution Since dynamic aspects of Web usage have recently become important it is desirable to study and discover Web usage patterns. News Dude [8] is an intelligent agent built to adapt to changing user’s interests by learning two separate user models that represent short-term and long-term interests. In [9], a user profiling system was developed based on monitoring, the user’s Web browsing and e-mail habits. In [9], as user-profiling system was developed based on monitoring the user’s web browsing and e-mail habits. This system used a clustering algorithm to group user interests into several interest themes, and the user profiles had to adapt to changing interests of the users over time. The above approaches are based on supervised learning framework and the present work is based on an unsupervised learning framework that tries to learn mass anonymous user profiles on the server side. In this work the web logs are fully mined for each period and then the subsequent results are compared. III. PROFILE DISCOVERY BASED ON WEB USAGE MINING
1) Preprocess Web log file to extract user sessions 2) Cluster the user sessions by using Hierarchical Unsupervised Niche Clustering 3) Summarize session clusters / categories into user profiles 4) Enrich the user profiles with additional facets by using additional web log data and external domain knowledge 5) Track current profiles against existing profiles. A. Preprocessing the Web Log File and Clustering sessions into an optimal number of categories Each log entry in the file consists of access time, IP address, URL viewed, REFERRER etc. The main step in preprocessing is that it maps the URLs on a web site to distinct indices. Sessions are implemented as lists instead of vectors thus saving memory and computational costs. For Clustering Unsupervised Niche Clustering algorithm is used. H-UNC uses Genetic Algorithm and handles noise in the data and automatically determines the number of clusters. After grouping the sessions into clusters, session categories are summarized in terms of user profile vectors [6] [7]. The vector captures the relevant URL in the profile. The profiles are then converted into binary vectors. Each profile is discovered along with a measure that represents the amount of variance or dispersion of user sessions in a given cluster around the cluster representative. This measure is useful in determining the boundary of each cluster and also determines whether two profiles are compatible or not.
Fig. 1. Web usage mining process and discovered profile facets
IV. TRACKING EVOLVING USER PROFILES Each profile is discovered along with an automatically determined measure that determines the amount of variance or dispersion of the user sessions in a given cluster around the cluster representative. This measure determines the boundary around each cluster and also determines whether two boundaries overlap. This comparison leads to events such as Persistence, Birth and Death. Aggregating these events helps in tracking profiles over many periods. Profiles retrieved should be as close as possible to the original session data.
The automatic identification of user profiles is a knowledge discovery task consisting of periodically mining new contents of the user access log files is summarized on the following steps:
Velammal College of Engineering and Technology, Madurai
Page 175
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
V. WEB PERSONALIZATION Web personalization is defined as any action that adapts the information or services provided by a web site to the needs of a particular user or set of users, taking advantage of the knowledge gained from the user’s navigational behavior and individual interests, in combination with the content and structure of the web site. The overall process of usage-based web personalization consists of five modules, which correspond to each step of the process. They are 1) User Profiling-it is the process of gathering information specific to each visitor either explicitly or implicitly. 2) Log analysis and Web usage mining- Information stored in Web server logs is processed by applying data mining techniques in order to a. Extract statistical information and discover interesting usage patterns b. Cluster the users groups according to their navigational behavior c. Discover potential correlations between the web pages and user groups. 3) Content Management- It is the process of classifying the content of a web site in semantic categories in order to make information retrieval and presentation easier for the users. 4) Web site publishing- It is used to present the content stored locally in a Web server and / or some information retrieved from other Web resources in a uniform way to the end-user. 5) Information acquisition and searching-Since the users are interested in information from various Web sources searching and relevance ranking techniques must be employed both in the process of acquisition of relevant information and in the publishing of the appropriate data to each group of users. VI. THE PERSONALIZATION SOLUTION Web usage mining techniques when combined with the multiagent architecture gives a personalization solution to web sites. Multi-agent architecture consists of a set of autonomous agents interacting together to fulfill the main goal of the system. Agent’s taps into the communication stream between a user’s web browser and the web itself. Agents observe the data flowing along the stream, observe the data flowing along the stream and alter the data as it flows past. These agents can learn about the user, influence what the user sees by making up pages before passing them on, and provide entirely new functions to the user through the web browser. Agents are divided into modules that have well defined tasks and that are further divided into two working groups such as data mining module and personalization module. The personalization agent uses the user model knowledge along with the previously discovered sequential patterns and applies a set of personalization rules in order to deliver
Velammal College of Engineering and Technology, Madurai
personalization tasks or functions. The functions based upon the personalization policy provide a complete personalization solution .The multi-user personalization policy used is static and are adjusted to the browsing context of the user. VII.
INTERACTIONS AMONG AGENTS
Agents co-operate and co-ordinate their work by sending messages among them. The interface agent captures the navigational behavior of the user and informs the user data agent and online classification agent about it. The user data agent identifies the current user, creates and records a new navigational session for him. The classification agent classifies the active session in one of the groups discovered by the clustering agent using the information embedded in his or her session. The classification agent informs the evaluation agent about the group. This is notified to the decision maker agent. The decision maker agent informs the personalization agent about the personalization functions to execute according to the adaptation rules associated to that group of user. The personalization agent sends the results to the interface agent, which displays them to the user as response to his or her request.
Page 176
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 2. Functional structure of PWUM Architecture
VIII. RESULTS H-UNC [10] was applied on a set of Web sessions preprocessed from Web log data for several months. After filtering, the data was segmented into sessions based on the client IP address. After filtering the irrelevant URLs unique sessions were obtained. H-UNC partitioned the web user sessions of each period into several clusters and each cluster was characterized by one of the profile vectors. Web Usage Mining multi-agent system for web personalization enhances the quality of discovered models and hence optimizes the personalization process. The software agents of PWUM have been implemented using multi agent platform called JADE [11]. The results obtained by using both of multi agents systems and WUM techniques were very encouraging. IX CONCLUSIONS We presented our system and described the mechanism necessary for Web Usage Mining and Personalization tasks. The combination of more than one technique of WUM enhances the quality of discovered models, so this optimizes the personalization process. The use of multi-agent paradigm reduces the time complexity. We are looking forward testing our approach in tourism web sites as part of national research projects. REFERENCES [1] H. Lieberman. Letizia. An agent that assists web browsing. In Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence, pages 924-929, 1995. [2] T.W.yan, M.Jacobsen, H.Garcia-Molina, and U.Dayal. From user access patterns to dynamic hypertext linking. In Proceeding of the fifth International World Wide Web Conference, Paris, 1996. [3] T.Joachims, D.Freitag and T. Mitchell. Web watcher: a tutor guide for the World Wide Web. In proceedings of the fourteenth International Joint Conference on Artificial Intelligence, pages 924-929, 1995. [4] D.Madenic. Machine Learning used by personal web watcher. In proceddings of the workshop on Machine learning and Intelligent Agents (ACAI-99), Chania, Greece, July 1999. [5] R.Cooley, B.mobasher, and J.Srivastava, “ Web mining: Information and Pattern Discovery in the World Wide Web,” Proc. Ninth IEEE Int’l Conf. Tools with Ai (ICTAI ’97),pp.558-567,1997. [6] O.Nasraoui, R.Krishnapuram, H.Trigui and A.Joshi, “Extracting Web User profiles Using Relational Competitive Fuzzy Clustering” Int’l J. Artificial Intelligence Tools, vol 9,no. 4, pp.509-526,2000. [7] J.Srinivastava, R.Cooley, M.Deshpande, and P.N.Tan, “Web Usage Mining: Discovery and Applications of Usage Patterns from Web Data,”SIGKDD Explorations, vol. 1, no.2,pp1-12, Jan 2000. [8] D.Billus and M.J.Pazzani , “ A Hybrid user Model for News Classification, “ Proc . Seventh Int’l Conf. User Modeling (UM ’99), J.Kay,ed., pp.99-108,1999 [9] I.Grabtree and S.Soltysiak, “Identifying and Tracking Changind Interests”, Int’l J. Digital Libraries, vol.2, pp.38-53 10] O.Nasraoui, R.Krishnapuram, “A New Evolutionary Approach to Web Usage and Context Sensitive Associations Mining” Int’l J.Computational Intelligence and Applications, special issue on Internet intelligent systems, vol.20, no.3, pp.339-348, Sept 2002. [11] Bellifemine F.etal, 2004 JADE Basic Documentation: Programmer’s Guide.
Velammal College of Engineering and Technology, Madurai
Page 177
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Automated Test Case Generation and Performance Analysis for GUI Application Ms. A.Askarunisa#1, Ms. D. Thangamari*2 #
Assistant Proffessor,Department of Computer Science and Engineering, Affiliated to Anna University, Thirunelveli Thiagarajar College of Engineering, Madurai, Tamilnadu, India. 1
[email protected],
*
Department of Computer Science and Engineering, Affiliated to Anna University, Thirunelveli Thiagarajar College of Engineering, Madurai, Tamilnadu, India. 2
[email protected]
Abstract— A Common method for GUI testing is the Capture and Replay (CR) technique.GUIs is complex pieces of software. Testing their correctness is challenging for several reasons: 1. Tests must be automated, but GUIs are designed for humans to use.2. Conventional unit testing, involving tests of isolated classes, is unsuitable for GUI Components. 3. GUIs respond to user-generated events 4. Changes in the GUI’s layout shouldn’t affect robust tests.5.Conventional test coverage criteria, such as “90 percent coverage of lines of code”. This paper proposes GUI Automation testing framework to test GUI-Based java programs as an alternative to the CR technique. The framework develops GUI-event test specification language for GUI application written using java swing APIs, which initiates an automated test engine. Visual editor helps in viewing the test runs. The test engine generates GUI events and captures event responses to automatically verify the results of the test cases. The testing framework includes the test case generation, test case execution and test case verification modules. The testing efficiency is measured by determining coverage metric based on Code coverage, Event coverage and Event Interaction coverage, while may be useful during Regression Testing. The paper uses Abbot and JUnit tools for test case generation and execution and Clover tool for code coverage. We have performed testes on various GUI applications and the efficiency of the framework is provided. Keywords— Automated testing, Coverage, GUI Testing, Test Suite Reduction
I. INTRODUCTION Test automation of GUI means mechanizing the testing process where testers use software in controlling the implementation of the test on the new products and comparing the expected and the actual outcomes of the product application. Scheduled testing tasks on a daily basis and repeating the process without human supervision is one advantage of automated testing. With all the mass production of gadgets and electronic GUI devices, the testing period seems quite demanding. Electronic companies must ensure quality products to deliver excellent products and maintain customer preferences over their products. In running automatic tests for the GUI application, the tester saves much time, especially when he is in a huge production house and needs to be multi-tasking. There are actually four strategies to test a GUI.1. Window mapping assigns certain names to each element, so the test is more manageable and
Velammal College of Engineering and Technology, Madurai
understandable.2. Task libraries sort the step sequence of the user task when they appear in multiple tests.3. The Datadriven type of test automation separates the limitation of the test case and test script so the test script will be reusable.4. Keyword-driven test automation converts tests as spreadsheets or tables, as it creates parsers to decode and perform the description of the test. The reminder of this paper is organized as follows: Section 2 covers the Background material for this proposal, Section 3 describes the proposed approach for GUI testing framework. Section 4 briefly highlights the implementation details. Section 5 gives the Conclusion and future enhancement. II. BACKGROUND AND RELATED WORK Existing works on GUI testing are mainly concerned with test automation assisting tools supporting the capture/replay technique. Researches have considered various techniques for GUI application testing, and coverage metric for test case. White et al [2, 3] and Belli [4] developed model based testing for GUI application under test. Each responsibility is simply the desired response for the user and can be specified as a complete interaction sequence (CIS) between the user and the GUI application under test. Then a finite-state machine is developed for each CIS, which generates the required tests and materializes the CIS. In the work of Memon et al [5], the GUI under test is modelled as a finite-state machine with hierarchical structure. The test case generation problem of GUI testing then follows the goal-oriented philosophy and is treated as an AI(Artificial Intelligence) planning problem. The approach can be viewed as a global one in the sense that a single or global finite-state machine is constructed for all test cases of interest. In the work of Cai et al [6], a GUI test case is defined as a word defined over a finite alphabet that comprises primitive GUI actions or functions of concern as symbols. Meyer [7] defined Capture/Replay Testing technique, a test case as an input with its expected output. Binder [12] defined a test case to consist of a pretest state of the software under test (including its environment), a sequence of test inputs, and a statement of expected test results. Memon et al [1,5] defined a test case to consist of an initial state and a legal action sequence. The problem here is how to define the state of the
Page 178
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
GUI. A few approaches have been proposed for GUI testing. The most known one may be the so-called capture/replay technique [8]. In the work of Sun [21] propose a specification-driven approach to test automation for GUI based Java programs as an alternative to the use of capture/replay. However the rationale of existing test case definitions is rarely discussed. An exception is the recent work of Rothermel et al [15], which assesses the effects of test suite granularity and test input grouping on savings in test execution time and costs in defectdetection effectiveness in the context of regression testing. Test suite granularity pertains to the size of the test cases so grouped or the number of inputs per test case, whereas test suite grouping pertains to the content of test cases or the degree of diversity among the inputs that compose a test case. Memon, Pollack and Sofia [5, 11, 12] exploit planning techniques developed and used extensively in artificial intelligence to automate the generation of test cases for GUIs. Given a set of operators, an initial state and a goal state, the planner produces a sequence of the operators that will transform the initial state to the goal state.
for extending the basic tester classes for new components, and has a well defined method for naming those new actions so they are automatically available at the scripting level. It also provides extensibility for Converting strings from scripts into arbitrary classes, and introducing new Individual steps into scripts. Scripts can call directly into java code (the script is actually just a thin veneer over method calls). Abbot provides both a script environment and a JUnit [13] fixture, both of which handle setup and teardown of the complete GUI environment.
III. PROPOSED APPROACHES FOR TESTING GUI APPLICATION In the Previous Existing methods, none of the tool does not support both capture replay technique and programmatic technique. Graphical user interface or GUI is the interactive graphical representation of its underlying application. Front end users find it easy to use this interface to operate the software since it is not necessary for them to understand the programming language used in the software. Testing such software for its reliability becomes complex since the testers have to test the software as well as the GUI for its design functionality. The proposed testing is based on functionality and in order to perform functional testing, different kinds of automated tools are available in market which will be quite hard for the tester to choose for satisfying his requirements. In this paper represents Abbot tool that supports both capture replay techniques (black box Testing) and Glass box technique (White box Testing) is used for test case narration and execution. This process is also called as Gray Box Testing. A.TestCase Generation Framework Testing Framework [9] application Involves, (Fig1) Test Case Generation, Testing Oracle, Automatic Executions [21], Test Result and Performance Analysis. Abbot uses a dedicated class to represent GUI components .It stores a range of attributes to be used when looking for a matching component in the hierarchy, and does a fuzzy lookup to avoid spurious errors due to component movement or minor changes to the GUI. The lookup mechanism is very general due to the fact it is used by the scripting layer which has no a priori knowledge of the GUI hierarchy or what needs to be looked up. The framework also provides a number of utilities to facilitate inspecting the GUI hierarchy itself. Abbot provides
Velammal College of Engineering and Technology, Madurai
Fig 1 Testing Framework
1) Test Case Generation Test case Generation and Execution is done using Abbot Tool. Abbot is a framework for driving java UI components programmatically. Costello is a script editor and a launcher which accompanies Abbot. They are based on the same capture/ playback/script principle that marathon uses, but Abbot can also be used programmatically. The framework can be invoked differently from Java code or accessed without programming through the use of 20 scripts. It is suitable for use both by developers for unit tests and QA for functional testing. The framework easily integrates with the JUnit test harness and therefore, during application development, the functional GUI tests can become a part of the test suite. All these features of Abbot make it an effective framework for rapidly creating a comprehensive test framework. Collections of test cases are generated by using Abbot and JUnit Tool. The test cases contain the sequence of user Input Events. GUI Model
Test Case Generation
JAR Files
Page 179
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Measurement is the process by which numbers or symbols are assigned to attributes of entities in the real world in such a way as to characterize them according to clearly defined rules. Coverage of GUI applications Require various tasks: Performance analysis and Coverage Report. Coverage measurement also helps to avoid test entropy. As your code goes through multiple release cycles, there can be a tendency for unit tests to atrophy. As new code is added, it may not meet the same testing standards you put in place when the project was first released. Measuring code coverage can keep your testing up to the standards you require. You can be confident that when you go into production there will be minimal problems because you know the code not only passes its tests but that it is well tested.
1) Code Coverage Code coverage analysis is sometimes called test coverage analysis. The two terms are synonymous. The academic world more often uses the term "test coverage" while practitioners more often use "code coverage". Likewise, a coverage analyser is sometimes called a coverage monitor. Code coverage is not a panacea. Coverage generally follows an 8020 rule. Increasing coverage values becomes difficult, with new tests delivering less and less incrementally. If you follow defensive programming principles, where failure conditions are often checked at many levels in your software, some code can be very difficult to reach with practical levels of testing. Coverage measurement is not a replacement for good code review and good programming practices. In general you should adopt a sensible coverage target and aim for even coverage across all of the modules that make up your code. Relying on a single overall coverage figure can hide large gaps in coverage. 2) Code Coverage with Clover Clover [24] uses source code instrumentation, because although it requires developers to perform an instrumented build; source code instrumentation produces the most accurate coverage measurement for the least runtime performance overhead. As the code under test executes, code coverage systems collect information about which statements have been executed. This information is then used as the basis of reports. In addition to these basic mechanisms, coverage approaches vary on what forms of coverage information they collect. There are many forms of coverage beyond basic statement coverage including conditional coverage, method entry and path coverage. Clover is designed to measure code coverage in a way that fits seamlessly with your current development environment and practices, whatever they may be. Clover's IDE Plug-in provide developers with a way to quickly measure code coverage without having to leave the IDE. Clover's Ant and Maven integrations allow coverage measurement to be performed in Automated Build and Continuous Integration systems, and reports generated to be shared by the team. The Clover Coverage Explorer: The Coverage Explorer allows you to view and control Clover's instrumentation of your Java projects, and shows you the coverage statistics for each project based on recent test runs or application runs. The main tree shows coverage and metrics information for packages, files, class and methods of any Clover-enabled project in your workspace. Clover will auto-detect which classes are your tests and which are your application classes - by using the drop-down box above the tree you can then restrict the coverage tree shown so that you only see coverage for application classes, test classes or both. Summary metrics are displayed alongside the tree for the selected project, package, file, class or method in the tree. The Clover Coverage Measurement: Clover uses these measurements to produce a Total Coverage Percentage for each class, file, and package and for
Velammal College of Engineering and Technology, Madurai
Page 180
Fig 2 Test Case Generation using Abbot Tool
This framework shown in Fig 2 chooses a specific model of the GUI application. This model satisfies the GUI application under test which is also the input of the Test Case generation. Collections of test cases are generated by using Abbot and JUnit Tool. The Test case contains the sequence of user Input events. The test cases are run with test runner. A test designer interacts with the GUI and generates mouse and keyboard events. 2) Test Case Execution Test cases from the repository are executed one by one automatically using Abbot and JUnit Tools as shown in Fig 3. Source Program
Jar files
Repository
Collection of test cases
Test Execution Automatically
Fig 3 Test Execution by JUnit Tool
3) Test case Verification: The expected results of various test cases are manually determined and stored in the GUI model of Testing Oracle. The Testing Oracle contains the expected state of sequences for each application. When the test cases start running, the start timer is initialized. The events are generated automatically and the actual results from the tool are verified with the expected results as shown in Fig 4. Pass/Failure testing report is generated accordingly. Actual State
Expected State
Automated Verified
Verified
Fig 4 Test Case Verification
B. Performance Analysis
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the project as a whole. The Total Coverage Percentage allows entities to be ranked in reports. The Total Coverage Percentage (TPC) is calculated as follows: TPC = (BT + BF + SC +MC)/(2*B + S + M) where BT - branches that evaluated to "true" at least once BF - branches that evaluated to "false" at least once SC – statements covered MC - methods entered B - total number of branches S - total number of statements M - total number of methods 3) Event Coverage These coverage criteria use events and event sequences to specify a measure of test adequacy. Since the total number of permutations of event sequences in any non-trivial GUI is extremely large, the GUI's hierarchical structure is exploited to identify the important event sequences to be tested. A GUI is decomposed into GUI components, each of which is used as a basic unit of testing. A representation of a GUI component, called an event flow graph [16], identifies the interaction of events within a component and intra component criteria are used to evaluate the adequacy of tests on these events. The hierarchical relationship among components is represented by an integration tree, and inter-component coverage criteria are used to evaluate the adequacy of test sequences that cross components. 4) Event Interaction Coverage The sequence of possible Event Interacts with an other Event. This type of event coverage is called as Event Interaction Coverage [22]. The event interaction coverage is consisting of 2 ways and 3 ways Combination. C. Coverage Report 1) Coverage HTML Report The clover html report task generates a full HTML report with sensible default settings. It is also generated prior to generation of the full report. 2) Coverage XML Report The clover xml report task generates a full HTML report with sensible default settings. It is also generated prior to generation of the full report. 3) Coverage PDF Report The clover pdf report task generates a PDF report with sensible default settings. It is also generated prior to generation of the full report. D. Coverage Metric The coverage metric CONTeSSi (n) (CONtext Test Suite Similarity) [23] for each model value factor is calculated and compared with the original pair of model. CONTeSSi (n) (CONtext Test Suite Similarity) that explicitly considers the context of n preceding events in test cases to develop a new “context-aware” notion of test suite similarity. This metric is an extension of the cosine similarity metric used in Natural Language Processing and Information Retrieval for comparing an item to a body of knowledge, e.g., finding a query string in a collection of web pages or determining the likelihood of finding a sentence in a text corpus (collection of documents).We evaluate CONTeSSi (n) by comparing four test suites, including suites reduced using conventional criteria, for four open source applications. Our results show
that CONTeSSi (n) is a better indicator of the similarity of test suites than existing metrics.
Velammal College of Engineering and Technology, Madurai
Page 181
This paper considers different models with varying frequencies of events like all individual events, all two pair events, all three pair events, etc. Coverage metric is calculated for various factors like statement, branch, method, etc. IV. IMPLEMENTATION This paper uses the GUI model of Calculator Application, which is written using Java Swing. A scientific Calculator Application contains a Collection of Standard Buttons, Control Buttons and Scientific Buttons. It is used to calculate the Arithmetic and Scientific data values. It performs several Basic operations such as Add, Sub, Mul and Div and Also Scientific Operations Such as Sin, Cos, Tan, Log, sqrt and etc,The Program also Contains Radio Buttons like Hexadecimal, Octal, Decimal, and Binary, which enable the The automatic execution of user input sequence is shown in Fig 5. All the operations are performed by Click Events .For this Application, the test cases are written by using java Swing. The unit testing is done by Abbot and JUnit Tool.
Figure 5: Automatic Execute of User Input Sequence
A. Automation Testing with Abbot The Specific Calculator Program written using Java Swing is selected for testing. The Calculator program is run with abbot and junit tool and the each action is performed using the button click event. The Calculator Application, test case generation makes use of the core components Component Test Fixture () and Component Tester () in the Abbot and also write the test cases are using action click (), Selectmen Item () , action Delay ().By making use of the components test cases are created. The assert method compares data value obtained from calculator program with data value automatically execute one by one using the test case sequences. The click event for Hexa Decimal option Button and the corresponding enabled buttons are shown in Fig 6
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1) Code Coverage Code coverage and event based performance analysis are done in the module. Clover coverage tool is used for collecting the metrics for frames, methods and package. Code coverage tool is used for statements, branches and loop.
Figure 6: Click Event for Hexa Decimal Button
B. Coverage View Manual analysis is done for event interaction and events and the relevant data is collected. Coverage metrics are collected for different units of application.The different explorer view of coverage report for all completed test cases is show in fig 7,8,9 Fig 10: Coverage view of Code Cover tool for Calculator program
Figure 7: Coverage Explorer View for all Test cases
2) Event Coverage Event Coverage is determined by the total number of click events generated by the user. For Eg table3 shows that Single event is generated for Basic button, 11 events are generated for calculating the area of circle (pi*r*r). 3) Event Interaction Testing It represents all possible sequences of events that can be executed on the GUI. In that Calculator Application contains as the collection Buttons such as standard, Control, and Scientific. In Figure 11 represents that, • e1 represents the clicking in the file Menu • e2 represents the clicking in the Menu Selection event • e4 represents the clicking in the Basic button after clicking the e2 event. It is similar to e5.
Fig 8: Project Coverage Report in XML Format
Fig 11 Shows that Event Based Sequences
Fig 9: Coverage Report in PDF Format
Velammal College of Engineering and Technology, Madurai
Table1 displays the report title and the time of the coverage contained in the report. The header displays metrics for the package files or project overview which is currently selected. Depending on the current selection, the metrics include all or a subset of: Number of Lines of Code (LOC),Number of Non-commented Lines of Code (NCLOC),Number of Methods, Number of Classes, Number of Files, Number of Packages.
Page 182
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
In considering this context, the event pair coverage suite in Table 4 is expected to be more similar to the original suite than the event coverage suite, since the event pair coverage suite is created based on the existence of event pairs. Table 4(b) shows the count of each event pair for each suite. This is the basis of CONTeSSi (n), for n = 1, since we are looking at events in the context of one other (previous) event. Now if we extend this example to compute CONTeSSi (2), we obtain the frequencies shown in Table 4(C). In general, as n increases, the frequencies for the event sequences decrease, as they appear less frequently in the test suites. Intuitively, comparing test suites on longer sequences will make it harder for the test suites to be similar. Therefore, if two test suites have a high similarity score with a larger n, they are even more similar than two suites being compared with a small n. By treating each row in Table 4 (a), (b), or (c) as a vector, CONTeSSi is computed as
follows:CONTeSSi(A,B) =(A · B) /(|A| × |B|) (1)where A and B are the vectors corresponding to the two test suites, A · B is the dot product of the two vectors, i.e.,Pj i=1(Ai×Bi) where j is the number of terms in the vector; and |A| = qPj i=1(Ai)2. The value of CONTeSSi lies between 0 and 1, where a value closer to 1 indicates more similarity. Hence, CONTeSSi (n) is computed as shown in Equation 1, creating a vector for each suite, representing the frequencies of all possible groups of n + 1 events. The inclusion of n previous events will increase the number of terms in the vector, thereby increasing j. The values in Table 5 show the values of CONTeSSi(n) for all our test suites, for n = 0, 1, 2, 3. From these values, we observe that if we ignore context, i.e., use n = 0, most of the reduced suites are quite similar to the original, as indicated by the high (> 0.9) value of CONTeSSi (0). However, the similarity between the test suites decreases as more context
TABLE:1 VIEW OF ALL TEST CASES VALUE USING CLOVER COVERAGE TOOL
Area Test Case
State ment
Bran ches
Metho ds
Classes
LOC
Circle Rectangle Parallelogram Triangle Trapezoid Total
280 293 291 298 297 339
132 132 132 132 132 132
21 24 24 24 24 28
1 2 2 2 2 2
625 667 666 672 672 738
Rectangle Prims Cylinder Cone Sphere Total
312 304 303 300 294 373
132 132 132 132 132 132
25 25 25 25 24 28
2 2 2 2 2 2
690 681 676 671 668 775
Rectangle Prims Cylinder Cone Sphere Pyramid
291 293 296 297 299 295
132 132 132 132 132 132
24 24 24 24 24 24
2 2 2 2 2 2
666 669 671 672 677 671
Total
351
132
29
2
NCLOC
762
483 513 511 518 517 570 Surface 535 527 525 522 513 602 Volume 511 513 516 517 519 515 586
Velammal College of Engineering and Technology, Madurai
Total Cmp
Cmp Densit y
Avg method Cmp
Stmt / methods
Methods / classes
Total Coverage (in %)
138 141 141 141 141 145
0.49 0.48 0.48 0.47 0.47 0.43
6.57 5.88 5.88 5.88 5.88 5.18
13.33 12.21 12.12 12.42 12.38 12.11
21 12 12 12 12 14
54.3 53.6 53.6 54.1 54.1 55
142 142 142 142 141 145
0.46 0.47 0.47 0.47 0.48 0.39
5.68 5.68 5.68 5.68 5.88 5.18
12.48 12.16 12.12 12 12.25 13.32
12.5 12.5 12.5 12.5 12 14
41.6 42.7 53.3 53.3 51.7 53.6
141 141 141 141 141 141
0.48 0.48 0.48 0.47 0.47 0.48
5.88 5.88 5.88 5.88 5.88 5.88
12.12 12.21 12.33 12.38 12.46 12.29
12 12 12 12 12 12
55 55 55 55 55.5 55
146
0.42
5.03
12.1
14.5
55
Page 183
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE: 2 VIEW OF ALL TEST CASES VALUE USING CODE COVERAGE TOOL
Test Case Basic Minus Add Mul Div Mod Hex Dec Oct Bin
Stmt 59.4 63.9 63.9 63.9 63.9 67.1 61.6 60.7 61.6 61.6
Circle Rectangle Parallelogram Triangle Trapezoid
68.9 63.9 63.9 65.3 65.8
Rectangle Prims Cylinder Cone Sphere
64.4 65.3 68.9 68.9 67.6
Rectangle Prims Cylinder Cone Sphere Pyramid Total
63.9 64.8 68.5 68.5 68.5 64.4 97.7
Branch 2.3 10.5 9.3 11.6 12.8 15.1 11.6 9.3 11.6 14 Area 36.0 12.8 12.8 17.4 18.6 Surface 14.0 16.3 34.9 34.9 30.2 Volume 11.6 15.1 34.9 34.9 33.7 15.1 93
Loop 10.1 8.7 8.7 8.7 8.7 10.1 14.5 13.0 14.5 13
Strict Con 6.0 7.7 7.7 7.7 7.7 9.4 19.7 7.7 15.4 9.4
10.1 8.7 8.7 8.7 8.7
14.5 11.1 10.3 13.7 15.4
8.7 8.7 10.1 10.1 10.1
13.7 13.7 16.2 15.4 9.4
8.7 8.7 10.1 10.1 10.1 8.7 34.8
8.5 10.3 14.5 12.8 12.8 11.1 94.9
Velammal College of Engineering and Technology, Madurai
TABLE: 3 VIEW OF ALL TEST CASE EVENT SEQUENCE
Test Plan
Event
Execution (in sec)
Execution Delay (1000 sec)
Basic Scientific Hex Dec Octal Binary
1 1 1 1 1 1
View 1.684 1.763 1.342 1.295 2.262 1.342
2.699 2.714 2.371 2.309 2.356 2.324
Circle Rectangle Parallelogram Triangle Trapezoid
11 7 6 12 12
Area 3.588 2.434 2.262 3.447 3.401
4.555 3.463 3.26 4.446 4.415
Rectangle Prism Cylinder Pyramid Cones Sphere
6 8 11 10 11 14
Volume 2.278 2.036 3.525 2.995 3.588 4.119
3.291 3.65 4.633 4.009 4.556 5.085
Rectangle Prims Cylinder Cones Sphere
26 18 17 13 9
Surface 6.006 4.524 4.68 3.916 3.183
6.989 5.506 5.647 4.898 4.165
Page 184
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE: 4 EXAMPLE TEST CASES YIELDED FROM SEVERAL REDUCTION TECHNIQUES ORIGINAL PAIR
EVENT PAIR
EVENT
STATEMENT
METHOD
BRANCH
e2,e5 e2,e4,e6 e2,e4,e8 e2,e4,e7 e2,e4,e9,e10,e9 e2,e4,e9 e2,e5,e11,e9,e10,e9 e2,e5,e11,e9 e9,e10,e9 e6,e9,e10,e9 e8,e9 e7,e9,e10,e9 e11,e9,e10,e9 e11,e9,e10 e6,e12 e2,e5,e11,e10 e9,e10,e2,e5,e11,e9
e6,e9 e7,e9 e8,e9 e2,e5,e11,e9,e10,e9 e2,e4,e9,e10,e9 e9,e10,e2,e5,e11,e9
e2,e4,e6 e2,e4,e8 e2,e4,e7 e2,e5,e11,e10 e9,e12 e8,e9
e9,e10,e2,e5,e11,e9 e2,e5 e6,e9 e7,e9 e8,e9 e2,e5,e11,e9,e10,e9
e2,e5,e11,e9 e9,e10,e9
e6,e9 e7,e9 e8,e9 e2,e5,e11,e9 e2,e4,e9,e10,e9
Illustrative Tests
e2 e4 e5 e6 e7 e8 e9 e10 e11 e12
TABLE: 4(A) FREQUENCY OF UNIQUE EVENTS OCCURRING IN THE TEST SUITE (LENGTH=0) TEST SUITE
E2
E4
E5
E6
E7
E8
E9
E10
E11
E12
ORIGINAL
10
5
5
3
2
2
18
9
6
1
EVENT PAIR
3
1
2
1
1
1
9
3
2
0
EVENT
4
3
1
1
1
2
2
1
1
1
STMT
3
0
3
1
1
1
7
2
2
0
METHOD
1
0
1
0
0
0
3
1
1
0
BRANCH
2
1
1
1
1
1
6
1
1
0
Illus. Suite
1
1
1
1
1
1
1
1
1
1
TABLE: 4( B) FREQUENCY OF ALL EVENTS PAIR OCCURRING IN THE TEST SUITE (LENGTH=1) TEST SUITE
E2,E4
E2,E5
E4,E6
E4,E7
E4,E8
E4,E9
E5,E11
E6,E9
E6,E10
E6,E12
E7,E9
E8,E9
ORIGINAL
5
5
1
1
1
2
4
1
1
1
1
1
EVENT PAIR
1
2
0
0
0
1
2
1
0
1
1
1
EVENT
3
1
1
1
1
0
1
0
0
0
0
1
STMT
0
3
0
0
0
0
2
1
0
0
1
1
METHOD
0
1
0
0
0
0
1
0
0
0
0
0
BRANCH
0
1
0
0
0
1
1
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
TEST SUITE
E2,E4
E2,E5
E4,E6
E4,E7
E4,E8
E4,E9
E5,E11
E6,E9
E6,E10
E6,E12
E7,E9
E8,E9
ORIGINAL
5
5
1
1
1
2
4
1
1
1
1
1
EVENT PAIR
1
2
0
0
0
1
2
1
0
1
1
1
EVENT
3
1
1
1
1
0
1
0
0
0
0
1
STMT
0
3
0
0
0
0
2
1
0
0
1
1
METHOD
0
1
0
0
0
0
1
0
0
0
0
0
BRANCH
0
1
0
0
0
1
1
1
0
0
1
1
0
0
0
0
0
0
0
0
0
0
0
0
Illus. Suite
TABLE: 4(C) FREQUENCY OF ALL EVENTS PAIR OCCURRING IN THE TEST SUITE (LENGTH=2)
Illus. Suite
Velammal College of Engineering and Technology, Madurai
Page 185
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE: 5 CONTESSI (N) VALUES FOR SUITE COMPARED TO ORIGINAL FOR ALL BUTTONS IN CALCULATOR APPLICATION IN GUI EXAMPLE SUITES n
EVENT PAIR
EVENT
STMT
METHOD
BRANCH
ILLUS. SUITE
0
0.97308
0.78513
0.95434
0.94405
0.94571
0.7816
1
0.9274
0.60669
0.788116
0.82365
0.685188
0.0000
2
0.9106
0.39509
0.79697
0.79018
0.79018
0.0000
3
0.9428
0.73786
0.82495
0.7071
0.68041
0.0000
4
0.9999
0.0000
0.8660
0.0000
0.5000
0.0000
5
0.9999
0.0000
0.9999
0.0000
0.0000
0.0000
V. CONCLUSIONS AND FUTURE SCOPE GUIs may introduce new types of error, increase complexity and make testing more difficult. The Automation testing reduces the work intensity of the developer and the tester. The main advantage of test automation is that as a result software developers can run tests more often, find and fix bugs on the early stage of development, before end users will face these bugs. The framework suggested in this project simplifies the creation and maintenance of robust GUI tests. Abbot is easy to learn and use, and provides some unique features that can make GUI development more productive. This project developed the method for (i) automatic test case generation, execution and verification (ii) performance analysis by calculating the coverage based on statement, method, branch, event and event interactions for GUI applications. Our results showed that CONTeSSi (n) is a better indicator of the similarity of Event Interaction test suites than existing metrics. In that metric also describes Event pair Coverage is the best Coverage Compared with Statement, Method, and Branch and Loop coverage. As a future enhancement is to conduct more intensive testing with the candidate tools by creating additional test cases and modifying default test settings to improve test coverage and conducting Automation testing. We will also further introduce the relationship between the test case lengths and the value of n used in CONTeSSi(n) to draw some Conclusions on how to pick the best value of n, starting with a value of n which matches the length of the most test cases in the input suite.
[4] F.Belli, "Finite State Testing and Analysis of Graphical User Interfaces", Proc. the 12th International Symposium on Software Reliability Engineering, 2001, pp34-43. [5] A.M.Memon, M.E.Pollack, M.L.So_a, "Hierarchical GUI Test Case Generation Using Automated Planning", IEEE Transactions on Software Engineering, Vol.27, No.2, 2001, pp144-155. [6] K.Y.Cai, L.Zhao, H.Hu, C.H.Jiang, "On the Test Case De_nition of GUI Testing", Proc. the 5th International Conference on Quality Software, IEEE Computer Society Press, 2005, [7] Brad A. Myers.” User interface software tools". ACM Transactions on Computer-Human Inter-action, 2(1):64103, 1995. [8] Capture-Replay Tool, 2003. http://soft.com. [9] T.Wall. "Abbot Java gui test framework", 2004.http://abbot.sourceforge.net/. [10] Tamas Daboci,Istvan Kollar,Tamas Megyeri,"How to Graphical User Interfaces" IEEE Instrumentation and Measurement Magazine,Sep 2003. [11] A. M. Memon, M. E. Pollack, and M. L. Soffa."Using a goal-driven approach to generate test cases for guis". In 10 Proceedings of the 21st international conference on Software engineering, IEEE Computer Society Press, 1999. [12] A. M. Memon, M. E. Pollack, and M. L. So_a. "Hierarchical gui test case generation using automated planning". IEEE Trans. Softw. Eng.,27(2):144-155, 2001. [13] E. Gamma and K. Beck."Gui tools for junit", 2004. http://www.junit.org/ [14] Atif M Memon."Advances in GUI testing". In Advances in Computers, ed. by Marvin V.Zelkowitz, volume 57. Academic Press, 2003.
REFERENCES [1] A. M. Memon."Gui testing: pitfalls and process". IEEE Computer, 35(8):87-88, 2002. [2] L.White, H.Almezen, "Generating Test Cases for GUI Responsibilities Using Complete Interaction Sequences", International Symposium on Software Reliability Engineering, 2001, pp54-63.Proc. the 11th International Symposium on Software Reliability Engineering, 2000, pp110-121. [3] L.White, H.Almezen, N.Alzeidi, "User-Based Testing of GUI Sequences and Their Interactions", Proc. the 12thInternational Symposium on Software Reliability Engineering,
[15] Monty L. Hammontree , Rothermel,Je_rey J. Hendrickson , Billy W.Hensley,"Integrated data capture and analysis tools for research and testingon graphical user interfaces", Proceedings of the SIGCHI conference on Human factors in computing systems, p.431-432, May 03-07, 1992, Monterey,California, United States [16] Atif M. Memon,"An Event-ow Model of GUI-Based Application for Testing" STVR2007. [17] Richard K. Shehady , Daniel P. Siewiorek," A Method to Automate User Interface Testing Using Variable Finite State
Velammal College of Engineering and Technology, Madurai
Page 186
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Machines, Proceedings of the 27th International Symposium on Fault-Tolerant Computing (FTCS'97), June 25-27, 1997. [18] Lee White , Husain Almezen,"Generating Test Cases for GUI Responsibilities Using Complete Interaction Sequences", Proceedings of the 11th International Symposium on Software Reliability Engineering. [19] Atif M. Memon, Qing Xie."Studying the Fault Detection Effectiveness of GUI Test Cases for Rapidly Evolving Software",Manuscript received 15 Mar. 2005; revised 22 June 2005; accepted 28 June 2005; [20] B. Marick,"When Should a Test Be Automated?" Proc. 11th Intl Software/Internet Quality Week, May 1998. [21] A.Sun, M. Finsterwalder, "Automating Acceptance Tests for GUI Applications in an Extreme Programming
Environment", Proc Second Intl Conf. EXtreme Programming and Flexible Processes in Software Eng. May 2001. [22] L. White, H. AlMezen, and N. Alzeidi, "User-Based Testing of GUI Sequences and Their Interactions", Proc. 12th Intl Symp. Software Reliability Eng., pp. 54-63, 2001. [23] A. Brooks and Atif M.Memon,"Introducing a Test Suite Similarity Metric for Event Sequence-based Test cases", 2007. [24] Clover coverage tool www.atlassian.com/software/ clover [25] www.planetsourcecode.com/calculator source code
Velammal College of Engineering and Technology, Madurai
Page 187
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
THE NEED OF THE HOUR - NOSQL Technology for Next Generation Data Storage K.Chitra1, Sherin M John2 Dept of Computer Applications, Thiagarajar School of Management Madurai, TN, India 1
[email protected] [email protected]
2
Abstract— A relational database is a table-based data system where there is no scalability, minimal data duplication, computationally expensive table joins and difficulty in dealing with complex data. The problem with relations in relational database is that complex operations with large data sets quickly become prohibitively resource intense. Relational databases do not lend themselves well to the kind of horizontal scalability that's required for large-scale social networking or cloud applications. NOSQL has emerged as a result of the demand for relational database alternatives. The biggest motivation behind NOSQL is scalability. NOSQL is meant for the current growing breed of web applications that need to scale effectively. This paper analyzes the need of the next generation data storage which is the need of the current large-scale social networking or cloud applications. We also analyze the capabilities of various NOSQL models like BigTable, Cassandra, CouchDB, Dynamo and MongoDB. Keywords— NOSQL, scalability, relational database alternatives, Next Generation Data Storage.
VI. INTRODUCTION A relational database management system (RDBMS) is a database management system (DBMS) that is based on the relational model as introduced by E. F. Codd. It supports a tabular structure for the data, with enforced relationships between the tables. Most popular commercial and open source databases currently in use are based on the relational model. The problem with RDBMS is not that they do not scale, it’s that they are incredibly hard to scale. The most popular RDBMS are Microsoft SQL Server, DB2, Oracle, MYSQL etc. Many Web applications simply do not need to represent data as a set of related tables that means all applications need not be a traditional relational database management system (RDBMS) that uses SQL to perform operations on data. Rather, data can be stored in the form of objects, graphs, documents and retrieved using a key. For example, a user profile can be represented as an object graph (such as pojo) with a single key being the user id. Another example: documents or media files can be stored with a single key with indexing of metadata handling by a separate search engine.
Velammal College of Engineering and Technology, Madurai
These forms of data storage are not relational and lack SQL, but they may be faster than RDBMS because they do not have to maintain indexes, relationships, constraints and parse SQL. Technology like that has existed since the 1960s (consider, for example, IBM’s VSAM file system). Relational databases are able to handle millions of products and service very large sites. However, it is difficult to create redundancy and parallelism with relational databases, so they become a single point of failure. In particular, replication is not trivial. To understand why, consider the problem of having two database servers that need to have identical data. Having both servers for reading and writing data makes it difficult to synchronize changes. Having one master server and another slave is bad too, because the master has to take all the heat when users are writing information. So as a relational database grows, it becomes a bottleneck and the point of failure for the entire system. As mega e-commerce sites grew over the past decade they became aware of this issue - adding more web servers does not help because it is the database that ends up being a problem.
Fig. 2 Scalability Issues with Relational Database
VII.
NEXT GENERATION DATA STORAGE MODEL
Page 188
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Trends in computer architectures and modern web-scale databases are pressing databases in a direction that requires: • storage of non-relational data, which are distributed to be available even when the nodes are crashed • open-source • effective scalability (Horizontal & Vertical) • schema-free • replication support • easy API • eventual consistency
NOSQL-style data stores attempt to address these requirements. It is the future of database technology that is suitable particularly for web applications that need to scale effectively. Many NOSQL products store data as BLOBs. Actually, NOSQL is an offhand term for a collection of technologies, some of which are DBMS (DataBase Management Systems) and some of which are not. All they have in common is: • They have something to do with managing data • They are not relational DBMS, if by “Relational DBMS” you mean “DBMS that want you to talk to them in SQL.” • select fun, profit from real_world where relational=false; XXXXXXX.
NOSQL (Not Only SQL)
The NoSQL movement is a combination of an architectural approach for storing data and software products that can store data without using SQL. Hence, the term NoSQL. The NOSQL products that store data using keys are called Key-Value (KV) stores. Because these KV stores are not relational and lack SQL they may be faster than RDBMS. The downside of NOSQL is that you cannot easily perform queries against related data. Prominent closed-source examples are Google's BigTable and Amazon's Dynamo. Several opensource variants exist including Facebook's Cassandra, Apache HBase, CouchDB LinkedIn's Project Voldemort and many others. Apache's open source CouchDB offers a new method of storing data, in what is referred to as a schema-free document-oriented database model. 4) The technology There are three popular types of NoSQL databases.
Column-oriented databases: Rather than store sets of information in a heavily structured table of columns and rows with uniform sized fields for each record, as is the case with relational databases, column-oriented databases contain one extendable column of closely related data. Facebook created the high-performance Cassandra to help power its website. The Apache Software Foundation developed Hbase, a distributed, open source database that emulates Google’s Big Table. Document-based stores: These databases store and organize data as collections of documents, rather than as structured tables with uniform sized fields for each record. With these databases, users can add any number of fields of any length to a document. VIII. NOSQL MODELS FOR DATA STORAGE Organizations that collect large amounts of unstructured data are increasingly turning to non-relational databases, now frequently called NoSQL databases. YYYYYYY. Bigtable: A Distributed Storage System for Structured Data Bigtable is Google’s internal database system. Bigtable is a distributed storage system for managing structured data that is designed to scale to a very large size i.e. petabytes (1 petabyte = 1.12589991 × 1015 bytes) of data across thousands of commodity servers. Many projects at Google store data in Bigtable, including web indexing, Google Earth, and Google Finance. These applications place very different demands on Bigtable, both in terms of data size (from URLs to web pages to satellite images) and latency requirements (from backend bulk processing to real-time data serving). Despite these varied demands, Bigtable has successfully provided a flexible, high-performance solution for all of these Google products. A Bigtable is a sparse, distributed, persistent multidimensional sorted map. The map is indexed by a row key, column key, and a timestamp; each value in the map is an uninterpreted array of bytes. Each cell in a Bigtable (like field in DBMS) can contain multiple versions of the same data; these versions are indexed by timestamp (Microseconds). Bigtable timestamps are 64-bit integers. Different versions of a cell are stored in decreasing timestamp order, so that the most recent versions can be read first. Bigtable depends on a cluster management system for scheduling jobs, managing resources on shared machines, dealing with machine failures, and monitoring machine status. ZZZZZZZ. Cassandra: A Decentralized, highly scalable, eventually consistent DB
Key-value stores: As the name implies, a key-value store is a system that stores values indexed for retrieval by keys. These systems can hold structured or unstructured data. Amazon’s SimpleDB is a Web service that provides core database functions of information indexing and querying in the cloud. It provides a simple API for storage and access. Users pay only for the services they use.
Cassandra is highly scalable second-generation distributed database. Cassandra was open sourced by Facebook in 2008 and is currently being developed as an Apache Incubator project. The system offers a fault tolerant, high availability, decentralized store for data which can be scaled up by adding
Velammal College of Engineering and Technology, Madurai
Page 189
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
hardware nodes to the system. Cassandra implements an "eventually consistent" model which trades-off consistency of data stores in the system for availability. Data is automatically replicated to multiple nodes for faulttolerance. Replication across multiple data centers is supported. Failed nodes can be replaced with no downtime. Cassandra is in use at Rackspace, Digg, Facebook, Twitter, Cisco, Mahalo, Ooyala, and more companies that have large, active data sets. The largest production cluster has over 100 TB of data in over 150 machines. AAAAAAAA. CouchDB – Extremely scalable, highly available and reliable Storage System CouchDB, is a free and open source document-oriented database written in the Erlang programming language for its emphasis on fault tolerance, accessible using a RESTful JavaScript Object Notation (JSON) API. The term "Couch" is an acronym for "Cluster Of Unreliable Commodity Hardware", reflecting the goal of CouchDB being extremely scalable, offering high availability and reliability, even while running on hardware that is typically prone to failure. Hence CouchDB is: •
A document database server
•
Ad-hoc and schema-free with a flat address space
•
Distributed, featuring robust, incremental replication
with bi-directional conflict detection and management. •
Highly available even if hardware fails
•
Query-able and index-able, featuring a table oriented
reporting engine that uses Javascript as a query language CouchDB is Not: •
A relational database
•
A replacement for relational databases
•
An object-oriented database. Or more specifically,
meant to function as a seamless persistence layer for an OO programming language BBBBBBBB. Dynamo - A Distributed Storage System for Amazon website Dynamo is an internal technology developed at Amazon to address the need for an incrementally scalable, highlyavailable key-value storage system. The technology is designed to give its users the ability to trade-off cost, consistency, durability and performance, while maintaining high-availability.
Velammal College of Engineering and Technology, Madurai
Amazon runs a world-wide e-commerce platform that serves tens of millions customers at peak times using tens of thousands of servers located in many data centers around the world. Reliability is one of the most important requirements because even the slightest outage has significant financial consequences and impacts customer trust. In addition, to support continuous growth, the platform needs to be highly scalable. Dynamo has a simple key/value interface, is highly available with a clearly defined consistency window, is efficient in its resource usage, and has a simple scale out scheme to address growth in data set size or request rates. Each service that uses Dynamo runs its own Dynamo instances. CCCCCCCC. MongoDB MongoDB (“humongous”) is a document database designed to be easy to work with, fast, and very scalable. It was also designed to be ideal for website infrastructure. It is perfect for user profiles, sessions, product information, and all forms of Web content (blogs, wikis, comments, messages, and more). It’s not great for transactions or perfect durability, as required by something like a banking system. Good fits for MongoDB include applications with complex objects or real-time reporting requirements, and agile projects where the underlying database schema changes often. MongoDB does not suit software with complex (multiple objects) transactions. 1) Inside MongoDB MongoDB is an interesting combination of modern Web usage semantics and proven database techniques. In some ways MongoDB is closer to MySQL than to other so-called “NoSQL” databases: It has a query optimizer, ad-hoc queries, and a custom network layer. It also lets you organize document into collections, similar to sql tables, for speed, efficiency, and organization. To get great performance and horizontal scalability however, MongoDB gives something up: transactions. MongoDB does not support transactions that span multiple collections. You can do atomic operations on a single object, but you can’t modify objects from two collections atomically. MongoDB stores BSON, essentially a JSON document in an efficient binary representation with more data types. BSON documents readily persist many data structures, including maps, structs, associative arrays, and objects in any dynamic language. Using MongoDB and BSON, you can just store your class data directly in the database. MongoDB is also schema-free. You can add fields whenever you need to without performing an expensive change on your database. Adding fields is also quick, which is ideal for agile environments. You need not revise schemas from staging to production or have scripts ready roll changes back.
Page 190
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IX. ANALYSIS OF NOSQL In this section we compare and analyze the nonfunctional requirements of NOSQL databases named BigTable, Cassandra, CouchDB, Dynamo, MongoDB, HBase. Challenges of NOSQL NOSQL databases face several challenges. Overhead and complexity: NOSQL databases don’t work with SQL; they require manual query programming, which can be fast for simple tasks but time-consuming for others. In addition, complex query programming for the databases can be difficult. Reliability: NOSQL databases do not support ACID properties which are natively supported by relational databases. NOSQL databases thus do not natively offer the degree of reliability that ACID provides. If users want NOSQL databases to apply ACID restraints to a data set, they must perform additional programming. Consistency: Because NOSQL databases do not natively support ACID transactions, they also could compromise consistency, unless manual support is provided. Not providing consistency enables better performance and scalability but is a problem for certain types of applications and transactions, such as those involved in banking. Unfamiliarity with the technology: Most organizations are unfamiliar with NOSQL databases and thus may not feel knowledgeable enough to choose one or even to determine that the approach might be better for their purposes. Limited Eco structure: Unlike commercial relational databases, many open source NOSQL applications do not yet come with customer support or management tools. DB Type
Bigtable Cassandra CouchDB Dynamo
MongoDB
Scalability
Key-Value store DB Column Oriented DB
Highly Scalable
JSON Document Oriented DB Key-Value store DB
Easily scalable and readily extensible Incremental
BSON Document Oriented DB
Scalable
Highly Scalable
scalable. The original intention has been modern web-scale databases. Often more characteristics apply as: schema-free, replication support, easy API, eventually consistency, and more. Hence the misleading term "NoSQL" is now translated to "Not Only SQL (NOSQL)". NOSQL databases generally process data faster than relational databases. Developers usually do not have their NOSQL databases supporting ACID properties, in order to increase performance, but this can cause problems when used for applications that require great precision. NOSQL databases are also often faster because their data models are simpler. “There’s a bit of a trade-off between speed and model complexity, but it is frequently a tradeoff worth making,” Because they do not have all the technical requirements that relational databases have. Major NOSQL systems are flexible enough to better enable developers to use the applications in ways that meet their needs. REFERENCES [43] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E. Gruber, ‘Bigtable: A Distributed Storage System for Structured Data’, OSDI'06: Seventh Symposium on Operating System Design and Implementation, Seattle, WA, November, 2006. [44] Hamilton, James, "Perspectives: One Size Does Not Fit All", 13 November 2009. Lakshman, Avinash; Malik, Prashant, Cassandra - A Decentralized [45] Structured Storage System. Cornell University, 13 November 2009.
Availability
Performance
Highly Available High availability is achieved using replication Highly Available Highly Available High writeavailability
TABLE II NONFUNCTIONAL REQUIREMENTS ANALYSIS OF NOSQL DATABASES
X. CONCLUSION Next Generation Databases mostly address some of the points: being non-relational, distributed, open-source and horizontal
Velammal College of Engineering and Technology, Madurai
High performance Low Loading speeds are better than retrieval speeds Performance at massive scale is one of the biggest challenges Excellent solution for short read
Reliability
Provides reliability at a massive scale. At massive scale is a very big challenge Reliable and efficient system Reliability at massive scale is one of the biggest challenges Low
[46] Chang, Fay; Jeffrey Dean, Sanjay Ghemawat, Wilson C. Hsieh, Deborah A. Wallach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E. Gruber. Bigtable: A Distributed Storage System for Structured Data. Google. 13 November 2009. [47] Kellerman, Jim. "HBase: structured storage of sparse data for Hadoop", 13 November 2009. [48] Ian Eure, Looking to the future with Cassandra. Digg Technology Blog, September 2009
Page 191
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[49] Zhou Wei, Guillaume Pierre and Chi-Hung Chi. CloudTPS: Scalable Transactions for Web Applications in the Cloud. Technical report IR-CS-53, VU University Amsterdam, February 2010. [50] http://www.allthingsdistributed.com/2007/12/eventually_consistent . [51] http://s3.amzonaws.com/AllThings/Distributed/sosp/amazondynamo [52] http://labs.google.com/papers/bigtable.html http://www.cs.brown.edu/~ugur/osfa.pdf [53] [54] http://www.readwriteweb.com/archives/amazon_dynamo.php [55] http://nosql-database.org/ [56] http://couchdb.apache.org/
Velammal College of Engineering and Technology, Madurai
Page 192
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
AN EFFICIENT TURBO CODED ofdm system Prof. Vikas Dhere Hirasugar Institute of technology Nidasoshi ( Belgam ) [email protected]
Abstract: Orthogonal frequency division multiplexing (OFDM) is a promising technique for high data rate wireless communications. In this paper, we propose ‘An Efficient Turbo Coded OFDM system’ that can improve the performance by making all the Turbo coded symbols having the same reliability for OFDM transmission. We perform computer simulations using the Log-maximum-aposteriori (Log-Map) algorithm in order to access the performance of the proposed system. The simulation results show that, the Turbo coded OFDM system gives the best performance compared to conventional coded systems. Keywords: Turbo Codes, Orthogonal Frequency Division Multiplexing (OFDM).
invoking a longer interleaver size. We perform computer simulation using the LOG_MAP algorithm for iterative decoding to access the performance of the proposed system because of its accuracy and simplicity. The rest of the paper is organized as follows: section 2. Describes the OFDM fundamental and OFDM system model used to obtain the performance result presented in this paper. Section 3. Follows with a description of the turbo encoding and decoding and
1. Introduction Error control systems concern themselves with practical ways of achieving very low bit error rates after a transmission over a noisy band limited channel. Several error correcting techniques are deployed and their performances can be measured by comparing them each other and to the theoretical best performance given by Shannon channel information capacity theory. Following the invention of Turbo Codes in 1993, a number of turbo coded systems exploiting the powerful error correction capability have been developed. Many factors affect the performance of Turbo codes. They include the interleaver size and structure, the generator polynomial and constraint length, the decoding algorithm, the number of iterations, and so on. The presence of multipath with different time delays causes inter-symbol interference (ISI) and degrades the performance of the transmission system. To improve performance of the system, complex adaptive equalization techniques are required. Recently, OFDM systems have been attracting much attention, because of its robustness against frequency-selective fading. Some examples of existing systems, where OFDM is used, are digital audio and video broadcasting, asymmetricdigital-subscriber line modems, and wireless local-areanetworks systems, such as IEEE 802.11 and HiperLan/2. In mobile environment, since the radio channel is frequency selective and time-varying, channel estimation is needed at the receiver before demodulation process. Orthogonal Frequency Division Multiplexing (OFDM) could be tracked to 1950’s but it had become very popular at these days, allowing high speeds at wireless communications. Orthogonal frequency division multiplexing (OFDM) is becoming the chosen modulation technique for wireless communications. In this paper, we propose a scheme that can improve greatly the performance of turbo-coded OFDM systems without
2.
OFDM System
2.1
OFDM Fundamentals
Velammal College of Engineering and Technology, Madurai
turbo coded OFDM system. The computer simulation results are shown and discussed in section4. While section 5. Concludes the paper.
The basic principle of OFDM is to split a high rate data stream into a number of lower rate streams that are transmitted simultaneously over a number of sub-carriers. Because the symbol duration increases for the lower rate parallel sub carriers, the relative amount of dispersion in time caused by multi-path delay spread is decreased. Inter symbol interference (ISI) is eliminated almost completely by introducing a guard time in every OFDM symbol. In the guard time, the OFDM symbol is cyclically extended to avoid inter carrier interference.
Fig. 1. OFDM Transmitter and Receiver channel
A basic OFDM system consists of a QAM or PSK modulator/demodulator, a serial to parallel / parallel to serial converter, and an IFFT/FFT module. OFDM is a special case of multi-carrier transmission, In Fig.1. a block diagram of the baseband model of OFDM system is shown. The binary information is first mapped using baseband modulation schemes such as QAM or PSK. Then the serial-to-parallel conversion is applied to baseband modulated signals. The serial-to parallel converted data is modulated using Inverse Fast Fourier Transform (IFFT). This is one of
Page 193
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the main principles of the OFDM systems. After IDFT, the time domain signal is given as following equation;
where N is the length of FFT, S(k) is baseband data sequence. After IFFT, the guard interval called as cyclic prefix is inserted to prevent inter-symbol interference (ISI). This interval should be chosen to be larger than the expected delay spread of the multipath channel. The guard time includes the cyclically extended part of the OFDM symbol in order to eliminate the inter-carrier interference (ICI). The symbol extended with the cyclic prefix is given as follows;
Otherwise
Where: T Symbol length; time between two consecutive OFDM symbols
Where Nc is the length of the cyclic prefix. The resultant signal st(n) will be transmitted over frequency selective time varying fading channel with additive white Gaussian noise (AWGN). The received signal is given by following equation;
Where h(n) is the impulse response of the frequency selective channel and w(n) is AWGN. At the receiver, the cyclic prefix is first removed. Then the signal y(n) without cyclic prefix is applied to FFT block in order to obtain following equation;
After FFT block, assuming there is no ISI, demodulated signal is given by following equation; Where H(k) is FFT[h(n)], I(k) is ICI and W(k) is FFT [w(n)]. 2.2. OFDM Modeling Mathematically, the OFDM signal is expressed as a sum of the prototype pulses shifted in time- and frequency and multiplied by the data symbols. In continuous-time notation, the kth OFDM symbol is written as
( 7) (6)
Velammal College of Engineering and Technology, Madurai
Finally, continuous sequence of transmitted OFDM symbols is
Page 194
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
expressed as
Fig. 3.
(8)
Turbo decoder block diagram
3.2 MAP A map decoder chooses the most likely path so as to maximize the a posteriori probabilities and that can be mathematically presented as:
3. Turbo Encoding and Decoding A ½ rate recursive systematic convolutional code with constraint length three is used as the component of a turbo coding scheme. The encoder employs parallel concatenation as illustrated in Fig. 2. The binary information to be encoded is represented as uk, with ck,1 and ck,2 signifying the parity bits of the first and second component encoders, and where ‘D’ stands for a bit delay. Symbol ‘π’ represents pseudo-random interleaver, which accepts blocks of 8000 information bits.
which is the probability of the codeword a particular sequence
conditioned on that
was received. The quantity
is known as a posteriori for a MAP decoder.
is the joint probability 3.3
Design Criteria
Binary information, uk, is fed to the component encoders in order to generate the associated parity bits, ck,1 and ck,2,which are selected alternately by the puncturing block. In other words, for a binary input sequence of u1, u2, u3, u4, the encoded sequence would be u1, , u2, C1 ,C2, U3,1, u4, c4,2. The fact that the input sequence uk is retained at the output, is due to the ‘systematic’ nature of the encoder.
1. First, the constituent encoders were selected yielding good performance at low signal-to noise ratios, with particular attention to decoding complexity. The final choice was the optimum 4-state recursive systematic encoder. 2. Next, the turbo-code interleaver was designed based on the codeword weight distribution and on the achievable performance on the additive white Gaussian noise (AWGN) channel, using a maximum likelihood approach. 3. Finally, the puncturing schemes were selected based again on both the weight distribution and the achievable performance on the AWGN channel .
3.1. Turbo Decoding
3.4 Turbo Coded OFDM System
The following Fig.3 shows the turbo decoder which contains two component decoders concatenated in parallel, performing the log-MAP algorithm. Let ki is the extrinsic information related to the kth information symbol, rk, provided by the ith decoder, and yk,i is the parity symbol associated with rk, and fed into the ith component decoder.
The block Diagram of proposed Turbo coded OFDM system is as shown in Fig. 4.
Fig. 2. Turbo encoder block diagram
Fig. 4 A block Diagram of Turbo coded OFDM System
Velammal College of Engineering and Technology, Madurai
Page 195
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A generic turbo encoder first encodes the input data sequence . We use the component code with the constraint length K=3 and the generator polynomilas ( 7,5). We chose a code rate of ½ by puncturing two redundancy bits alternatively. After serial-to-parallel conversion for OFDM symbol stacking, each coded symbol is assigned to a distinct orthogonal code with the same length as the number of effective subcarriers to be transmitted. For the single turbo code adaptive OFDM scheme, the information bits are firstly turbo coded by a single turbo code, then modulated separately for each subband. In the same way, at the receiver, the signal from the output of the FFT is demodulated separately then decoded as a single turbo code frame. In the separate turbo code scheme , the OFDM frame is divided into several subbands , and the turbo code frame combines the same subband of several OFDM symbol. In the single turbo code scheme, the OFDM frame is also divided into several subbands, each subband using a different modulation scheme. 4.
4.1.2 Effect of varying number of iterations on performance Frame size = 700 Number of Iterations = 1 Eb/N0(dB)
0
0.2
0.4
0.6
BER
0.1316
0.1519
0.1098
0.1081
Eb/N0(dB)
0.8
1
1.2
1.4
BER
0.0926
0.0719
0.0663
0.0503
Tab. 2. Performance with number of iteration 1 Number of iterations = 4 Eb/N0(dB)
0
0.2
0.4
0.6
BER
0.1616
0.1298
0.0602
0.0321
Eb/N0(dB) BER
0.8 0.0106
1 0.0011
1.2 1.527e-4
1.4 2.6551e4
SIMULATION RESULTS
In this paper, we propose an efficient wireless transmission method that can improve greatly the performance of turbocoded OFDM systems. 4.1 SIMULATION SETUP 4.1.1 Turbo codes We are selecting the AWGN channel. The turbo encoder used in our simulation is of rate ½ by puncturing two redundancy bits alternatively and constraint length K =3 and generator polynomial GR ( D) = ( 7,5). The Log-MAP map algorithm is used for decoding with 08 iterations. The recursive least square algorithm is used for channel estimation. Simulation results for a turbo code are based on bit error rate (BER) performance over a range of Eb/No. The following table shows the rate 1/2 component RSC codes used in the simulation results. The following Tab. 1 shows the configuration used for simulations. Channel Component Encoders RSC Parameters Component Decoder Iterations
Tab.3. Performance with number of iteration 4 Number of Iterations 8 Eb/N0(dB) BER Eb/N0(dB) BER
0 0.1225 0.8 0.0034
0.2 0.0942 1 0.0014
0.4 0.0729 1.2 1.713e-4
0.6 0.0208 1.4 2.7042e4
Tab.4. Number of iterations = 8
Additive White Gaussian( AWGN) 2 Identical Recursive Convolutional codes ((RSC) n=2,k=1,k=2, G 0 = 7, G1 = 5 Log MAP 8
Tab. 1. Simulation parameters Turbo codes
Velammal College of Engineering and Technology, Madurai
Fig. 5 Performance comparison of different iteration times
Page 196
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The results show that as we increase the number of iterations for the same channel, the bit error rate decreases by increasing the signal to noise ratio. 4.2 . OFDM The following Tab.5 shows the configuration used for the simulations performed on the OFDM signal. QPSK 8-PSK 16-PSK FFT size Number of carrier used Guard time Guard Period Type
SNR = 10 dB SNR = 15 dB SNR = 25 dB 2048 800 512 samples Half zero signal , half a cyclic extension of the symbol
= 20 µs . the sample rate is 16 MHz which corresponds to a sample time of Ts= 0.0625 µs. The channel model is AWGN. The recurceive systematic encoder in the Turbo encoder is rate ½ with constraint length K = 3 and a generator polynomial GR ( D) = ( 7,5). In the simulation for both systems, we set the BER target to 10-2. Fig.7 illustrates the BER performance of a Turbo coded OFDM system with the optimal adaptation algorithm. The red solid line with circle in figure shows the BER performance for the separate turbo code system while the blue solid line with square in Fig.6 is the BER performance for the single code system. Both of the systems fulfill the BER target, but the single turbo code system is closer to it.
Tab.5. Simulation Parameters OFDM The following Fig. 6 shows the simulation results and the performance of different modulation scheme in an AWGN channel.
Fig. 7 BER of separate turbo code system and single turbo code system
5. CONCLUSION Red
QPSK
Green Blue
8-PSK 16-PSK
As shown in the numerical results the single turbo code adaptive system provides better performance for longer turbo codes.
Fig. 6 BER curves of different modulation schemes for OFDM system in AWGN channel
6.Refernces.
The results show that the lower order modulation schemes has low bit error rate as compared to higher order modulation at a given SNR.
1..Irving Reed, Xuemin Chen, Error-control coding for Data Networks,Kluwer Academic Publishers, Norwell, Massachusetts, USA, 1999. 2. Hosein Nikopour, Amir K. Khandani and S. Hamidreza Jamali, “Turbo coded OFDM transmission over a nonlinear channel,” IEEE transactions on vehicular technology, vol.54, no.4, (July 2005) 1361-1371.
4.3. TURBO CODED OFDM The simulation system uses Turbo Code –OFDM with N = 256 subcarriers out of which only 16 of them are pilot subcarriers. In the simulated system we use a Cyclic Prefix length of 65 samples , and OFDM symbol duration is TOFDM
Velammal College of Engineering and Technology, Madurai
Page 197
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3. B. Bougard, L.van der Perre, B. Gyselinckx, M. Engels,“Comparison of several FEC schemes for wireless OFDM transmission on frequency selective channels,”. 4. E.R Berlekamap, Algebraic coding Theory, New York. McGraw-Hill 1968. 5. C. Berrou and A. Flavieux, “Near-optimum error correcting coding and decoding. Turbo codes”, IEEE Trans. Common, Vol. 44, PP 1261-1271, Oct. 1996. 6. C. Kikkert, “Digital Communication Systems and their Modulation Techniques”,James Cook University, October 1995. 7. J. G. Proakis, Digital Communications, McGraw-Hill, New York, USA, 4th edition 2000. 8. Pyndiah, R.; Picart, A.; Glavieux, A., "Performance of block turbo coded 16-QAM and 64-QAM modulations," GLOBECOM '95., IEEE , vol.2, no.pp.1039-1043 vol.2, 14-16 Nov 1995 9. S. Le Goff, A. Glavieux and C. Berrou, “Turbo codes and high spectral efficiency modulation,” Proc. of the IEEE International Conference on Communications, vol. 2, pp. 645-649, May 1994.
Velammal College of Engineering and Technology, Madurai
Page 198
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Advancement in mobile technology Using BADA V.AISHWARYA* UG STUDENT DEPT OF ECE KALASALINGAM UNIVERSITY KRISHNAN KOIL [email protected]
J.MANIBHARATHI UG STUDENT DEPT OF ECE KALASALINGAM UNIVERSITY KRISHNAN KOIL [email protected]
Abstract : Bada is a complete smart phone platform. Bada runs high-performance native apps and services. With this use of BADA inour mobile phone the compatibility ,accessibility and the interface of the mobile phone version onescan be considerably increased. The use of Badaallows us to implement mobile with no control technology and speech with gesture recognitionfacility in the mobile devices. Also further the development of BADA allows us to install hand tracking detection in our mobile phones.in this paper ,feasibilities of using Bada in mobile is discussed indetail
Keywords: Software development kit(SDK),Application program interface(API),WII.
1.0 Introduction There were nearly 2000 million cell phones sold each year compared with fewer than 200 million PCs — and the gap was widening. Increasingly, phones were the way people wanted to connect with each other and with everything else. Phones were going to replace PCs as the main gateway to the Internet, and they were going to do it soon. The cell phones ran on different software, had less memory, and operated under the constraints of pay-per-byte wireless networks, the mobile Web was a stripped-down, mimeographed version of the real thing. To avoid this, Bada had the solution. Bada is a free, open source mobile platform that any coder could write for and any handset maker could install It would be a global, open operating system for the wireless future. Bada, (bɑdɑ, the Korean word for “ocean”,) is a new smartphone platform that allows developers to create featurerich applications that elevate the user experience in mobile spaces. The bada platform is kernel-configurable and it can run either on the Linux kernel or real-time OS kernels and it was first developed by Samsung.
1.1Building a better mobile phone
Velammal College of Engineering and Technology, Madurai
Dr.S.DURAI RAJ PROFESSOR & HOD DEPT OF ECE KALASALINGAM UNIVERSITY KRISHNAN KOIL [email protected]
To build a better mobile phone, a mobile must have the ability to record and watch videos with the camcordermode.Uploading videos to YouTube and pictures to Picasa directly from the phone .A new soft keyboard with an "Auto complete" feature. Ability to automatically connect to a Bluetooth headset within a certain distance. New widgets and folders that can populate the desktop. Animations between screens and moving sensors to play WII games .Expanded ability of Copy and paste to include web pages. Ability to embed the ADOBE flash player and webkit internet browser directly. Bada enables developers to take full advantage of mobile device capabilities to easily create compelling applications. Different service applications can share information such as personal profiles, social relations, schedules, and contents with simple user confirmation, all in order to provide services with greater personal relevance and cooperative service extensions. For example, social networking applications can share user information with commerce or location-aware applications, and share photos from other content publishing applications. The bada platform is kernel-configurable so that it can run either on the Linux kernel or real-time OS kernels, which makes bada applicable to a wider range of devices than any other mobile operating system.
2. Bada Runtime Bada run on any platform . Java run timeand dalvik virtual machine and other computing platform. 2.1 Dalvik Virtual Machine Every bada application runs in its own process, with its own instance of the Dalvik virtual machine. Dalvik has been written so that a device can run multiple VMs efficiently. The Dalvik VM executes files in the Dalvik Executable (.dex) format which is optimized for minimal memory footprint. The VM is register-based, and runs classes compiled by a Java language compiler that have been transformed (as ByteCodes, which is good for run fast) into the .dex format by the included "dx" tool.
Page 199
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.2 Linux Kernel Bada relies on Linux version 2.6 for core system services such as security, memory management, process management, network stack, and driver model. It also act as an abstraction layer between the hardware and rest of the software stack.
3.0 Supported Operating Systems 1. Windows XP (32-bit) or Vista (32- or 64-bit)- 1.4 GB RAM or higher . 2. 1.8 GB free disk space or higher. Simulator screen size is 480 x800, Mac os and linux.
3.1 Supported Development Environments 1. Eclipse IDE , Eclipse 3.3 (Europa), 3.4 (Ganymede) . Recommended Eclipse IDE packages: Eclipse IDE for Java EE Developers, Eclipse IDE for Java Developers, Eclipse for RCP/Plug-in Developers. Eclipse JDT plugin. 2. Eclipse Classic IDE package is not supported. 3. 4. JDK 5 or JDK 6 (JRE alone is not sufficient). JAVA FX 1.1 and HELIX. 3.2 Supported web browser environment Internet Explorer 6.0 or higher, Firefox 3.0 or higher and Safari 3.2 or higher.
4.0
Bada Architecture
Bada has a four-layer architecture: kernel, device, service, and framework layers. The kernel layer can be the Linux kernel or a real-time OS kernel, depending on the hardware configuration. The device layer provides the core functions as a device platform, such as system and security management, graphics and windowing system, data protocols and telephony, as well as audio-visual and multimedia management. The service layer provides service-centric functions that are provided by application engines and webservice components interconnecting with bada Server. Only the framework layer can export the C++ open API. The framework consists of an application framework and functions exported by the underlying layers.
5.0 Software Development Kit (SDK) It is a typically a set of development tools that allows a engineer to create applications for a certain software package, software framework, hardware platform, computer system, video game console, operating system or similar platform.
5.2
Basic data types
Object, String, Date Time, Byte Buffer, uid, and other base types. It has Wrapper classes for C++ primitive types, such as Integer, Short, Double, and Long. Its has certain condition for run time. It is interms of timer thread and its Synchronization with Mutex, Semaphore, and Monitor. The other basic data types are Collection and its utility. The SDKfeatures in data types are, ArrayList, HashMap, Stack, and other collection types. Object-based collections and template-based collections.. Math, StringUtil, StringTokenizer, and Uri with Standard library support. C++ STL (Standard Template Library)
5.3 Osp::Io Namespace for input/output data handling. It contains file and Directory. The database involves transaction such as begin, commit and rollback. Its also Support of both SQL queries and enumerations with DB statements. It has registry data IO which involves system-provided data store in which applications store and retrieve settings.
5.4 Osp::Text and Osp::Locales It has namespaces for internationalization and localization. Conversion between the major character encoding schemes, such as UTF8, UCS2, Latin1, GSM, and ASCII and it has separate locale by Identifying a specific language code, country code, and variant code and its Formatting localesensitive information, such as date, number, and currency and it has the ability to Convert a date, time, and number into a string Unicode.
5.5 Osp::System Namespace for controlling devices and getting system information. It has the separate device control with register an alarm. 1. Screen power management: 2. It has Keep LCD on with or without dimming . 3. Vibrator activation with level, on/off period, and count 4. System information: 5. Getting uptime and current time in UTC, Standard, and Wall time mode. The features of sdk in Bada relates with all mobile phone applications like connection of networking which is in the form of Osp::Net, and for multimedia purpose in Osp::Media . It has separate security systems and it is different from all other mobile phones. 6.
5.6 Osp::Security 5.1 Bada SDK features The sdk features is represented in osp: base model. It has several osp for each and every application. There are separate Namespaces for basic types, execution environment, and utilities.
Velammal College of Engineering and Technology, Madurai
It has separate namespace for cryptography keys, algorithms, and certificate management. It has Cryptography keys and PRNG (Pseudo Random Number Generator).The Cryptographic algorithms, such as HMAC, Hash, AES, DES, and RSA occur in basic bada. The X.509 certificate information retrieval and certificate path validation.
Page 200
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6.0 Future SDK Development In Bada The bada system is going to deliver the SDK, which gives compatible to bada powered handsets. The system has fully comfort and include the map external library. The bada system delivers an updated version of the frame work API after the bada s8500 devices. As with previous sensors, the bada has sensor bar with softwares of WII games and also sensor software. The updated API is assigned an integer identifier that is stored in the system itself. It is used to determine whether the application is compatible with the system, prior to installing the applications.
7.0
InBuilt Applications
In the developer application the inbuilt of sensors with software provide the best playment of WIIgames and other 3D games. Some spare parts are software of wii games which is installed through API and sensors to capture from the wii remote. Other applications are as same as the previous bada SDK. The spare parts are the includement of map external library and other external libraries. The software of wii and sensors are transferred through Bluetooth so it brings autopairing effect. The wii sensor is also effective in password security, secret fixing of cameras, effective tool to speech and hearing impaired people . Hacking can be avoided by these sensors. It is applicable to WII 2 games also.
a.
Wii Remote
The main feature of wii remote is the gesture recognition and pointing devices. The gesture recognition detects motion and rotation in three dimension through the use of accelerometer technology. By separating the controller from the gaming console, the accelerometer data can be used as an input for the gesture recognition. The accelerating sensors represented the gestures by characteristic patterns of incoming signals data ie. Vectors, representing the current accelerations of controller in all three dimensions. By filtering and quantizer technique the unwanted vector data can be identified and cleared. This can be done by k-mean algorithm. In this recognition all different size objects, and other images in 3D can be recognized.
b.
Vector Quantization
Like other acceleration-sensors the one integrated into the Wiimote delivers too much vector data to be put into a single HMM. In order to cluster and abstract this data the common kmean algorithm is applied with k being the number of clusters or codes in the so-called codebook. A codebook size delivering satisfying results and Its empirically identified k = 8 for gestures in a two-dimensional plane. However, adopt the idea of arranging the 8 cluster centres on a circle by extending it to the 3D case. Instead of distributing the centres uniformly
Velammal College of Engineering and Technology, Madurai
on a two-dimensional circle its possible to put them on a threedimensional sphere, intersecting two circles orthogonal to each other. Consequently this leads to k = 8 + 6 = 14 centres. The radius of each circle/sphere dynamically adapts itself to the incoming signal data.
9.0 Implementation In Bada The implementation of the gesture recognition in Java using the standardization of Java APIs for Bluetooth Wireless Technology (JABWT) defined by the JSR-82 specification. Using Java ensures platform independency, for developing and testing purposes the use of GNU/Linux platform with the Avetana Bluetooth implementation2 is needed. The recognition process is realized as a reusable and extensible gesture recognition library based on an event-driven design pattern. The library provides an interface for basic functions, e.g. acceleration readout with the Wii Listener interface, as well as recognition functions using a Gesture-Listener interface. Through its modularity it is easy to adapt the prototype to other acceleration-based controllers.
9.1 Sensing The wii remote has the ability to sense acceleration along with three axes through the use of an ADXL 330 accelerometer. The wii remote also has the feature of pix art optical sensors, allowing it to determine where the wii remote is pointing. Unlike a light gun that senses light from a mobile screen, the wii remote senses light from the sensor which is inbuilted in the form of chip in the bada phones allows consistent usage regardless of the mobile sizes. The infrared LED’s in the sensor gives an accurate pointer. The sensor allows the wii remote to be used as an accurate pointing device 30 meters away from the mobile. The wii remote’s image sensor is used to locate the sensor’s points of light in the wii remote’s field of view. The inbuilted wii software in the mobile devices calculates the distance between the wii remote and sensor using triangulation. Games can be programmed to sense whether the image sensor is covered, which is demonstrated in a microgame of smooth moves, where if the player does not uncover the sensor, the champagne bottle that the remote represents will not open.
9.2 Map External Library The map external library is made to add powerful mapping capabilities to this sdk and the good application programme interface add on includes a map external library. The classes of the map library offer built in downloading, rendering, caching of map files as well as the variety of display options and controls. The key class in the maps library is map view and a subclass of view group is in the bada standard library.capture key presses and touch gestures to pan and zoom the map automatically, including handling network requests for additional map files. It also provides all of the UI elements necessary for users to control the map.
Page 201
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
10.0 Summary The implementation of Bada in the mobile phones leads to lot of facilties in each and every applications. The introduction of WII in Bada ensures gesture recognition and its sensing is used as the basic technique for no control technology. Bada has the special feature to have Bluetooth in the standardized Java API, in which no other mobile phones have that features. This analysis lead to the highest cost reduction in WII games if it is implemented.
11.0 References 1. Wii: The Total Story IGN. http://wii.ign.com/launchguide/hardware1.html Retrieved 2006-11-202 2. Consolidated Financial Highlights(PDF). Nintendo. 200907-30. p. 9. http://www.nintendo.co.jp/ir/pdf/2009/090730e.pdf#page=9, Retrieved 2009-10-29. 3 Electronic Arts (2008-01-31). Supplemental Segment Information http://media.corporateir.net/media_files/IROL/88/88189/Q3FY08SupSeg.pdf#page= 4 Retrieved 2008 (PDF). Thomson Financial. p. 4. -02-09.. 4. Wii Remote Colors. news.com. http://news.cnet.com/i/ne/p/2005/0916nintendopic4_500x375.j pg. Retrieved 2006-07-15. 5. http://badawave.com, what is bada, Retrieved 7 April 2010.
Velammal College of Engineering and Technology, Madurai
Page 202
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Mixed-Radix 4-2 Butterfly FFT/IFFT For Wireless communication A.Umasankar, S.Vinayagakarthikeyan EEET- Department, ECE-Department Yanbu Industrial College, Yanbu, Kingdom Of Saudi Arabia, Vel’s Srinivasa College of Engineering and Technology, Chennai, India [email protected] [email protected]
Abstract The technique of orthogonal frequency division multiplexing (OFDM) is famous for its robustness against frequency-selective fading channel. In general, the fast Fourier transform (FFT) and inverse FFT (IFFT) operations are used as the modulation/demodulation kernel in the OFDM systems, and the sizes of FFT/IFFT operations are varied in different applications of OFDM systems. The modified Mixed Radix 4-2 Butterfly FFT with bit reversal for the output sequence derived by index decomposition technique is our suggested VLSI system architecture to design the prototype FFT/IFFT processor for OFDM systems. In this paper the analysis of several FFT algorithms such as radix-2, radix-4 and split radix were designed using VHDL. The results show that the proposed processor architecture can greatly save the area cost while keeping a high-speed processing speed, which may be attractive for many real-time systems. Key words: OFDM, FFT/IFFT, VLSI, VHDL, Mixed Radix with bit reversal.
I.
Introduction
Fast Fourier transform (FFT) are widely used in different areas of applications such as communications, radars, imaging, etc. One of the major concerns for researchers is the enhancement of processing speed. However according to use of portable systems working with limited power supplies, low-power techniques are of great interest in implementation of this block. FFT and FFT blocks are used in OFDM links such as very high speed digital subscribe line (VDSL), Digital Audio Broadcasting (DAB) systems and microwave portable links. Of course there are many ways to measure the complexity and efficiency of an algorithm, and the final assessment depends on both the available technology and the intended application. Normally the number of arithmetic multiplications and additions are used as a measure of computational complexity. Several methods of for computing FFT (and IFFT) are discussed in [1],[2],[3]. These are basic algorithms for implementation of FFT and IFFT blocks. ORTHOGONAL FREQUENCY DIVISION MULTIPLIXING
Velammal College of Engineering and Technology, Madurai
The FDM system above had been able to use a set of subcarriers that were orthogonal to each other; a higher level of spectral efficiency could have been achieved. The guard bands that were necessary to allow individual demodulation of sub carriers in an FDM system would no longer be necessary. The use of orthogonal subcarriers would allow the subcarriers’ spectra to overlap, thus increasing the spectral efficiency. As long as orthogonality is maintained, it is still possible to recover the individual subcarriers’ signals despite their overlapping spectrums. If the dot product of two deterministic signals is equal to zero, these signals are said to be orthogonal to each other. Orthogonality can also be viewed from the standpoint of stochastic processes. If two random processes are uncorrelated, then they are orthogonal. Given the random nature of signals in a communications system, this probabilistic view of orthogonality provides an intuitive understanding of the implications of orthogonality in OFDM. Later in this article, we will discuss how OFDM is implemented in practice using the Fast Fourier transform (FFT) show fig 1. Recall from signals and systems theory that the sinusoids of the FFT form an orthogonal basis set, and a signal in the vector space of the FFT can be represented as a linear combination of the orthogonal sinusoids. One view of the FFT is that the transform essentially correlates its input signal with each of the sinusoidal basis functions. If the input signal has some energy at a certain frequency, there will be a peak in the correlation of the input signal and the basis sinusoid that is at that corresponding frequency. This transform is used at the OFDM transmitter to map an input signal onto a set of orthogonal subcarriers, i.e., the orthogonal basis functions of the DFT. Similarly, the transform is used again at the OFDM receiver to process the received subcarriers show fig 2. The signals from the subcarriers are then combined to form an estimate of the source signal from the transmitter. The orthogonal and uncorrelated nature of the subcarriers is exploited in OFDM with powerful results. Since the basis functions of the FFT are uncorrelated, the correlation performed in the FFT for a given subcarrier only sees energy for that corresponding
Page 203
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
subcarrier. The energy from other subcarriers does not contribute because it is uncorrelated. This separation of signal energy is the reason that the OFDM subcarriers’ spectrums can overlap without causing interference.
have different radices. For instance, a 64-point long FFT can be computed in two stages using one stage with radix-8 PEs, followed by a stage of radix-2 PEs. This adds a bit of complexity to the algorithm compared to radix-r, but in return it gives more options in choosing the transform length. The Mixed-Radix FFT algorithm is based on sub-transform modules with highly optimized small length FFT which are combined to create large FFT. However, this algorithm does not offer the simple bit reversing for ordering the output sequences. Mixed-Radix FFT Algorithms with Bit Reversing The mixed-radix 4/2 [9] butterfly unit is shown in Figure 2. It uses both the radix-2^2 and the radix-2 algorithms can perform fast FFT computations and can process FFTs that are not power of four. The mixed-radix 4/2, which calculates four butterfly outputs based on X (0) ~X (3). The proposed butterfly unit [10], [11] has three complex multipliers and eight complex adders. Four multiplexers represented by the solid box are used to select either the radix-4 calculation or the radix-2 calculation.
Fig.1 OFDM Transmitter and Receiver
II. FAST FOURIER TRANSFORM The Discrete Fourier Transfer (DFT) plays an important role in many applications of digital signal processing including linear filtering, correlation analysis and spectrum analysis etc. The DFT is defined as:
(1) Fig 2: The basic butterfly for mixed-radix 4/2 DIF FFT algorithm.
is the DFT coefficient. Where Evaluating the Equation (1) directly requires N complex multiplications and (N-1) complex additions for each value of the DFT. To compute all N values therefore requires a total of N^2 complex multiplications and N(N-1) complex additions. Since the amount of computation, and thus the computation time, is approximately proportional to N2, it will cost a long computation time for large values of N. For this reason, it is very important to reduce the number of multiplications and additions. This algorithm is an efficient algorithm to compute the DFT [4],[5], which is called Fast Fourier Transform (FFT) algorithm or radix-2 FFT algorithm, and it reduce the computational complexity from O(N²) to O(N log₂(N ). Mixed Radix 4-2 A mixed radix algorithm is a combination of different radix-r algorithms. That is, different stages in the FFT computation
Velammal College of Engineering and Technology, Madurai
In order to verify the proposed scheme, 64-points FFT based on the proposed Mixed-Radix 4-2 butterfly with simple bit reversing for ordering the output sequences is exampled. As shown in the Figure 2, the block diagram for 64-points FFT is composed of total six-teen Mixed-Radix 4-2 Butterflies. In the first stage, the 64 point input sequences are divided by the 8 groups which correspond to n3=0, n3=1, n3=2, n3=3, n3=4, n3=5, n3=6, n3=7 respectively. Each group is input sequence for each Mixed-Radix 4-2 Butterfly. After the input sequences pass the first Mixed-Radix 4-2 Butterfly stage, the order of output value is expressed with small number below each butterfly output line in the figure 3. The proposed Mixed-Radix 4-2 is composed of two radix-4 butterflies and four radix-2 butterflies. In the first stage, the input data of two radix-4 butterflies which are expressed with the equation B4 (o, n3, kj) B4 (i, n3, k1), are grouped with the x(n3),
Page 204
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
x(N/4±n3), x(N/2±n3), x(3N/4±n3) and x(N/ 8±n3), x(3N/8±n3), x(5N/8±n3), x(7N/8±n3) respectively. After the each input group data passes the first radix-4 butterflies, the outputted data is multiplied by the special twiddle factors. Then, these outputted sequences are inputted into the second stage which is composed of the radix-2 butterflies. After passing the second radix-2 butterflies, the outputted data are multiplied by the twiddle factors. These twiddle factors WQ (1+k) is the unique multiplier unit in the proposed MixedRadix 4-2 Butterfly with simple bit reversing the output sequences. Finally, we can also show order of the output sequences Fig above. The order of the output sequence is 0,4,2,6,1,5,3 and 7 which are exactly same at the simple binary bit reversing of the pure radix butterfly structure. Consequently, proposed mixed radix 4-2 butterfly with simple bit reversing output sequence include two radix 4 butterflies, four radix 2 butterflies, one multiplier unit and additional shift unit for special twiddle factors.
Fig.3 Proposed Mixed-Radix 4-2 Butterfly for 64 point FFT
III.
RESULT
Employing the parametric nature of this core, the OFDM block is synthesized on one of Xilinx’s Virtex-II Pro FPGAs with different configurations. The results of logic synthesis for 64 point FFT based orthogonal frequency division multiplexing (OFDM) using Radix-2, Radix-4, split Radix and mixed radix 4-2 are presented in Table 1. We analysis the 64-point FFT based OFDM is chosen to compare the number of CLB slices and Flip Flop for different FFT architectures .
Velammal College of Engineering and Technology, Madurai
Table 1: Comparison of FFT Algorithm based on CLB Slices, Function Generators.
IV.
DFF, and
Conclusion
In this paper, we design an OFDM for different algorithms implemented in OFDM modem are identified. It was found during the algorithm design that many blocks need complex multipliers and adders and therefore special attention needs to be given to optimize these circuits and maximize reusability. In particular, the models have been applied to analyze the performance of mixed-radix FFT architectures used in OFDM. Actual hardware resource requirements were also presented and simulation results were given for the synthesized design. The 64-point Mixed Radix FFT based OFDM architecture was found to have a good balance between its performance and its hardware requirements and is therefore suitable for use in OFDM systems. V. References [1] Shousheng. He and Mats Torkelson, "Design and Implementation of a 1024-point Pipeline FFT Processor", IEEE Custom Integrated Circuits Conference, May. 1998, pp. 131-134. [2] Shousheng He and Mats Torkelson, "Designing Pipeline FFT Processor for OFDM (de)Modulation", IEEE Signals, Systems, and Electronics, Sep. 1998, pp. 257-262. [3] Shousheng He and Mats Torkelson, "A New Approach to Pipeline FFT Processor", IEEE Parallel Processing Symposium, April. 1996, pp.776-780. [4] C. Sidney Burrus, "Index Mapping for Multidimensional Formulation of the DFT and Convolution", IEEE Trans. Acoust., Speech, and Signal Processing, Vol. ASSP-25, June. 1977, pp. 239-242. [5] Lihong Jia, Yonghong GAO, Jouni Isoaho, and Hannu Tenhunen, "A New VLSI-Oriented FFT Algorithm and Implementation", IEEE ASIC Conf., Sep. 1998, pp. 337-341.
Page 205
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[6] Martin Vetterli and Pierre Duhamel, "Split-Radix Algorithms for Length-ptmDFT's", IEEE Trans. Acoust, Speech, and Signal Processing, Vol. 37, No. 1, Jan. 1989, pp. 57-64. [7] Daisuke Takahashi, "An Extended Split-Radix FFT Algorithm", IEEE Signal Processing Letters, Vol. 8, No. 5, May. 2001, pp. 145-147.
Velammal College of Engineering and Technology, Madurai
[8]Y.T. Lin, P.Y. Tsai, and T.D. Chiueh, "Low-power variable-length fast Fourier transform processor", IEE Proc Comput. Digit. Tech, Vol. 152, No. 4, July. 2005, pp. 499506..
Page 206
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Towards Energy Efficient Protocols For Wireless Body Area Networks Shajahan Kutty#1, Laxminarayana J.A*2 #
Departmen of E&TC Engineering, Goa College of Engineering Farmagudi, Ponda, Goa , India 1
[email protected]
Department of Computer Engineering, Goa College of Engineering Farmagudi, Ponda, Goa , India [email protected]
Abstract The major medium access control (MAC) layer protocol design challenges for wireless body area network (WBAN) involve Quality of Service assurance for high reliability and guaranteed latency requirement for real time data, especially vital signs that needs a deterministic structure with special care for emergency reaction alarm; flexibility since it must support various types (periodic, non-periodic, medical, entertainment) of traffic/data rate and importantly, energy efficiency, since energy consumption especially for implanted device must be limited. These requirements necessitate design of efficient active/inactive scheduling. This paper evaluates the needs and performance of energy efficient MAC protocols proposed for wireless body area networks. Keywords- body sensor nodes; BAN co-ordinator node; beacon; preamble; superframe.
Introduction The major sources of energy wastage in a typical sensor network are due to consumptions occurring in three domains, viz; sensing, data processing and communication [1]. Of these, the energy communication wastage is a significant contributor and occurs due to idle listening as dominant factor in most applications, collision of packets, overhearing and control packet overhead. Mainly design of energy efficient MAC protocols target a reduction or elimination of the energy communication waste in particular idle listening. The central approach to reduce idle listening is through duty cycling. Suitable wake-up mechanisms can save significant amount of energy in Wireless Body Area Networks (WBAN) and increase the network lifetime. In these schemes, the device wake-ups only when necessary, otherwise it sleeps thereby saving energy. Coordinated and controlled data transmission can therefore reduce energy consumption. Data traffic in a WBAN is classified into [2]:
Velammal College of Engineering and Technology, Madurai
Normal traffic: Based on normal operation between device and coordinator using a defined pattern. On-demand traffic: Initiated by Coordinator to know certain information. Emergency traffic: In case of critical condition.
Fig (1) Data traffic in a WBAN The various MAC issues that WBAN has to contend with are [3]: Heterogeneous traffic: due to normal, on-demand and emergency traffic, Interoperability due to multiple frequency bands and correspondingly multiple PHY techniques and connecting different devices working on different bands/PHYs Scalability: due to variable data rate, kbps to Mbps and variable number of devices The options for implementing energy saving mechanisms involve choosing between synchronous or asynchronous transmission, beacon or preamble and different periods of wake up and sleep. The three main approaches adopted are: 1. Low power listening (LPL) 2. Scheduled contention 3. Time Division Multiple Access (TDMA)
Page 207
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Low power listening (LPL) involves channel polling where nodes wake up very briefly to check channel activity without receiving data. If the channel is idle, the node goes back to sleep, otherwise it stays awake to receive data. This action is performed regularly but not synchronised among nodes. To rendezvous with receivers, senders send a long preamble before each message to intersect with polling . Since the technique is asynchronous it is sensitive to tuning for neighbourhood size and traffic rate. The result is poor performance when traffic rates vary greatly and thus it is effectively optimised for known periodic traffic. As can be seen, the receiver and polling efficiency is gained at the much greater cost of senders. Scheduled contention involves schedule coordinated transmission and listen periods. Here the nodes adopt common schedule, synchronised with periodic control messages. The receiver listens to brief contention periods while senders contend. Only nodes participating in data transfer remain awake after contention periods while others can sleep. Since the method is synchronous it is sensitive to clock drift. However, improved performance with traffic increase results in similar cost incurred by sender and receiver. As seen the technique is scalable, adaptive and flexible. TDMA is a contention-free/cluster based method where the base station (BS)/cluster head (CH) allocates time slots to every node. Only one node is allowed to transmit in a slot. Timing and synchronisation provided by BS/CH. Normally this requires nodes to form clusters with one node as CH/BS Communication is only between nodes and cluster head; no peer to peer communication exists. The advantage is low duty cycle and fine grained time synchronisation. However the system is very sensitive to clock drift and results in limited throughput. Moreover, it requires clustering which means the cost incurred more on cluster head. Limited scalability and adaptivity to changes on number of nodes are the other issues with this scheme. A WBAN consists of low-power invasive and non-invasive Body Sensor Nodes (BSN). BSNs can be full functional, eg when dealing with multiple PHYs or reduced functional that are mostly common in the in-body networks [4]. One WBAN device is selected as a BAN Network Co-ordinator (BNC).
Traffic between BSNs and the BNC and On-Demand Traffic between the BNC with BSNs and BSNs with BSNs. The BNC can wake-up all the time and can support normal, emergency, and on-demand traffics. On the battery power supply side it has certain limitations and should adopt lowpower scavenging techniques. It should also calculate its own wakeup pattern based on the BSN’s wake-up patterns. This necessitates maintenance of the pattern-based wake-up table. The BSN devices operate on limited power and support a default normal wake-up state. They wake-up and sleep according to a pattern-based wake-up table. The BSN wakesup upon receiving an ‘on demand’ request from a BNC. The BSN also wakes-ups by itself to handle emergency events. .
Fig (2) Traffic characteristic in a WBAN MAC LAYER FRAME FORMAT A typical frame structure includes the advertisement which specifies the synch, interval, address, the CAP specifying reservation and non-periodic data, Beacon with synch, length of slot, reservation status announcement, DTP (Data Transmit Period), DTS (Data Transmit Slot) Continuous, Periodic data, ETS (Emergency Data Transmit Slot) Emergency data & Periodic data Advertisement
Velammal College of Engineering and Technology, Madurai
DTS
ETS ( CAP )
Batch Ack
( Optional ) Inactive Period Or CAP Extension (Optional )
CAP
The traffic characteristic in a WBAN varies from low to high with periodic or non-periodic data and vice versa. The dynamic natures of BSNs does not urge synchronized periodic wake-up periods. Communication flows include Normal Traffic between BSNs and the BNC, Emergency
Beacon
TDMA Data Transmit Period
Fig (3) A typical BAN superframe (BSF)
Page 208
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
BANs may be coordinated most of the time. The coordinator can allocate a negotiated bandwidth to QoS demanding nodes (data rate, BER, latency) with a beaconenabled mode including TDMA and Slotted-ALOHA (for UWB) with relaying. The inherent advantage of TDMA architecture for power saving over non-TDMA structure is that the coordinator goes into inactive mode if no activity is scheduled after CAP or DTP Sensor (continuous data). Once DTS reservation is done, the node stays inactive period except for beacon and its reserved transmit time. Sensors (routine data) wakeup only for report and transmit during CAP or CAP extension whichever available first. Sensors (Alarm) wakeup only for report and transmit during CAP, CAP extension or ETS whichever is available first. Some applications imply a lower channel load. A beacon-free mode is more efficient. NETWORK ARCHITECTURE THE PRIMARY REQUIREMENT FOR WBAN IS TO COPE WITH LOW-POWER EMISSION AND SEVERE NON LINE OF SIGHT (NLOS) CONDITIONS MAINTAINING HIGH RELIABILITY WITH HIGH LEVEL OF QOS FOR CRITICAL / VITAL TRAFFIC FLOWS [5]. A SUITABLE ARCHITECTURE IS THE MESH NETWORK CENTRALIZED ON THE GATEWAY. A FULL MESH
The Beacon-based MAC mode Superframe structure is based on 802.15.4 with an inclusion of a control portion for management messages. The control portion is large enough to allow dynamic changes of topology. The CAP is minimized, mostly used for association using slotted ALOHA Control portion
Beacon period
Request period
Topo mgmt period
Data portion
CAP
CFP
Inactive
Fig (5) Beacon-based MAC mode Superframe structure In the Fine structure the superframe is divided equally into slots. Use of minislots in the control portion provides flexibility and adaptation to different frame durations. Guaranteed Time MiniSlot (GTMS) is introduced in CFP
TOPOLOGY BASED ON A SCHEDULING TREE WITH GUARANTEED ACCESS FOR MANAGEMENT AND DATA MESSAGES (REAL TDMA) IS PROPOSED.
111111111 000000000 000000000 111111111 Preamble
Packet source
DATA
111111 0 1 0 1 000000 1111111111111111111111111111 0000000000000000000000000000 111111 0 1 0 1 000000 TCI
Node #1
Node #i
1 0 0 1
11 00 00 11
1111111 0000000 0000000 1111111
0 1 0 1 000 111 0 1 0 1 000 111 0000000000000000000000000000 1111111111111111111111111111 0 1 0 1 000 111 0 1 0 1 1 0
Node #N
Receive mode
Transmit mode
Fig (4) Mesh network and scheduling tree structures
Velammal College of Engineering and Technology, Madurai
Fig (6) Fine superframe structure In the Control portion structure beacon period, the beacons are relayed along the scheduling tree. The beacon-frame length is minimized and the beacon alignment procedure is used During the request period, a set of GTS dedicated to allocation demands. In the topology management period, Hello frames, for advanced link state procedure are introduced with a scheduling tree based update Beacon-free MAC mode is used in situations that might not require a beacon-enabled MAC protocol. Examples are a symmetric or asymmetric network, low communication rate and small packets, network set-up, coordinator disappearance, etc. In such conditions, the downlink is an issue : nodes must be powered-up to receive data from the coordinator (e.g. polling/by-invitation MAC scheme) A Preamble Sampling MAC protocol for Ultrawideband (UWB) involves nodes that periodically listen to the channel. If it is clear, they go back to sleep; conversely, they keep on listening until data. Nodes are not synchronized but share the wake up period (TCI) and a packet is transmitted with a preamble as long as TCI
Page 209
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
70 60
Fig (7) The Preamble Sampling Scheme Preamble sampling protocols are known to be the most energy efficient uncoordinated MAC protocols since TCI depends on the traffic load and the TRX consumption (TX/RX/listen modes). This can be of the order of 100 ms. The preamble is a network specific sequence, either a typical wake-up signal or a preamble with modulated fields (e.g. @, time left to data)
Rate (%)
50 40 30 20 10
Prioritized Backoff with Slotted Aloha Slotted Aloha
0 1
2
3
4
5
Number of Node per Class
PERFORMANCE EVALUATION A.. Beacon-free MAC mode : The simulation is carried out for three setups viz ; Mode when only one burst per device is emitted, Mode 1 when burst is re-emitted by a device, under the LDC limit, until a neighbor relays and Mode 2 which is same as Mode 1 plus former relays, with emission credits left, can relay again.
Fig(9) Average number of acknowledged packets Simulation Parameters: Simulation time: 60 sec, Inter-arrival time: 1pkt/sec (Poisson), Packet size: 16byte, BO=6, SO=3, CAP size = 8 slot, 1 to 5 node per each class DISCUSSION AND CONCLUSION The responsibilities of the upper layers are to enable/disable beacons switching procedure as follows: (i) from the beacon-free to beacon-based mode, BAN formation (coordinator election), new coordinator election if former coordinator leaves that, for example, can be triggered by a user requesting a link with high QoS (ii) from the beacon-based to beacon-free mode which is the fallback mode if the coordinator leaves or if the required BW (rate of GTS requests) is below a given threshold and requested QoS is adequate. A combination of beacon-based true TDMA mode including fast relaying and mesh support and a beacon-free mode using preamble sampling and extra features adapted to UWB is suitable for WBAN systems
Fig (8) PDP vs Wake up period plots for the three modes B. Access Policy in CAP for UWB PHY: Since CCA or carrier sense is not feasible with UWB PHY Aloha type of access is the only possible solution. While Aloha is a non-beacon mode, Slotted Aloha (Beacon mode) is better than Aloha but still the probability of collision is still high. Slotted Aloha with Prioritized Back-off is a new scheme.
Velammal College of Engineering and Technology, Madurai
References [1] J.S Yoon, Gahng S. Ahn, Myung J Lee, Seong-soon Joo “Versatile MAC for BAN proposal responding to TG6 Call for Proposals (15-08-0811-03-0006-tg6 call for proposals” [2] A. El-Hoiydi and J. D. Decotignie, “WiseMAC: An Ultra Low-Power MAC Protocol for Multi-hop Wireless Sensor Networks”, Proceeding of First International Workshop on Algorithmic Aspects of Wireless Sensor Networks (ALGOSENSOR), July 2004. [3] E. A. Lin, J. M. Rabaey, and A. Wolisz, “PowerEfficient Rendezvous Schemes for Dense Wireless Sensor Networks”, Proceedings of ICC, June 2004. TICER/RICER [4] The French ANR "BANET" project [5] M. J. Miller and N. H. Vaidya, “A MAC Protocol to Reduce Sensor Network Energy Consumption Using a Wakeup Radio”, IEEE Transactions on Mobile Computing, 4, 3, May/June 2005. [6] 15-08-0644-09-0006-tg6-technical-requirementsdocument.doc [7] Timmons, N.F. Scanlon, W.G. "Analysis of the performance of IEEE 802.15.4 for medical sensor body area
Page 210
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
networking", IEEE SECON 2004, Santa Clara, Ca, October 2004 [8] European ICT SMART-Net project deliverables [9] Jean Schwoerer , Laurent Ouvry , Arnaud Tonnerre, “PHY and MAC proposal based on UWB impulse radio for the IEEE 802.15.6 CFP-Response to IEEE 802.15.6 call for proposals” [10] J. Polastre, J. Hill and D. Culler, “Versatile Low-Power Media Access for Wireless Sensor Networks”, in Proc. of ACM Sensys, November 2004. B-MAC [11] E. Shih, P. Bahl, and M. J. Sinclair, “Wake On Wireless: An Event Driven Energy Saving Strategy for Battery Operated Devices”, Proceedings of ACM MobiCom, September 2002. [12] HyungSoo Lee, Jaehwan Kim, et al; “ Merged Baseline Proposal for TG6 -response to IEEE 802.15.6 call for proposals”
Velammal College of Engineering and Technology, Madurai
Page 211
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The Medical Image Segmentation R.Kalaivani M.phil (final year) Department of Computer Science Dr.N.G.P Arts and Science College Coimbatore ABSTRACT The Medical Image Segmentation is developed to extract the suspicious region from the patient image. The proposed system design aim is to give second opinion for the radiologist in tumor deduction. Image Acquisition is used to collect the image from a specified location or database, then the image data is converted to digital data or digital matrix. Image Preprocessing removes the labels and unwanted areas from the image using tracking algorithm. Image Enhancement covers the study related the removal of film artifacts and noises from the image using filters. Image Registration correlates the source and target image at same position and place. It also crop unwanted image from the original image by using the point mapping technique. Image Blocking blocks existing registered source and target images. Image Comparison compares the sources and average intensity of each and every block of the source and target image. Image segmentation the suspicious region is segmented from the target image if comparing the difference from source and target image. KEYWORDS Image acquisition, Image preprocessing Image enhancement, Image registration, Image blocking, Image comparison, Image segmentation, Performance evolution.
1. INTRODUCTION In the clinical context, medical image processing is generally equated to radiology or "clinical imaging" and the medical practitioner responsible for interpreting (and sometimes acquiring) the images are a radiologist. Diagnostic radiography designates the technical aspects of medical imaging and in particular the acquisition of medical images. The radiographer is usually responsible for acquiring medical images of diagnostic quality, although some radiological Medical imaging is often perceived to designate the set of techniques that no invasively produce images of the internal aspect of the body. In this restricted sense, medical imaging can be seen as the solution of mathematical inverse problems. This means that cause (the properties of living tissue) is inferred from effect (the observed signal). In the case of ultra sonography the probe consists of ultrasonic pressure waves and echoes inside the tissue show the internal structure. In the case of projection radiography, the probe is X-ray radiation
Velammal College of Engineering and Technology, Madurai
Guided by, Hemalatha, Lecturer Department of Computer Science Dr N.G.P Arts and Science College Coimbatore which is absorbed at different rates in different tissue types such as bone, muscle and fat. Medical Image diagnosis is considered as one of the fields taking advantage of high-technology and modern instrumentation. Ultra sound, CT, MRI, PET etc., have played an important role in diagnosis, research, and treatment. Image registration is a process of overlaying two or more images of the same scene taken at different times, from different viewpoints, and/or by different sensors. The present differences between images are introduced due to different imaging conditions. 2. METHOD Assuming that we have two images of the same Object, a structural image and a functional image. The Process of registration is composed of following steps: • Collect the image from a specified location or database, then the image data is converted to digital data or digital matrix. • Removes the labels and unwanted areas from the image using tracking algorithm. • Removal of film artifacts and noises from the image using filters. • Image Registration correlates the source and target image at same position and place. • Crop unwanted image from the original image by using the point mapping technique. • Blocks existing registered source and target images. • Compares the sources and average intensity of each and every block of the source and target image • The suspicious region is segmented from the target image if comparing the difference from source and target image. 2.1 Image Acquisition Image Capturing is a process to acquire the digital image into the Mat lab. In this module medical image was given as input and it is the base input to the whole project. All types of medical images can be acquired in this module. Images of a patient obtained
Page 212
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
are displayed as an array of pixels (a 2D unit based on the matrix size and the field of view) and stored in memory. In mat lab there are various formats of image encoding Binary Images Large matrix whose entries are one of the two values will specify a black and white image.0 corresponds to black, 1 to white. Grey Scale/Intensity Images A grey scale image can be specified by giving a large matrix whose entries are numbers between 0 and 1.A black and white image can also be specified by giving a large matrix with integer entries. The lowest entry corresponds to black, the highest to white. Fig shows typical Image Captured in Mat lab from patient database
For removing the unwanted portions of the image we have developed an algorithm named Tracking Algorithm. Fig shows a typical preprocessed image after the removal of film artifacts
2.3 Image Enhancement Image enhancement methods inquire about how to improve the visual appearance of images from Magnetic Resonance Image (MRI), Computed Tomography (CT) scan; Positron Emission Tomography (PET) and the contrast enhancing brain volumes were linearly aligned. Fig : Shows the Median Filter output images
2.2 Image Pre-processing Preprocessing indicates that the same tissue type may have a different scale of signal intensities for different images. Processing functions involve those operations that are normally required prior to the main data analysis and Extraction of information, and are generally grouped as geometric corrections. Radiometric corrections include correcting the data for sensor irregularities and unwanted sensor or atmospheric noise, removal of non-brain voxels and converting the data so they accurately represent the reflected or emitted radiation measured by the sensor. Tracking Algorithm In preprocessing module image acquired will be processed for correct output. Medical images surely will have some Film Artifacts like labels and marks which are detected and removed for better result.Pre-processing was done by using tracking algorithm. If these images are too noisy or blurred, they should be filtered and sharpened.
Velammal College of Engineering and Technology, Madurai
(a).3*3filtered image (b).5*5 filtered image
(c). 11*11 filtered image Fig : shows the enhanced output
Page 213
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.4 Image Registration Image registration is the process of transforming the different sets of data into one coordinate system. Registration is necessary in order to be able to compare or integrate the data obtained from different measurements. In medical imaging (e.g. for data of the same patient taken at different points in time or by different angles) registration often additionally involves elastic (or non rigid) registration to cope with elastic deformations of the body parts imaged. Non rigid registration of medical images can also be used to register a patient's data to an anatomical atlas, such as the Talairach atlas for neuro-imaging. Algorithm Classifications •
Area-based vs. Feature-based.
•
Transformation model.
•
Search-based vs. direct methods.
•
Image nature.
•
Other classifications.
2.5.1 Block Based Method In Block based method, both the original image and the sensed image has been divided into several blocks. Then block wise subtraction has been done between the two images. This subtracted value is then checked with the threshold value, Block Based method Image. Fig shows the original image in blocks
Fig : shows the tumor image in blocks
Fig shows the image before Correlation 2.5 Image Blocking(Correlation) Here, we have taken two reference points first, in front view and second in the top view of the image. The enhanced image has to be resized to the original image size by fixing the same reference points as in the original image. Since in our technique, the size of the original image is 256*256, the enhanced image has been resized to (256-x)*(256-y) by removing the extra portions in the image. Fig shows the Correlated image
Velammal College of Engineering and Technology, Madurai
3. And
Conclusions future work
Relevance of these techniques is the direct clinical application for registration. This provides a diagnosis and preoperative planning surgery. Alternatively, it is a useful tool in image processing training. Algorithms of finding a transformation between two images is of importance in the methodology, spatial and density resolution have to be achieved to reduce geometric errors In this survey paper various automatic detection methods of brain tumor through MRI has been studied and compared for the period of one year. This is used to focus on the future of developments of medical image processing in medicine and health care. We have described several methods in medical
Page 214
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
image processing and discussed requirements and properties of techniques in brain tumor detection. This paper are used to give more information about brain tumor detection and registration. In this paper various step in detection of automatic system (a) Image capturing (b) Image preprocessing (c) Image Enhancement (d) Image Registration. The program needs to be compiled to run in an independent environment and algorithms need to be optimized inorder to reduce the time consuming and performing and increase in accuracy. The program could be developed in ability of access to host computer or hospital via intranet to get patients information and images. 4. References 1. Amini L, Soltanian-Zadeh H, Lucas’s,:“Automated Segmentation of Brain Structure from MRI”, Proc. Intl.Soc.Mag.Reson.Med.11 (2003). 2. BrainTumors: Gd-DTPA-Enhanced, Emodynamic, and Diffusion MR Imaging Compared with MRSpectroscopic Imaging”, Neuroradiol 23, February 2002. 3. Azadeh yazdan-shahmorad, Hesamoddin Jahanian, Suresh Patel, Hammed Soltanian- Zadeh,:”Automatic Brain Tumor Segmentation .Using Diffusivity Characteristics”, IEEE, ISBI,2007. 4. Boada F.E, Davis.D, Walter.K, Torres- Trejo.A, Kondziolka.D, Bartynski.W, Lieberman. “Triple Quantum Filtered Sodium MRI of Primary Brain Tumors, “IEEE, USA, 2004. 5. Dimitris N Metaxas, Zhen Qian, Xiaolei Huang and Rui Huang,Ting Chameleon Axal.:”Hybrid Deformable Models for Medical Segmentation and Registration, “IEEE Transactions on Medical Image processing, ICARCV ,2006. 6 Gilbert Vezina:”MR Imaging of Brain Tumors- Recent Developments”, Childhood Brain Tumor Foundation, Germantown, Washington. Evaluating Image Segmentation Algorithm”, IEEE, Medical Image Processing. 7 Hideki yamamoto and Katsuhiko Sugita, Noriki Kanzaki, Ikuo Johan and Yoshio Hiraki,Michiyoshi Kuwahara , : “Magnetic Resonance Image Enhancement Using VFilter”, IEEE AES Magazine, June 1990.
Velammal College of Engineering and Technology, Madurai
Page 215
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Selection of a Checkpoint Interval in Coordinated Checkpointing Protocol for Fault Tolerant Open MPI #
Mallikarjuna Shastry P.M.#1, K. Venkatesh #2
Research Center, Department of Computer Science & Engg., M. S. Ramaiah Institute of Technology, M.S.R Nagar, Bangalore-54, Karnataka, India. 1
[email protected] 2 [email protected]
Abstract — The goal of this paper is to address the selection of efficient checkpoint interval which reduces the total overhead cost due to the checkpointing and restarting of the applications in a distributed system environment. Coordinated checkpointing rollback recovery protocol is used for making the application programs fault tolerant on a stand-alone system under no load conditions using BLCR and OPEN MPI at system level. We have presented an experimental study in which we have determined the optimum checkpoint interval and we have used it to compare the performance of coordinated checkpointing protocol using two types of checkpointing intervals namely fixed and incremental checkpoint intervals. We measured the checkpoint cost, rollback cost and total cost of overheads caused by the above two methods of checkpointing intervals Failures are simulated using the Poisson distribution with one failure per hour and the inter arrival time between the failures follow exponential distribution. We have observed from the results that, rollback overhead and total cost of overheads due to checkpointing the application are very high in incremental checkpoint interval method than in fixed checkpoint interval method. Hence, we conclude that fixed checkpointing interval method is more efficient as it reduces the rollback overhead and also total cost of overheads considerably.
Keywords: Checkpoint, Checkpoint Interval, Fault tolerance, Marker, Checkpoint Overheads.
I. INTRODUCTION Since, the recent trends in HPC and even stand alone systems employ a very large number of processors to execute the large size application programs in a distributed system environment, it is required to provide the fault tolerance to such applications.
Velammal College of Engineering and Technology, Madurai
As the complexity of the program increases, the number of processors to be added to the cluster / HPC / Super Computer also increases which in turn decreases the MTBF (mean time between failures) of the processors or the machines. It means that the probability of failure of one or more processors will be very high before the completion of the execution of the long running application being executed parallely on several processors. When a processor fails, we need to restart the entire application on all the processors from the beginning. Hence, it is required to address the issues like the scalability and fault tolerance. Fault tolerance provides the reliability and availability to the large size applications programs executed in a distributed system environment. Fault tolerance is achieved using coordinated checkpointing rollback recovery protocol in which an initiator takes a checkpoint by synchronizing with all other processes of MPI application [1]. For MPI applications, a cluster consisting of a group of processes interacting with each other is formed and each individual process in the cluster is checkpointed and a global state is formed out of it. The global state contains the “set of checkpoints exactly one from each processor”. The global state is consistent if and only if for each message received by a processor (receiver), there is a corresponding sender. The latest consistent global state is known as the recovery line [2]. The checkpoint / restart scheme has been widely used in [3]-[9] to address the failure of execution of an application. Checkpoints can be taken using either fixed checkpoint interval or variable checkpoint interval [10]. In case of
Page 216
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
fixed checkpoint interval, checkpoint interval size remains same between any two successive checkpoints. But, in case of the incremental checkpoint interval method discussed in this paper, the second checkpoint interval size is 2 times the 1st one and third checkpoint interval is 3 times the 1st one and so on in each cycle A cycle is the execution time interval of the application between two successive failures. Since, these two methods of checkpoint intervals are not compared in the literature as we understand; we have carried out an experiment to determine the behaviour of the coordinated checkpointing protocol using the fixed and incremental checkpoint interval methods. The rest of the paper is organized as follows. Section 2 introduces the related works carried out in checkpoint and restart schemes using different checkpoint intervals. Section 3 presents the different notations used in this paper. The implementation of the coordinated checkpointing protocol, determination of optimum checkpoint interval, complete description of fixed and incremental checkpoint intervals are discussed in section 4. Computation of cost of overheads is discussed in section 5. Section 6 presents the experimental setup and the results. Section 7 presents the conclusion.
3. R - restart time required to resume the execution of an application from the most recent checkpoint. 4. F - the number of failures during the execution of the application. 5. TS - time required to save each checkpoint on to a local disk. 6. Ni - the number of checkpoints taken in ith cycle. 7. tc - checkpoint interval without restart cost. 8. TC - optimum checkpoint interval size and is used as fixed checkpoint interval. 9. TCi– ith checkpoint interval which is incremental. 10. CCi - the cost of checkpoints in ith cycle. 11. CC – total cost of checkpoints 12. P - the number of processes / processors used for parallelism. 13. λ – Number of failures per hour. 14. TF – time to failure. 15. Ti - the time at which the ith failure occurs. IV. IMPLEMENTATION OF COORDINATED CHECKPOINTING PROTOCOL A. Protocol
II. RELATED WORKS
Master MPI process with rank i=0 takes the tentative checkpoint and then sends the marker to MPI process with rank (i+1) % N. When MPI process i > 0 receives the marker from (i + N-1) % N, takes its tentative checkpoint and sends the marker to MPI process with rank (i + 1) % N.
Young [5] has presented an optimum checkpoint and restart model and has shown that the total waste time due to checkpointing can be reduced using fixed checkpoint interval. But, this model [5] does not consider the restart time required to resume the application from the most recent checkpoint after a failure.
When the MPI process with rank 0 receives the marker from MPI process N-1, a global consistent checkpoint is formed out of all the local checkpoints and then sends the checkpoint file to the local disk and then initiates the next checkpoint cycle after the checkpoint interval as specified by the user.
An optimal checkpoint / restart model presented by Yudun liu [11] uses varying checkpoint interval with different failure distributions. But, varying checkpoint interval does not yield optimal rollback and checkpoint cost. R. Geist et. Al [12] discusses the selection of checkpoint interval in a critical task environment, but it does not present any optimal solution for selecting the checkpoint interval.
The checkpoint period can be either fixed or incremental which will be discussed in the following subsections.
J.T. Daly [6], [9] presents a method for determining the optimum checkpoint interval but they do not discuss the comparison of the performance of the coordinated checkpointing protocol with respect to fixed and incremental checkpointing interval methods. III. NOTATIONS USED 1. 2.
Rbi - the cost of rollback in ith cycle. Rb - total rollback cost.
Velammal College of Engineering and Technology, Madurai
B. Computing Optimal Checkpoint Interval Knowing the number of processors (P) used for computation and the failure rate (λ) of the processors, we can compute the time to failure [14] of the application during run time as follows.
TF = 1 / (P λ )
(1 )
Once, we determine the time to failure TF and the checkpoint overhead TS (time required to save each checkpoint onto local disk), the checkpoint interval tC can be computed [ 5] as follows.
Page 217
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
tC =
2 TS TF
(2)
i
Ci = ∑ TCK + (i − 1) TS
(5)
k =1
But, the above equation does not consider the restart time of the application after a failure. Hence, considering the restart time (R) also, the above equation can be modified by adding the restart time (R) multiplied by the number of failures (F) to tC as follows to get the optimal checkpoint interval TC.
TC =
2 TS T F + R F
(3)
And the ith checkpoint interval is computed as follows.
TC i = i TC
( 6)
C1
C2 TC1
TS
CNi TC2
TS
TS
Fig. 2 Incremental Checkpoint Interval
C. Fixed Checkpoint Interval The total execution time of the application program is divided into n checkpoint intervals of length TC as computed in the previous section. The first checkpoint is initiated by the master MPI process after completion of Tc minutes of execution of the application program and second checkpoint is initiated after completion of 2 TC + TS minutes and so on as shown in figure 1. In general, the starting time of computed as follows.
ith checkpoint Ci is
Ci = i TC +(i −1) TS
(4)
V. COST OF OVERHEADS DUE TO CHECKPOINTING AND RESTARTING During recovery from failure, the master MPI process coordinates with all the other MPI processes and restarts by rolling back the application to the most recent consistent global checkpoint. The cost of rollback, cost of checkpointing and the restart cost are the 3 components which are used to determine the waste time in each cycle of the application. If the application undergoes F failures, the execution of the application will have F cycles. We present in the following sections the determination of these costs using two different methods of checkpoint intervals such as fixed and incremental checkpoint intervals.
But, the length of each checkpoint interval is fixed.
A. Cost of Overheads in Fixed Checkpoint Interval
i.e., TC1 = TC, TC2 = Tc, TC3 = Tc and so on.
Failure of a fault tolerant application using fixed checkpoint interval is shown in figure 3. Since, all the checkpoint intervals have same length; the number of checkpoints to be taken in ith cycle (before ith failure occurs) is computed as follows.
In general, TCN = TC C1 TC
TS
TC
C2
CNi
TS
TS
TC
Fig. 1 Fixed Checkpoint Interval
D. Incremental Checkpoint Interval Figure 2 shows the checkpointing of the application with an incremental checkpoint interval. In this case, in each cycle, the first checkpoint is initiated after (Tc1) or TC minutes of execution of the application and second checkpoint is initiated after TC1 + TC2 + TS minutes and the third checkpoint is initiated after TC1 + TC2 + TC3 + 2TS and so on as shown in figure 2. In general the starting time of ith checkpoint Ci is computed as follows.
Velammal College of Engineering and Technology, Madurai
N i = ⎣Ti / (TC + TS )⎦
(7 )
Then, the cost of checkpoint in ith cycle is computed as follows.
CC i = N i TS
(8)
The cost of rollback in ith cycle is then computed as follows
Rbi = (Ti − N i (TC + TS ))
(9)
The time lost in ith cycle TLi due to a failure can be obtained by adding checkpoint cost, rollback cost, restart cost together as follows.
Page 218
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TLi = CC i + Rbi + R
(10)
The total waste time due to F failures is then computed as follows. F
TL = ∑ TLi
(11)
i =1
Suppose, if the first failure occurs at 40 minutes of execution and checkpoint interval size TC is 4 minutes, time to save a checkpoint is 35 seconds and number of checkpoints taken before failure occurs is 8, The different overhead costs are determined as follows. i) Rollback cost: By applying equation (8), we can determine the rollback cost as follows.
ii) Checkpoint cost = 8 checkpoints * cost of each checkpoint = 6 * 35 = 280 seconds iii) It was found from the experimental setup that the time required to restart an application after a failure is just about 24 seconds. So, the total time lost due to a failure of application in fixed checkpoint interval case is TL1=cost of rollback (200 seconds) + cost of checkpoints (280 seconds) + restart cost (24 seconds) = 504 seconds (about 21 % of execution time of 1st cycle is wasted due to checkpointing, rollback and restart). It was observed that, the cost of rollback is dependent on the amount of time elapsed since the last checkpoint. Time required to restart
Restart
Application Failure C2 TC
TS
CNi TC
TS
a. .if T1 < (TC1 + TS ), b . .if (
n
∑K k =1
(12)
T C + n T S ) < T i and
Ti < ( for
N1 = 0
n +1
∑K k =1
n = 1, 2 , 3 , ..
T C + ( n + 1) T s ) and
for
(13 )
i = 1, 2 , 3 , .. F
The number of checkpoints to be taken in ith cycle can be computed as follows.
Rbi = (2400 – 8 * (240 +35)) = 200 seconds
C1
Failure of a fault tolerant application using incremental checkpoint interval is shown in figure 4. In this, method, size of first checkpoint interval is TC minutes and size of other checkpoint intervals is TC minutes more than its previous checkpoint interval in each cycle. Hence, the number of checkpoints (Ni) to be taken in ith cycle will vary in this method and it is computed as follows.
TS
TC
Execution time till a failure (Ti ) Cycle i Fig. 3 Cost of Overheads in Fixed Checkpoint Interval
B. Cost of Overheads in Incremental Checkpoint Interval
Velammal College of Engineering and Technology, Madurai
Ni = n
(14)
Though, the checkpoint interval length keeps increasing, from one checkpoint to another checkpoint, the per checkpoint cost remains same as per the experimental results that we have obtained. So, the total checkpoint cost in ith cycle can be computed as follows.
CCi = Ni TS
(15)
The cost of rollback in ith cycle is then computed as follows. n
Rbi = (Ti − ( N i TS + ∑ K TC ))
(16)
k =1
The time lost in ith cycle TLi due to a failure can be obtained by adding checkpoint cost, rollback cost, and restart cost together using equation (9). Total waste time due to F failures is computed using equation (10). It was observed that, the cost of rollback depends on two factors like the time of failure and the checkpoint interval size. This is because, in this method, the checkpoint interval size varies from one checkpoint to another checkpoint. Suppose, if the first failure occurs at 40 minutes of execution and initial checkpoint interval TC is 4 minutes,2nd checkpoint interval is 8 minutes and 3rd checkpoint interval is 12 minutes and the 4th checkpoint interval is 16 minutes (during which the failure occurs), time to save a checkpoint is 35 seconds and number of checkpoints taken before failure occurs is 3 after applying the equation (12) and (13). Then, the different overhead costs are determined as follows.
Page 219
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
i ) rollback cost : after applying the equation (15), we get Rbi =(2400 – (105 + 1440 )) = 855 seconds ii) checkpoint cost is = 3 checkpoints * cost of each checkpoint=3*35 = 105 seconds iii) and it was found from the experimental setup that the time required to restart an application after a failure is just about 24 seconds. So, the total time lost due to a failure of application in incremental checkpoint interval case is TL1 =cost of rollback + cost of checkpoints + restart cost = 984 seconds (41 % of execution time of 1st cycle is wasted due to checkpointing, rollback and restart). Time required to restart Restart Application Failure C1 TC1
TS
C2 TC2
During the restart or recovery state, the MPI application rolls back to the most recent checkpoint as discussed and resumes the execution of the application from that point. BLCR checkpoint and restart library [13] is used to implement the blocking coordinated checkpointing protocol to checkpoint the application. The application was run 10 times for different number of processors varying from 1 to 10. We observed that the checkpoint cost and the restart cost increase linearly with the increase in the number of processors. The results obtained for 10 processors are presented in this paper. The number of arrivals N (t) in a finite interval of length t obeys the Poisson (גt) distribution.
P { N (t ) = n} = (λ t ) n e − λt / n ! The inter arrival times of the failures are independent and obey the exponential distribution.
f ( x )
= λ e = 0,
− λ x
for
x >=
0
otherwise
CNi
TS
TS
TCNi
Execution time till a failure (Ti) Cycle i Fig. 4 Cost of Overheads in Incremental Checkpoint Interval
VI. EXPERIMENTAL SET-UP AND RESULTS We have taken the application program that multiplies 2 integer matrices of size 7000 * 7000. The above application is written in C language and run on a standalone system using scattered method of MPI under no load conditions. In scattered method, one of the matrices, say the matrix B is broadcasted across all the processors using MPI_Bcast(). The matrix A is divided equally among the number of processors used for parallelism and each of the processors gets only a portion of matrix A allocated for it for the computation using MPI_Scatter(). The above application was run on a system with 6 GB of RAM, Intel ® Core ™ 2 Duo CPU,E7200 @ 2.53 GHz and 110 GB of HDD and the execution time of the application considered in our experimental setup on this system is 67 minutes without checkpointing. The monitor program is written in a shell script which runs at the background and keeps monitoring whether the MPI processes grouped under mpirun are running or not. Once, monitor program learns that an MPI process has failed, it calls the restart() routine of BLCR to restart the application.
Velammal College of Engineering and Technology, Madurai
We have used Poisson distribution with 1 failure per hour (λ = 1) for 10 processors and generated the probability distributions for the inter arrival times of failures. Failures are simulated using these probability distributions and the results are presented in the form of graphs. Figures 5a and 5b present the comparison of total cost of all the overheads due to checkpointing and restarting of the application with different checkpoint interval sizes when the application considered in our experimental setup fails at different timings. We have tested for 10 different failures (whose failure timings are generated using Poisson arrival process) and determined the total cost of overheads incurred due to checkpointing and restarting using 5 different checkpoint intervals of size 2,3,4,5 and 10 minutes as shown in figures 5a and 5b. In 7 different cases out of 10 failures at different timings, checkpoint interval size with 4 minutes was found to be optimal as it yields minimum total cost of all the overheads as shown in figures 5a and 5b when the application fails at 10,15,20,42,47,50 and 60 minutes of execution of the application. From our experiment, we determined that, per checkpoint cost TS is 35 seconds for 10 processors and it remains same for all the other checkpoints taken at different timings for the same number of processors. Per checkpoint cost is determined by taking the average value of checkpointing the application at 20 different timings on 10 processors. Restart cost is found to be only 24 seconds to resume the execution of the application on 10 processors after a failure occurs. This restart cost is determined by taking the average of 20 different restart costs measured when the application failed at different timings.
Page 220
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Restart cost R is added only once because, we have considered only one failure i.e F = 1 during the execution of our application. This checkpoint interval value is almost matching with the optimum checkpoint interval of 4 minutes obtained from the figures 5a and 5b based on our experimental results. The value of TC calculated should yield almost the exact result, if the value of TF is quite large in which case R << TF. Hence, the equations (1), (2) and (3) are validated based on our experimental results and discussion. As, we have obtained 4 minutes as the optimum checkpoint interval size from our experimental analysis, in our further analysis and discussion (figures 6 to 9), fixed checkpoint interval size taken is 4 minutes and in incremental checkpoint interval method, the first checkpoint interval size taken is 4 minutes, second checkpoint interval size taken is 8 minutes and third checkpoint interval size is 12 minutes and so on. We have presented the results in the form of graphs for one failure in an hour with λ = 1 using Poisson distribution for arrival of failures. Figure 6 presents the comparison of number of checkpoints taken in fixed and incremental checkpoint interval methods. Figures 7, 8, and 9 present the comparison of i ) cost of checkpoints, ii ) cost of rollback and iii) total cost of overheads caused by fixed and incremental checkpoint interval methods respectively.
2 Min.
34
4 Min.
5 Min.
Total Cost (Sec)
500 400 300 200
10 Min.
42 47 50 Failure Time (Min)
60
Incremental
Fixed 14 12 10 8 6 4 2 0
5 10 15 20 25 30 35 40 45 50 55 60 Failure Time (Min) Fig 6. Comparison of Number of Checkpoints of Fixed and Incremental Checkpoint Intervals
Comparison of Checkpoint Cost Fixed
Incremental
500 400 300 200 100 0 5
100
5 Min.
Comparison of Number of Checkpoints
10 Min.
600
4 Min.
Fig 5b. Comparison of Total Cost of Overheads with Different Checkpoint Interval Size.
Checkpoint Cost (Sec)
3 Min.
3 Min.
900 800 700 600 500 400 300 200 100 0
Comparison of Total Cost of Overheads with Different Checkpoint Interval Size 2 Min. 700
Different
Comparison of total cost of overheads with different Checkpoint Interval Size
Total Cost (Sec)
TF = 1 / (P * λ) = 1 / ( 10 * (1/60)) minutes = 6 minutes and TC = sqrt (2 TS TF ) + R F = sqrt( 2 * (35 /60) * 6) + (24 seconds * 1) = 3.2 minutes
Fig 5a. Comparison of Total Cost of Overheads with Checkpoint Interval Size.
Number of Checkpoints
We have used the equations (1), (2) and (3) to determine the optimum checkpoint interval size when the time to failure TF and checkpoint cost TS and the restart cost R are known. The value of TC obtained from equation (3) shows that the optimum checkpoint interval size is 3.2 minutes as shown below.
10 15 20
25 30 35 40 45 50 55 60
Failure Time ( Min )
0 10
12
15
20
30 Fig 7. Comparison of Checkpoint Cost of Fixed and Incremental Checkpoint Intervals
Velammal College of Engineering and Technology, Madurai
Page 221
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Comparison of Rollback Cost Fixed
Incremental
Rollback Cost ( Sec)
1200 1000 800 600 400 200 0 5
10 15 20 25 30 35 40 45 50 55 60 Failure time (Min)
Fig 8. Comparison of Rollback Cost of Fixed and Incremental Checkpoint Intervals
Comparison of Total Cost 1400
Fixed
Incremental
Total Cost (Sec)
1200 1000 800 600 400 200 0 5 10 15 20 25 30 35 40 45 50 55 60 Failure Time (Min)
Fig 9. Comparison of Total Cost of overheads caused by Fixed and Incremental Checkpoint Intervals
VII. CONCLUSIONS Figures 5a and 5b show the comparison of total cost of overheads with different checkpoint interval size. From figures 5a and 5b, it is clear that the total cost of overheads is quite minimum when checkpoint interval size is 4 minutes. We have even validated the model developed by Young [5] to determine the optimum checkpoint interval. But, the model [5] does not consider the restart cost required to restart the application when it fails. We have added the restart cost multiplied by the number of times the application undergoes failures to Young’s model [5] to get the optimum checkpoint interval which yields the optimum overheads cost irrespective of number of failures for any application. An approximate estimate of the checkpoint interval can be calculated from equation (3). From figure 6, we see that the fixed checkpoint interval method causes more number of checkpoints than
Velammal College of Engineering and Technology, Madurai
incremental checkpoint interval method. So, the checkpoint cost is also quite high in fixed checkpoint interval method as compared to the incremental checkpoint interval method when the application fails after first checkpoint as shown in figure 7. But, the rollback cost and the total cost of overheads produced by fixed checkpoint interval are quite low as compared to the incremental checkpoint interval when the application fails after first checkpoint as shown in figure 8 and figure 9 respectively. Fixed checkpoint interval reduces more than 50% of total overhead cost as compared to the incremental checkpoint interval. Hence, we conclude that using fixed checkpoint interval for checkpointing an application would be more advantageous than using incremental checkpoint interval because fixed checkpoint interval reduces both rollback cost and the total cost of overheads significantly. ACKNOWLEDGEMENT The authors thank the Head of Research Centre, CSE dept. and head of ISE dept M.S Ramaiah Institute of Technology, Bangalore for their constant encouragement. The Computer system on which the experimental analysis has been carried out is acquired under the project sanctioned by BRNS, INDIA, bearing sanction No. 2008/37/15/BRNS. REFERENCES [1] Luis Moura Silva and Joao Gabriel Silva, “The Performance of Coordinated and Independent Checkpointing”, IEEE Trans, 1999. [2] G.E. Fagg, A. Bukovsky and J.J. Dongarra, “Harness and Fault Tolerant MPI”, Parallel Computing, 27(11):1479-1495, 2001. [3] K.M. Chandy, “A survey of analytic models of roll-back and recovery strategies,” Computer 8, 5 (May 1975), 40-47. [4] K.M. Chandy, J.C. Browne, C. W. Dissly, and W. R. Uhrig, “Analytic models for rollback and recovery stratagems in data base systems,” IEEE Trans Software Engg. SE-1, ( March 1975), 100-110 [5] J.W. Young, “A first order approximation to the optimum checkpoint interval,” Communications of ACM 17, 9(Sept 1974), 530531. [6] J.T.Daly, “A Higher Order Estimate of the Optimum Checkpoint Interval for Restart Dumps,” Future Generation Computer Systems [Elsevier ], Amsterdam, 2004. [7] E. Elnozahy, J. Plank, “Checkpointing for Peta Scale Systems: A Look into the Future of Practical Rollback-Recovery,” IEEE Trans. Dependable Sec. Comput.1 (2):97-108(2004). [8] M.Treaster, “A survey of fault-tolerance and fault-recovery techniques in parallel systems, “Technical Report cs.DC / 0501002, ACM computing Research Repository (CoRR), January 2005. [9] J. T. Daly, “A Model for Predicting the Optimum Checkpoint Interval for Restart Dumps,” ICCS 2003, LNCS 2660, Proceedings 4 (2003) 3-12. [10] Yudun Liu, Raja Nassar, Chokchai (box) Leangsuksum, Nichamon Naksinehaboon, Mihaels Paun, Stephen L. Scott, “An Optimal Checkpoint /Restart Model for a Large Scale High Performance Computing System,” IEEE Trans. 2008. [11] Y. Liu, “Reliability Aware Optimal Checkpoint / Restart Model in High Performance Computing, PhD Thesis,” Louisiana Tech university, Ruston, LA, USA, May-2007.
Page 222
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[12] R. Geist, R. Reynolds, and J. Westall,” Selection of a checkpoint interval in a critical-task environment,” IEEE Trans. Reliability, 37, (4), 395-400 (1988). [13] H. Paul Hargrove and C. Jason Duell, “Berkeley lab checkpoint / restart (BLCR) for Linux clusters”, Journal of Physics, Conference series 46 (2006), 494-499, SciDAC 2006. [14] James S. Plank and MichG.Thomason, “The Average Availability of Parallel Checkpointing Systems and Its Importance in th Selecting Runtime Parameters”,29 Internatioonal symposium on Fault Tolerant Computing , Madison WI, Jun-1999, pg 250-259.
Velammal College of Engineering and Technology, Madurai
Page 223
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Simplification Of Diagnosing Disease Through Microscopic Images Of Blood Cells T.Benazir Fathima#1 #1,2
K.V.Gayathiri Devi #2
M.Arunachalam*3
M.K.Hema*4
Final year Computer Science Engineering , K.L.N.College of Information Technology. [email protected],[email protected]
*3,4
Computer Science Engineering Department, K.L.N.College of Information Technology Pottapalayam,Sivagangai dt,TamilNadu,India [email protected],[email protected]
Abstract— This paper is the implementation of a simple, fast and reliable method for automatically diagnosing diseases through digitized images of blood. Hematopathology reveals that there is an intrusion of the disease cell in the blood having identical characteristics for each disease. Principal effectors in the blood that are Erythrocyte, Leukocyte and Platelets of the blood play a crucial role in supplying blood to the various parts of the body, resisting foreign particles and in clotting of blood respectively. The effectors are identified and skipped for our observation. In microscopic images of the diseased blood cells, the diagnosis is based on the evaluation of some general features of the blood cells such as color, shape, and border, and the presence and the aspect of characteristic structures. Perception of these structures depends both on magnification and image resolution. The introduction of vision system associated with image processing techniques enables the numerical and objective description of some pattern cells features. Matlab image processing software is used for the experimental work. A new algorithm is introduced which will efficiently reduce the overall running time of object recognition. Keywords - Medical imaging,BPN,Pattern matching
I.INTRODUCTION This project is the implementation of a simple, fast and reliable method for automatically diagnosing diseases through digitized images of blood. Hematopathology reveals that there is an intrusion of the disease cell in the blood having identical characteristics for each disease. Principal effectors in the blood that are Erythrocyte, Leukocyte and Platelets of the blood play a crucial role in supplying blood to the various parts of the body, resisting foreign particles and in clotting of blood respectively. The effectors are identified and skipped for our observation. In microscopic images of the diseased blood cells, the diagnosis is based on the evaluation of some general features of the blood cells such as color, shape, and border, and the presence and the aspect of characteristic structures. Perception of these structures depends both on magnification and image resolution. The introduction of
Velammal College of Engineering and Technology, Madurai
vision system associated with image processing techniques enables the numerical and objective description of some pattern cells features. Matlab image processing software is used for the experimental work. A new algorithm is introduced which will efficiently reduce the overall running time of object recognition. Concept of Back Propagation network is used to set appropriate error values during object recognition. Experimental results using proposed technique gives a way for new research area in Bio mechanics. Nevertheless, this would be an interesting future investigation, since our approach is sufficiently general to be applied to modeling tasks from other two dimensional or three dimensional application domains. The extension of this may open a new area of research in bio mechanics (modeling and simulation). II.IMPLEMENTATION A.EXPERIMENTAL SETUP The technology used in solid state imaging sensors is based principally on charged coupled devices (CCD), which can convert the images from analog to digital format. Simple strategies can be adapted to grab the images from the Electron Microscope. The experiment apparatus consists of structured light source, Electron Microscope, CCD camera and vision system. The relative cells of the blood are first measured. Before the actual diagnosis is made, the real coordinates and world coordinates must be determined in the camera coordinate system by a calibration process. The camera acquisition requires the knowledge of calibration analysis. The camera can be viewed simply as a device for mapping a 3D image into a 2D image. A detailed experimental study is carried out in diseased cell structure with the normal one. The real time experiment is to illustrate the actual implementation of the estimation method developed by spatial encoding technique. The innovative hardware and vision-processing algorithm with menu driven user interface was used to analyze the condition of the blood cells in the diseased cell structure. The experiment is planned taking into consideration of the
Page 224
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
above requirements and detection of diseases through microscopic images of the diseased cell structure. The software facilities capturing of high resolution images in color or in black and white, comparison of built up model (Template) with that of subjected model and the images can be sized and stored for future reference. The last stage of the experimental setup in which only one of the diseases would be activated from the output layer of the back propagation algorithm denoting the disease that is to be determined.
image is made to fix into the standard dimensions as shown below.For example INPUT SLIDE SIZE STANDARDIZATION
Î
Color/ Grey Image
Image capturing
(Reduced)
Î (Enlarged) Fig2.1.1inputslidesize Standardization
Segmentation
Binarized Image
Analysis (BP Algorithm)
• Color/Gray image The image is then passed to the above component from which the color image is converted to gray scale using the necessary formula. Weights of the Red, Blue and Green are mixed in proper proportions and then we calculate a single value as required in gray scale. • Segmentation The image is then segmented according to the required values of height and width. When there is some exceeding part of the block they are filled with appropriate values of padding. •
Disease Detection
Feed back to control (Vision System) Figure 2.1 Experimental Setup •
Image capturing
The blood strip taken from the patient’s body is placed in the slide. This is viewed through the electron microscope. A frame is retrieved using the frame grabber. From the frame the region of interest is extracted according to the height and width in number of pixels as required. Thus the
Velammal College of Engineering and Technology, Madurai
Binarized Image
The Image is then binarized into 0 and 1 using necessary approximations.For example terms with values 0.65 are approximated to 1 and terms with 0.35 are considered as 0. •
Back propagation algorithm
Back propagation algorithm is explained in detail later which possesses certain formula so that an image that almost resembles a particular image of pattern is considered to be that pattern. For example if the difference between the two images is negligible or fixed to a certain error rate is considered to have the same pattern. Comparisons are made repeatedly so that the inputted slide is found to either pattern matched or not pattern matched. •
Disease detection
Page 225
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The last stage of the experimental setup in which only one of the diseases would be activated from the output layer of the back propagation algorithm denoting the disease that is to be determined.
Chronic Myeloid Leukemia
Gaucher's Disease
Hypersplenism
Fig 2.2 Slides Containing Blood Cells B. BLOOD: THE VITAL FLUID Blood is a red fluid pumped by the heart to various parts of the body through the circulatory system. No other artificial fluid can replace the function of blood. Blood contains hemoglobin which makes the blood appear red in color. Blood contains 45% of solid content and the remaining liquid. Components of blood Blood contains plasma that is 12% consisting of the solid particles and the remaining liquid. The three major components of blood is • Erythrocytes • Leukocytes • Platelets as shown in the below figure.
organs of the human body. The size of red blood cell is 7.5 microns in diameter. It is spherical in shape. In a human body weighing 75 kg containing 5 litres of blood can have about 1000 red blood cells in it. Red blood cells are almost same in diameter. • Leukocytes Leukocytes are also termed as White blood cells. The function of Leukocytes is to fight against the foreign organisms or the cells getting into the blood due to the various diseases. When compared to Erythrocytes, Leukocytes are large in number. The main way in differencing Erythrocytes and Leukocytes is the nucleus that is only possessed by the Leukocytes. The size of Leukocytes is also bigger in diameter compared to that of Erythrocytes. • Platelets Platelets play a vital role in performing the clotting of blood. Platelets do not have a definite structure. Digital Image Processing Digital image processing is the processing of the image data for storage, transmission and representation for autonomous machine perception. A digital image is composed of a finite number of elements each of which has a particular location and value. These elements are referred to as picture elements, image elements and pixels. “Pixels” is the term most widely used to denote the elements of the digital image. Vision is the most advanced of our senses, so it is not surprising that images play the single most important role in human perception. Thus digital image processing encompasses a wide and varied field of applications. There are fields such as computer vision whose ultimate goal is to use computers to emulate human vision, including learning and being able to make inferences and take actions based on visual inputs. This area is itself a branch of artificial intelligence whose objective is to emulate human intelligence. The area of image analysis also called as “Image understanding” is in between image processing and computer vision. Lets us now discuss about Medical Imaging that is part of Digital image processing.
C.Medical Imaging
Fig 2.2.1
Components Of Blood
• Erythrocytes Erythrocytes are also termed as Red blood cells. The function of Erythrocytes is to carry oxygen to various
Velammal College of Engineering and Technology, Madurai
The parts of the digital image processing that are exclusively used in medical related fields define Medical Imaging. There are numerous applications of the Digital image processing concepts used in medical fields like that of Gamma ray and X-ray imaging. D.Neural network concepts involved The architecture of the human brain is significantly different from the architecture of a conventional computer. The conventional computers are not suited to perform pattern recognitions problems. We therefore borrow features from physiology of the brain as the basis for our new process models. Hence the technology has come to be known as Artificial Neural Systems technology or simple Neural Networks. Thus we
Page 226
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
are going to a concept involved in the Neural Networks termed as Back Propagation networks that plays a vital role in our project in the pattern recognition problem. Back propagation network Back propagation network can be used to solve complex pattern matching problems. The network learns a predefined sets of input/output example pairs by using a two phase propagate adapt cycle. After an input pattern has been applied as a stimulus to the first layer of the network units, it is propagated through each upper layer until an output is generated. This output pattern in then compared with the desired output and an error signal is computed for each output unit. In our project the input is shown in table containing the input slides and the output is the table containing the diseased cell pictures. The significance of the process is that as the network trains the nodes in the intermediate layers organize themselves such that different nodes learn to recognize different features of the total input space. After training when presented with the arbitrary input pattern that is noisy or incomplete, the units is the hidden layer of the network will respond with the active output with the new input contains the pattern that resembles the feature the individual units learn to recognize during training. The error term is calculated using the formula: M Ep = (1/2) ∑δ2pk k=1 where δpk is the error term for the output
activated indicated by 1. Else if the error difference is exceeding Ep then the output is not activated and is indicated by 0. As we know in the Back Propagation network we fix the error value. Using the error value or the threshold we could determine whether the inputted image matched the already trained image. If the difference is below the error or the threshold value then it is obvious that the inputted image is one among the trained sample of the blood. From the figures in the above table we could find the following • When the obtained image is heavily blurred exceeding the error value and thus the result is 0. • When the obtained image is less blurred exceeding the error value and thus the result is 0. E. PATTERN MATCHING Basic principle of pattern matching: Lets consider the following case in which the first matrix in the top to be the slide containing only the diseased cell. Let the second matrix be considered as the slide retrieved from the frame grabber containing the diseased cell. As shown below we can observe that the first matrix is present in the second containing the values 53, 44, 67, 55. In ordinary or classic methods of pattern matching we would obviously think of solving the problem by sliding the first matrix with the second.
units. The above error value Ep is set so that the difference between the obtained and the actual image is fixed to a certain quantity. Result analysis of Error differences by BPN Actual image
Obtained image
Result
0
1
Fig 2.4.1 Result Analysis Of Error Differences By BPN Thus only if the obtained image has an error difference within the allowed error term that is Ep the output is
Velammal College of Engineering and Technology, Madurai
53 67 12 22 54 22
44 55 11 33 76 44
13 53 67 88
14 44 55 55
So the step involved in the normal case is as below: For i= 1 to m2 For j= 1 to n2 For k=1 to m1 For l= 1 to n1 { Processing statements } Thus the running time is as below: Runtime= O(m2*n2*m1*n1) Where ‘O‘denotes the BIG O notation representing the lower bound of any algorithm.
Page 227
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Enhanced pattern matching We now find a method from which we could obviously reduce the running time of the pattern matching involved. The steps involved is described as below • No of pixels of a sample and the slide for height and width is collected for every sample • The first element of the sample is compared with each and every element of the slide • Comparison is done along with the error value of the BPN already prefixed • If at all the first element is equal (considering the error value) to a element then the remaining m2 elements in that row are found for match • If a row is matched a counter is set. • Now comparison is done simply by skipping m1 value thus saving runtime. COMPONENTS OF BLOOD • Pattern is matched if the counter value equals to n2. • Depending upon the matched values that exceeds the threshold value gives a true result. Thus the Enhanced pattern matching analysis is done as below. Running time=O(m1*m2) that is comparatively less to O(m2*n2*m1*n1) Thus time saved =O(m2*n2*m1*n1) - O(m1*m2) Pattern matching using selected points: Cell of Chronic Myeloid Leukemia
Fig 2.5.3 Bloc k Extr actio n
In orde r to still reduce the running time of the algorithm we can select some points may be in some fashion and find out appropriate match. This is represented in the following diagrams. This is represented as below: Considering Slide dimensions as below the total no of comparisons required are 100X100 =10000 TABLE 2.5 IMPROVED ALGORITHM ANALYSIS Block dimensions
NO of Percentage of Total no of blocks running time comparisons required in saved
Velammal College of Engineering and Technology, Madurai
minimum
20X20
8
3200
68
10X10
16
1600
84
5X5
32
800
92
The above is graphically represented below:
Graph Representation 100 90 80 70 60 % 50 40 30 20 10 0
Time saved
50X50 20X20 10X10
5X5
Dimensions
Inappropriate
Less accuracy
Fig2. 5 Graphical Representation Of Algorithm Analysis Thus improved running time considering dimensions 10X10 = 16% [O(m1*m2)] Thus total time saved [Improved Technique] O(m2*n2*m1*n1) - 16%[O(m1*m2)] that is considerably more to that of O(m2*n2*m1*n1) - O(m1*m2) III.RESULTS AND DISCUSSION Blood cell recognition, disease diagnosis and some of the morphological methods are briefly reviewed in this section. This includes experimental investigations and image processing algorithms that have been developed to diagnosis diseases. To overcome the problems associated with blood cells, disease constituents and effectors during the diagnosis process, the image processing technique have been formulated with vision system to provide closed loop control. Several experiments were carried out to characterize the disease of the main modeling steps, i.e., establishing corresponding cells from a set of learning shapes and deriving a statistical model. We applied the modeling procedure to distinct the incoming image with the normal one. We test our methods on 15 images of healthy and 30 images with disease. In our approach, we established the corresponding cells prior to pose and scaling estimation during statistical modeling. This assumes that the position of the cells does not change significantly with different pose and scaling, which is justified by the high accuracy of the model.
Page 228
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
To assess how well the statistical model generalizes to unseen objects, we measured how close a shape not part of the learning sample can be approximated with the statistical shape model. To compute the robustness of the method with respect to different parameter settings, the entire modeling was carried out for a range of parameter at a time while keeping the remaining parameters at their optimal values. The images used are presented here, which have exhibited a wide variety of lesions and confusing manifestations. On this difficult data set, our methods successfully detect the disease in most of the reliable cases. IV.CONCLUSION In this paper, we have presented a method to automatically construct statistical blood cells from segmented microscopical images. Standard models are used to diagnosis the disease that can also be used for haematological applications. Corresponding parameters on the surface of the shape are established automatically by the adaptation of back propagation algorithm to segmented volumetric images. To assess the robustness of the method with respect to the parameter settings, the entire statistical model was carried out for range of parameter values, varying one at a time while keeping the remaining parameters at their optimal values. MS Visual C++ and Matrox Inspector image processing software are used for the experimental work. Nevertheless, this would be an interesting future investigation, since our approach is sufficiently general to be applied to modeling tasks from other two dimensional or three dimensional application domains. The extension of this may open a new area of research in bio mechanics (modeling and simulation). REFERENCES
cells”, Science Direct, Pattern Recognition, Volume 16, Issue 6, Pages 571-577, 1983. 5)Landeweerd.G.H, Timmers.T and Gelsema.E.S,” Classification of normal and abnormal samples of peripheral blood by linear mapping of the feature space”, Science Direct, Pattern Recognition, Volume 16, Issue 3, Pages 319-326, 1983. 6)Lester.J.M, Williams.H.A , Weintraub.B.A and Brenner.J.F, " Two graph searching techniques for boundary finding in white blood cell images”, Science Direct, Computers in Biology and Medicine Volume 8, Issue 4, Pages 293-308, 1978. 7)Parthenis.K, Metaxaki-Kossionides.C and Dimitriadis.B, "An automatic computer vision system for blood analysis ",Science Direct, Microprocessing and Microprogramming, Volume 28, Issues 1-5, Pages 243-246, March 1990. 8)Philip E. Norgren Ashok V. Kulkarni and Marshall D, Graham,“Leukocyte image analysis in the diff3 system”, Science Direct, Pattern Recognition, Volume 13, Issue 4, Pages 299-314, 1981. 9)Schönfeld.M and Grebe.R,”Automatic shape quantification of freely suspended red blood cells by isodensity contour tracing and tangent counting”, Science Direct, Computer Methods and Programs in Biomedicine, Volume 28, Issue, Pages 217-224, April 1989. 10)Xubo B. Song, Yaser Abu-Mostafa, Joseph Sill, Harvey Kasdan and Misha Pavel, “Robust image recognition by fusion of contextual information Information Fusion”, Science Direct, Volume 3, Issue 4, Pages 277-287, December 2002.
1)Artmann.G, Schneider.G and Henning, "Automated image processing system for shape recognition of single red blood cells based on out-of-focus images", Science Direct, Biorheology, Volume 32, Issues 2-3, Pages 237238, March-June 1995. 2)Brenner.J.F, Selles.W.D and Zahniser.D.J, "Pattern recognition of nucleated cells from the peripheral blood”, Science Direct, Pattern Recognition, Volume 16, Issue 2, Pages 131-140, 1983. 3)Daniela Mayumi Ushizima Sabino, Luciano da Fontoura Costa, Edgar Gil Rizzatti and Marco Antonio Zago "A texture approach to leukocyte recognition”, Science Direct, Real imaging in Bioinformatics, Volume 10, Issue 4, Pages 205-216, August 2004. 4)Ldeweerd.G.H, Timmers.T and Gelsema.E.S, ” Binary tree versus single level tree classification of white blood
Velammal College of Engineering and Technology, Madurai
Page 229
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Cloud Computing And Virtualization Mr. Nilesh Madhukar Patil Lecturer (IT Dept.), MCT’s Rajiv Gandhi Institute of Technology, Mumbai, State: Maharashtra, Country: India Email: [email protected]
Mr. Shailesh Somnath Sangle Lecturer (IT Dept.), MCT’s Rajiv Gandhi Institute of Technology, Mumbai, State: Maharashtra, Country: India Email: [email protected] Abstract Cloud Computing is defined as a pool of virtualized computer resources. Based on this Virtualization the Cloud Computing paradigm allows workloads to be deployed and scaled-out quickly through the rapid provisioning of virtual machines or physical machines. A Cloud Computing platform supports redundant, self-recovering, highly scalable programming models that allow workloads to recover from many inevitable hardware/software failures. The paper throws light on how to use cloud computing as a technical design center; how to take advantage of the economics of cloud computing in building and operating cloud services; and how to plan for the technical and operational scale that cloud computing makes possible. Keywords: SaaS, PaaS, IaaS
1. Cloud Computing Fundamentals For an organization to execute its operations, it requires components or hardware and software abided with it to furnish them. For instance, consider that in a software company each and every service is dependent upon other. Starting from Data center, server requires un- interrupted power supply, coolant solutions for its components to deliver effectively, Redundant Servers, Wide Range of Bandwidth to access them, complicated software’s to manage or monitor them, Team of Expert Employees monitoring or refurbishing them, development, production, fail over network. Upon all that suppose a new version or update is introduces than the whole system crashes down or requires a reallocation of resources. Ultimately the investment is found to be high but the results doesn’t seem stable or ever requires technical expert’s involvement. The search was up for a service which is agile, could be spread wider as time grows and could contract with the same speed as it grew, automatic updates without any user intervention, no knowledge or little requirement of technical trouble shooting expertise.
Velammal College of Engineering and Technology, Madurai
Figure 1.1: Cloud Computation
Cloud computing is using virtualized resources as a service over the internet with provision of scaling up dynamically. The improvement is that the users need not have to have knowledge of, expertise in, or control over the technology infrastructure over the “Cloud” that supports them. It operates through browsers while the software and the data are stored in the servers. A technical definition is "a computing capability that provides an abstraction between the computing resource and its underlying technical architecture (e.g., servers, storage, networks), enabling convenient, on-demand network access to a shared pool of configurable computing resources that can be rapidly provisioned and released with minimal management effort or service provider interaction." There have been many well known participants (ZOHO, Microsoft,
Amazon,
Rackspace
Cloud,
SalesForce,
Sun)
to
the
Yahoo, Cloud
Google, Computing
technologies, blowing the trumpet at high peak by introducing customary packages about the services offered by the Cloud Computing.
Page 230
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Figure 2.1: Virtualization Bundle Figure 1.2: Cloud Hosters
2. Virtualization Whenever an application or a service, is being developed by a company, its reliability to deliver at the native environment requires all its dependent services or resource to be bundled together due to the reason that the Application and Infrastructure are found to be dependent. Here technologies advanced to find a solution like bundling application with all it required services like Database, subsequent immediate services as well as Operating System into one, thereafter making Application independent of the Infrastructure into which the application is deployed. Yes very close, fueling the application to work independent of the Infrastructure is the architecture terms as virtualization. It refers to the Abstraction of computer resource. With the emergence of Virtualization, the internet cloud gathered multiple resources to equip for the Virtualization. Most of the time, servers don't run at full capacity. That means there's unused processing power going to waste. It's possible to fool a physical server into thinking it's actually multiple servers, each running with its own independent operating system. The technique is called server virtualization. By maximizing the output of individual servers, server virtualization reduces the need for more physical machines. Maneuvering virtualization technique according to the demands of the vendors is the simplest definition for Cloud Computing.
Velammal College of Engineering and Technology, Madurai
GMAIL is a Cloud Computing services. Now creating GMAIL account does not takes much time rather than a username and password. That’s it. This is the benefit of cloud computing. No server restart, no software updated, Just a simple username, password authentication, and then the resource is ready to be accessed by the end user. Cloud computing doesn’t only help customer application like “Account”, “Web Sites” but also support Business Architecture As such. Thereby Entitling it self as Enterprise Cloud Computing. Business Associates could pay for the resources they use. No additional cost/No wastage of Cost/Less cost. Talking about scalability, the cloud computing is termed as Elastic, enables customers to increase or decrease capacity within minutes, not hours or days. You can commission one, hundreds or even thousands of server instances simultaneously. Of course, because this is all controlled with web service APIs, your application can automatically scale itself up and down depending on its needs. To the End user, it is approached as a Multi-tenancy architecture, where in each compartment is handled individually without sharing resource for other compartment, but the fact is it is a shared platform where in individual Resource is being shared with all the available vendors compartmented by policies and permissions. Users have started rapidly using Cloud Computing technologies, and its usage is being encouraged. Few of the hosting service providers has prompted usage of Cloud Computing to help its reach to the common people have provided multiple shared services for free like Google Document (Sharing / Storing Documents in the cloud) and ZOHO Document (Sharing / Storing Documents in the cloud) are few to be mentioned.
Page 231
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3.1 Virtualized Infrastructure Virtualized Infrastructure provides the abstraction necessary to ensure that an application or business service is not directly tied to the underlying hardware infrastructure such as servers, storage, or networks. This allows business services to move dynamically across virtualized infrastructure resources in a very efficient manner, based upon predefined policies that ensure specific service level objectives are met for these business services. 3.2 Virtualized Applications Virtualized applications decouple the application from the underlying hardware, operating system, storage, and network to enable flexibility in deployment. Virtualized Application servers that can take advantage of grid execution coupled with Service Oriented Architectures and enable the greatest degree of scalability to meet the business requirements. 3.3 Enterprise Management Figure 2.2: Multitenancy Conceptual
Architecure
3. Cloud Building Blocks The building blocks of cloud computing are rooted in hardware and software architectures that enable innovative infrastructure scaling and virtualization. Many data centers deploy these capabilities today. However, the next infrastructure innovations are around more dynamic provisioning and management in larger clusters both within and external to the conventional corporate data center. There are also implications for next generation application design to make optimum use of massively parallel processing and fault tolerance. The figure 3 below illustrates some common architectural components:
Enterprise management provides top-down, end-to-end management of the virtualized infrastructure and applications for business solutions. The enterprise management layer handles the full lifecycle of virtualized resources and provides additional common infrastructure elements for service level management, metered usage, policy management, license management, and disaster recovery. Mature cloud service management software allows dynamic provisioning and resource allocation to allow applications to scale on demand and minimize the waste associated with underutilized and static computing resources. 3.4 Security and Identity Management Clouds must leverage a unified identity and security infrastructure to enable flexible provisioning, yet enforce security policies throughout the cloud. As clouds provision resources outside the enterprise’s legal boundaries, it becomes essential to implement an Information Asset Management system to provide the necessary controls to ensure sensitive information is protected and meets compliance requirements. 3.5 Development tools
Figure 3: Cloud Components
Velammal College of Engineering and Technology, Madurai
Next generation development tools can leverage cloud’s distributed computing capabilities. These tools not only facilitate service orchestration that can leverage dynamic provisioning, but also enable business processes to be developed that can harness the parallel processing capabilities available to clouds. The development tools
Page 232
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
must support dynamic provisioning and not rely on hard coded dependencies such as servers and network resources. 4. Cloud Computing Infrastructure Models There are many considerations for cloud computing architects to make when moving from a standard enterprise application deployment model to one based on cloud computing. There are three basic service models to consider: Public, private, and hybrid clouds IT organizations can choose to deploy applications on public, private, or hybrid clouds, each of which has its trade-offs. The terms public, private, and hybrid do not dictate location. While public clouds are typically “out there” on the Internet and private clouds are typically located on premises, a private cloud might be hosted at a colocation facility as well. Companies may make a number of considerations with regard to which cloud computing model they choose to employ, and they might use more than one model to solve different problems. An application needed on a temporary basis might be best suited for deployment in a public cloud because it helps to avoid the need to purchase additional equipment to solve a temporary need. Likewise, a permanent application, or one that has specific requirements on quality of service or location of data, might best be deployed in a private or hybrid cloud. 4.1 Public clouds
virtual machine images, but also servers, storage systems, network devices, and network topology. Creating a virtual private datacenter with all components located in the same facility helps to lessen the issue of data locality because bandwidth is abundant and typically free when connecting resources within the same facility.
Figure 4.1: Public Cloud
4.2 Private clouds Private clouds are built for the exclusive use of one client, providing the utmost control over data, security, and quality of service (Figure 4.2). The company owns the infrastructure and has control over how applications are deployed on it. Private clouds may be deployed in an enterprise datacenter, and they also may be deployed at a colocation facility. Private clouds can be built and managed by a company’s own IT organization or by a cloud provider. In this “hosted private” model, a company such as Sun can install, configure, and operate the infrastructure to support a private cloud within a company’s enterprise datacenter. This model gives companies a high level of control over the use of cloud resources while bringing in the expertise needed to establish and operate the environment.
Public clouds are run by third parties, and applications from different customers are likely to be mixed together on the cloud’s servers, storage systems, and networks (Figure 4.1). Public clouds are most often hosted away from customer premises, and they provide a way to reduce customer risk and cost by providing a flexible, even temporary extension to enterprise infrastructure. If a public cloud is implemented with performance, security, and data locality in mind, the existence of other applications running in the cloud should be transparent to both cloud architects and end users. Indeed, one of the benefits of public clouds is that they can be much larger than a company’s private cloud might be, offering the ability to scale up and down on demand, and shifting infrastructure risks from the enterprise to the cloud provider, if even just temporarily. Portions of a public cloud can be carved out for the exclusive use of a single client, creating a virtual private datacenter. Rather than being limited to deploying virtual machine images in a public cloud, a virtual private datacenter gives customers greater visibility into its infrastructure. Now customers can manipulate not just
Velammal College of Engineering and Technology, Madurai
Figure 4.2: Private Cloud
4.3 Hybrid clouds Hybrid clouds combine both public and private cloud models (Figure 4.3). They can help to provide on-demand, externally provisioned scale. The ability to augment a private cloud with the resources of a public cloud can be used to maintain service levels in the face of rapid workload fluctuations. This is most often seen with the use
Page 233
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of storage clouds to support Web 2.0 applications. A hybrid cloud also can be used to handle planned workload spikes. Sometimes called “surge computing,” a public cloud can be used to perform periodic tasks that can be deployed easily on a public cloud.
market, including the Google Apps offering of basic business services including email and word processing. Although salesforce.com preceded the definition of cloud computing by a few years, it now operates by leveraging its companion force.com, which can be defined as a platform as a service.
Hybrid clouds introduce the complexity of determining how to distribute applications across both a public and private cloud. Among the issues that need to be considered is the relationship between data and processing resources. If the data is small, or the application is stateless, a hybrid cloud can be much more successful than if large amounts of data must be transferred into a public cloud for a small amount of processing.
Figure 5: Cloud computing means using IT infrastructure as a service
5.2 Platform as a service (PaaS) Platform as a service encapsulates a layer of software and provides it as a service that can be used to build higherlevel services. There are at least two perspectives on PaaS depending on the perspective of the producer or consumer of the services:
Figure 4.3: Hybrid Cloud
5. Architectural layers of cloud computing In practice, cloud service providers tend to offer services that can be grouped into three categories: software as a service, platform as a service, and infrastructure as a service. These categories group together the various layers illustrated in Figure 5, with some overlap. 5.1 Software as a service (SaaS) Software as a service features a complete application offered as a service on demand. A single instance of the software runs on the cloud and services multiple end users or client organizations. The most widely known example of SaaS is salesforce.com, though many other examples have come to
Velammal College of Engineering and Technology, Madurai
• Someone producing PaaS might produce a platform by integrating an OS, middleware, application software, and even a development environment that is then provided to a customer as a service. For example, someone developing a PaaS offering might base it on a set of Sun™ xVM hypervisor virtual machines that include a NetBeans™ integrated development environment, a Sun GlassFish™ Web stack and support for additional programming languages such as Perl or Ruby. • Someone using PaaS would see an encapsulated service that is presented to them through an API. The customer interacts with the platform through the API, and the platform does what is necessary to manage and scale itself to provide a given level of service. Virtual appliances can be classified as instances of PaaS. A content switch appliance, for example, would have all of its component software hidden from the customer, and only an API or GUI for configuring and deploying the service provided to them.
Page 234
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PaaS offerings can provide for every phase of software development and testing, or they can be specialized around a particular area such as content management. Commercial examples of PaaS include the Google Apps Engine, which serves applications on Google’s infrastructure. PaaS services such as these can provide a powerful basis on which to deploy applications, however they may be constrained by the capabilities that the cloud provider chooses to deliver. 5.3 Infrastructure as a service (IaaS) Infrastructure as a service delivers basic storage and compute capabilities as standardized services over the network. Servers, storage systems, switches, routers, and other systems are pooled and made available to handle workloads that range from application components to highperformance computing applications. Commercial examples of IaaS include Joyent, whose main product is a line of virtualized servers that provide a highly available on-demand infrastructure. 6. Benefits of Cloud Computing 1. Decoupling and separation of the business service from the infrastructure needed to run it (virtualization) 2. Flexibility to choose multiple vendors that provide reliable and scalable business services, development environments, and infrastructure that can be leveraged out of the box and billed on a metered basis—with no long term contracts. 3. Elastic nature of the infrastructure to rapidly allocate and de-allocate massively scalable resources to business services on a demand basis. 4. Cost allocation flexibility for customers wanting to move Capital Expenditures (CapEx) into Operational Expenditures (OpEx). 5. Reduced costs due to operational efficiencies, and more rapid deployment of new business services. 7.
Issues in Cloud Computing
The main concerns about cloud computing are security and privacy. The thought of handing your important data over to something called a “cloud” can be daunting. Nervous corporate executives might hesitate to take advantage of a cloud computing system because they feel like they’re surrendering control of their company’s information. Data inside the ‘cloud’ is outside a company’s firewall and that brings with it an intrinsic threat of risk, because services that companies outsource evade the physical, logical and
Velammal College of Engineering and Technology, Madurai
personnel controls that I.T. shops wield over data maintained in-house. Other fears include: 1. Risk of data breaching 2. Appeal to cyber crooks 3. Lack of specific standards for security and data privacy 4. Questions about jurisdiction. European concern about U.S. privacy laws led to creation of the U.S. Safe Harbor Privacy Principles, which are intended to provide European companies with a degree of insulation from U.S. laws 5. Data location. Cloud users probably do not know exactly where their data is hosted—not even the specific country. 8.
Conclusion
Cloud computing offers real alternatives to IT business for improved flexibility and lower cost. Markets are developing for the delivery of software applications, platforms, and infrastructure as a service to IT business over the “cloud”. These services are readily accessible on a pay-per-use basis and offer great alternatives to businesses that need the flexibility to rent infrastructure on a temporary basis or to reduce capital costs. Architects in larger enterprises find that it may still be more cost effective to provide the desired services in-house in the form of “private clouds” to minimize cost and maximize compatibility with internal standards and regulations. If so, there are several options for future-state systems and technical architectures that architects should consider to find the right trade-off between cost and flexibility. Using an architectural framework will help architects evaluate these trade-offs within the context of the business architecture and design a system that accomplishes the business goal. 9. References [1] Michael Miller, “Cloud Computing: Web-Based Applications That Change the Way You Work and Collaborate Online”, Que, 2008. [2] David Linthicum, “Cloud Computing and SOA Convergence in Your Enterprise: A Step-by-Step Guide”, 1st edition, Addison-Wesley Professional, 2009. [3] Toby Velte, Anthony Velte, Robert Elsenpeter, “Cloud Computing, A Practical Approach”, 1st edition, McGrawHill Osborne Media, 2009. [4]http://www.exforsys.com/tutorials/cloudcomputing/cloud-computing-users-perspective.html
Page 235
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[5]http://www.watblog.com/2008/03/25/yahoocomputational-research-laboratories-team-up-for-cloudcomputing-research/ [6]http://computersoftware.suite101.com/article.cfm/cloud _computing_and_virtualization_impact [7]http://www.microsoft.com/virtualization/en/us/privatecloud.aspx
Velammal College of Engineering and Technology, Madurai
Page 236
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
New Method For Solving Fuzzy Linear Programming With TORA S. Sagaya Roseline , A. Faritha Asma , E.C. Henry Amirtharaj, Department of Mathematics , Bishop Heber College (Autonomous) Tiruchirappalli -17, India. Email: [email protected]
[email protected]
Abstract - Since many real-world engineering systems are too complex to be defined in precise terms, imprecision is often involved in any engineering design process. Fuzzy linear programming problems have an essential role in fuzzy modeling, which can formulate uncertainty in actual environment. In this study we present a new method for solving FLPP with TORA software. Here the coefficient of objective function , constraint coefficient matrix, and right hand side of the constraints are in fuzzy nature. The coefficients of the objective function have taken as trapezoidal fuzzy number, the coefficients of the constraints and the right hand side of the constraints are considered as a triangular fuzzy number. Here our main scope is to find some non-negative vector x which maximizes the objective function z = c x so that A x = b . Finally the defuzzified LPP is solved by using TORA.
Keywords - Fuzzy sets, Fuzzy number, Trapezoidal fuzzy number, Triangular fuzzy number, Fuzzy linear programming.
I. INTRODUCTION Fuzzy set theory has been applied to many fields of Operations Research. The concept of fuzzy linear programming (FLP) was first formulated by Zimmermann. And then various types of the FLP problems are considered by many authors and several approaches for solving these problems are proposed. In this paper we consider a LPP in which the coefficients of the objective function, constraint coefficients and right hand side of the constraints are fuzzy. Here we explain the concept of the comparison of fuzzy numbers by introducing a linear ranking function. Our main contribution here is the establishment of a new method for solving the Fuzzy linear programming problems with TORA. Moreover we illustrate our method with an example. II. PRELIMINARY
[email protected]
A. FUZZY SETS Definition Let X be a classical set of objects called the universe whose generic elements are denoted by x. The membership in a crisp subset of X is often viewed as characteristic function μA(x) from X to {0, 1} such that μA(x) = 1 , if x Є A = 0 , otherwise. where {0, 1} is called valuation set. If the valuation set is allowed to be the real interval [0, 1], A is called a fuzzy set. μA(x) is the degree of membership of x in A. The closer the value of μA(x) is to 1, the more x belong to A. Therefore, A is completely characterized by the set of ordered pairs: A = { ( x, μA(x)) / x Є X } Definition The support of a fuzzy set A is the crisp subset of X and is presented as : Supp A = { x Є X / μA(x) > 0 } Definition The α level (α – cut ) set of a fuzzy set A is a crisp subset of X and is denoted by Aα = { x Є X / μA(x) ≥ α) } Definition A fuzzy set A in X is convex if μA( λx + (1-λ)y) ≥ min { μA(x) , μA(y) } x , y Є X and λ Є [0, 1] . Alternatively, a fuzzy set is convex if all α level sets are convex . Note that in this paper we suppose that X = R. B. Fuzzy numbers
In this section we review some necessary backgrounds of the fuzzy theory which will be used in this paper. Below we give definitions and notations.
Definition A fuzzy number A is a convex normalized fuzzy set on the real line R such that 1) it exists atleast one x0 Є R with μA(x0) = 1. 2) μA(x) is piecewise continuous.
Velammal College of Engineering and Technology, Madurai
Page 237
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Among the various types of fuzzy numbers, triangular and trapezoidal fuzzy numbers are of the most important. Definition A fuzzy number is a trapezoidal fuzzy number if the membership function of it be in the following form :
x > 0 , x a = ( xaL , x aU , x α, x β ) x < 0 , x a = ( xaU , xaL , - x β, - x α ) IV. RANKING FUNCTIONS A convenient method for comparing of the fuzzy numbers is by use of ranking functions. A ranking function is a map from F(R) into the real line. Now , we define orders on F(R) as following:
μ a
a ≥ b if and only if R (a) ≥ R (b) R x
1 x 0
aL - α
aL
aU
aU + β
We show that trapezoidal fuzzy number by a = (aL , aU , α, β ) where the support of a is ( aL - α , aU + β ) , and the modal set of a is [aL , aU ] . Let F(R) be the set of trapezoidal fuzzy numbers. Definition Triangular fuzzy number A can be represented by three real numbers, s, l, r whose meanings are defined as
1
l
r
R
a > b if and only if R (a) > R (b) R a = b if and only if R (a) = R (b) R where a and b are in F(R). It is obvious that we may write a ≤ b if and only if b ≥a R R Since there are many ranking function for comparing fuzzy numbers we only apply linear ranking functions. So, it is obvious that if we suppose that R be any linear ranking function, then i) a ≥ b if and only if a - b ≥ 0 if and only if - b ≥ - a R R ii) a ≥ b and c ≥ d , then a + c ≥ b + d R R R One suggestion for a linear ranking function is following: R (a) = aL + aU + (β - α) where a = ( aL , aU , α, β ) .
0 s-l
s
s+r
Using this representation we write A = (s, l, r ) Let F1 (R) be the set of all triangular fuzzy numbers. III. ARITHMETIC ON TRAPEZOIDAL FUZZY NUMBERS Let a = ( aL , aU , α, β ) and b = ( bL , bU , γ , θ ) be two trapezoidal fuzzy numbers and x Є R. Then the results of applying fuzzy arithmetic on the trapezoidal fuzzy numbers are shown in the following : U
L
Image of a : - a = (- a , - a , β, α ) Addition : a + b = ( aL + bL , aU + bU , α + γ, β + θ ) Scalar Multiplication:
Velammal College of Engineering and Technology, Madurai
V. FUZZY LINEAR PROGRAMMING PROBLEM Consider Fuzzy linear programming problem in which the coefficient of the objective function , the right hand side numbers and the coefficients of the constraint matrix are fuzzy numbers. Max z = xj R Sub to
xj ≤ Bi (i Є Nm) xj ≥ 0 (j Є Nn)
According to [2] the problem can be rewritten as Max z = R
xj
Page 238
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Sub to
,
,
)
( ti , ui, vi )
( i Є Nm )
Max z = ( 5, 8, 2, 5) x1 + ( 6, 10, 2, 6) x2
xj ≥ 0 ( j Є Nn )
Sub to 4x1 + 5x2 ≤ 24 4x1 + x2 ≤ 12 2x1 + 2x2 ≤ 19 3x1 +0.5x2 ≤ 6 5x1 + 6x2 ≤ 32 6x1 + 2x2 ≤ 15 x1 , x2 ≥ 0
where Aij = ( sij , lij , rij ) and Bi = ( ti , ui, vi ) are triangular fuzzy numbers.cj Є F(R) Summation and multiplication are operations on fuzzy numbers, and the partial order ≤ is defined by A ≤ B iff MAX ( A, B) = B. For any two triangular fuzzy numbers A = (s1, l1, r1) and B = (s2 , l2 , r2 ) A ≤ B iff s1 ≤ s2 , s1 - l1 ≤ s2 - l2 and s1 + r1 ≤ s2 + r2. Moreover, (s1, l1, r1) + (s2 , l2 , r2 ) = (s1+s2 , l1+l2 , r1+r2 ) and (s1, l1, r1)x = (s1x, l1x, r1x) for any non-negative real number x. Then the problem can be rewritten as Max z = R
R(z) = 86.515 Defuzzifing the objective function using ranking function , the FLPP becomes the following classical LPP.
xj
Sub to
Solving this problem we obtain, x1 = 1.23, x2 = 3.82 max z = ( 29.04 , 48, 10.09 , 29.04)
Max z = 14.5 x1 + 18 x2
≤ ti
Sub to 4x1 + 5x2 ≤ 24 4x1 + x2 ≤ 12 2x1 + 2x2 ≤ 19 3x1 +0.5x2 ≤ 6 5x1 + 6x2 ≤ 32 6x1 + 2x2 ≤ 15 x1 , x2 ≥ 0
≤ ti - ui ≤ ti + vi (i Є Nm) xj ≥ 0 ( j Є Nn ) Now the Fuzzy linear programming problem becomes Max z = c x R Sub to A x ≤ b
Solving this LPP using TORA we obtain, x1 = 1.23, x2 = 3.82 , max z = 86.52 (1)
x≥0 c T Є (F(R))n , b Є Rm , x Є Rn , R is a linear ranking function. Then according to [1] , the problem can be solved by using simplex method. By using ranking function , objective function of (1) can be defuzzified and then (1) is equivalent to the classical LPP which can be solved by using TORA.
VII. CONCLUSION In this paper we proposed a new method for solving the FLP problems with TORA. The significance of this paper is providing a new method for solving the fuzzy linear programming in which the coefficients of the objective function are trapezoidal fuzzy number and the coefficients of constraints, the right hand side of the constraints are triangular fuzzy number. We compared the solution of FLPP with the defuzzified LPP solution using TORA. REFERENCES
VI. A NUMERICAL EXAMPLE For an illustration of the above method we solve a FLP problem with TORA.
[1] [2]
Max z = ( 5, 8, 2, 5) x1 + ( 6, 10, 2, 6) x2
[3]
Sub to ( 4, 2, 1) x1 + ( 5, 3, 1) x2 ≤ ( 24, 5, 8) ( 4, 1, 2) x1 + ( 1, 0.5,1)x2 ≤ ( 12, 6, 3) x1 , x2 ≥ 0
[4] [5]
S. H. Nasseri, E. Ardil, A. Yazdani and R. Zaefarian, Simplex method for solving linear programming problems with fuzzy numbers, World academy of science, engineering and technology,2005 . George j. Klir / Bo Yuan, Fuzzy sets and fuzzy logic Theory and applications, 2006. S.H. Nasseri, A new method for solving fuzzy linear programming by solving linear programming, Applied mathematical sciences,vol.2, 2008. H. Rommelfanger, R. Hanuscheck and J. Wolf, Linear programming with fuzzy Objective funcion, Fuzzy sets and systems,1989. H.J. Zimmermann, Fuzzy programming and linear programming with several objective functions, Fuzy sets and systems, 1978.
we can rewrite it as
Velammal College of Engineering and Technology, Madurai
Page 239
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Latest Trends And Technologies In Enterprise Resource Planning – ERP Author–Dakshayani.B.S. Department of Computer Science, Amrita Vishwa Vidyapeetham, Mysore Campus, Karnataka. [email protected]
Abstract—This article deals with, a) The Latest trends in ERP, b) Recent technologies like Open Source, Web 2.0 services, SaaS, Cloud Computing and Green ERP Systems, Compatibility with dot net technologies c) Future of ERP Key words—Green ERP, Cloud Computing, SaaS, Wireless Technology, Web based ERP, Compatibility of ERP with Dot net technology
I INTRODUCTION ERP evolved from Manufacturing Resource Planning – MRP II (which originated from Material Resource Planning – MRP I). It has gained much prominence and utility with the intervention of open source, web enabled and wireless technologies. ERP has undoubtedly become an important business application to all industries. It has almost become a must for all organizations irrespective of the type of business manufacturing or service. In this context it becomes important to analyze the direction in which ERP is geared to progress or will ERP diminish in the future, emerging technologies etc. II LATEST TRENDS IN ERP ERP calls for constant modifications and up gradations. ERP developers are facing tremendous pressure both from vendors and companies. In this context it becomes important to analyze the ERP’s trends and modalities. The following are relevant issues in ERP. A. Need based application Organizations had to implement ERP throught their systems irrespective of the fact whether they help in all the functions or in one particular function. This was proving to be a big hurdle to the firms. In addition this remained as the main disadvantage or setback of ERP. They had to purchase the whole applications even if it meant that most of them would be idle except for the core function. The latest ERP software programs have overcome this menace. They offer need based applications.. They were given the liberty to purchase and install software Programs pertaining to that particular function. This advantage has
Velammal College of Engineering and Technology, Madurai
helped to increase the scope of ERP not only among large firms but also small and medium business as well. B. Expenditure ERP was a very costly affair. Thanks to the intrusion of internet and open source applications. This has helped S.M.E.'S to enter the market of prospective buyers. This has not only widened the horizon of S.M.E.'s but also increased the usage among large firms. These large firms were not able to invest huge money in spite of adequate funds. Now that the spending on ERP gets reduced there are no hesitations to show the green signal for monetary outlay. It is encouraging to notice the improving IT ERP trends. C. Implementation Time ERP was discouraged by companies because they took such a long time to get installed and set the whole process into action. Since this resource was spent excessively there were chances for reduction in potential business and losing manhours. The current day ERP applications are less complex to install and train. This has reduced the amount of time spent on ERP. Companies are thereby assured of spending lesser time for ERP. D. Open Source, Web enabled (Web 2.0 services), SaaS, Cloud Computing, Green ERP and wireless technologies These technologies are the results of finding solution to the issues discussed above. They have been dealt in detail in the following paragraphs. These technologies have helped in rejuvenation of the functioning of ERP and also in creating wide market base, III RECENT TECHNOLOGIES IN ERP A. Open Source ERP technologies Open Source Technologies have made the job of ERP easier. It has done away with the hassles of paying licence fee not only during installation but also whenever a modification is made. The company is relieved from depending even for mince matters. Following issues relating to open source ERP technology is worth discussing. 1) Influence of the cost
Page 240
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
factor: It is interesting to know the Price tag of Open Source ERP technologies. It is literally available at free of cost. All that the user has to do is to download the software and install it. An unbelievable fact is that even the source code comes freely. This in itself has encouraged lots of companies to go for ERP, as they are not burdened with investments. Of late, companies don't necessarily go for ERP because their attitude towards spending on ERP has undergone a sea change in the sense they don't mind to pay as long as they think ERP is worth the costs. Open Source accounting ERP and open source ERP payment are famous solutions. 2) Influence on the operational expenses: Open Source ERP technologies largely influence operational expenses. The company is relieved from paying the extra sum for facilities provided in moving to a new system. Similarly the company need not incur additional expenditures for renewal and purchase of licenses as and when some components are added in the framework. This gradually reduces the monetary outlay that has to be otherwise incurred for every update. Open Source accounting ERP has helped to simplify the financial practices. Open source ERP payment has helped in facilitating easy disbursement of cash. 3) Absence of Vendors help: Unlike the usual ERP applications it is not possible to avail the services of a vendor as the company handles everything independently. This has many dimensions. Firstly the company enjoys a sole liability. Secondly a simple error when not rectified (or the in-house personnel does not know to) could prove to be a costly affair for the company. Above all the company gets to learn from mistakes and without any external assistance 4) Litigations: Open source ERP has resulted in many lawsuit and incidental claims. There is still ambiguity in the copying aspects. The question of infringement and indemnification remains unanswered as seen from the previous cases. 5) Unsuitable for all applications: Open source has a limit when it comes to the places where they can be put to use. They don't find applicability in all areas especially for conventional practices. It is not appropriate to introduce open source in those areas without changing the way the systems work. In fact it could be a risky option to do it. This drawback discourages many functions from being Open source friendly
B. Web Based ERP This Web Based ERP is usually less expensive than a fullblown ERP software package which is usually priced out of the range for a small or medium company. Also with fullblown packages, there will probably be many modules that will not be implemented, even though those modules have been purchased. On the other hand, Web Based ERP can be purchased and implemented in stages. You only buy what you need. Since Web based ERP is based upon the web, it can be used from anywhere you have access to the internet or even
Velammal College of Engineering and Technology, Madurai
through an intranet. This means that financial information’s available whenever it is needed, from wherever you happen to be. This is an important criterion for people like Sales Reps to corporate CEO’s. Sales analysis, profitability, inventory, production, pricing, credit history: these are just some the information that’s available just by using the web to securely connect, in real-time, to a secure database SaaS – ERPshe Software as a Service (SaaS) model is a way of providing the same software to different customers via a network, usually the internet. In other words, the software is not hosted on the customer’s individual computers. Under the SaaS model, a vendor is responsible for the creation, updation, and maintenance of software. Customers buy a subscription to access it, which includes a separate license, or seat, for each person that who use the software. This model can add efficiency and cost savings for both the vendor and customer. Customers save time and money since they do not have to install and maintain programs. The customers do not have to hire staff, or use existing staff to maintain the software. They also generally do not have to buy any new hardware. This allows a customer to focus more resources on growing the business.Shifting the burden of software hosting and development to the vendor can also speed up the time it takes for the customer to see a return on the software investment. Using the SaaS model, the number of seats can be increased as the business grows. This is usually faster and cheaper than purchasing another license and adding to another computer, as with traditional software.Vendor usually only have to update and maintain the software on the network, versus updating different copies of the software on different computers. This allows the vendor to provide the latest updates and technology to each customer in a timely manner. The drawback for the customer is that they do not control the software and customization of programs may be limited. If an update is requested by a customer, it will most likely need to benefit other customers who are also using the same software. If the customer completely outgrows the software, however, the company can simply discontinue its subscription at the end of the current contract. In such a cancellation, applications typically do not have to be removed from the customer’s computers. Generally, the canceling customer maintains ownership of any proprietary data entered into the SaaS applications. SaaS model contracts may be terminated early with sufficient cause. The vendor not delivering the software in a timely manner, the software not working as specified in the contract, are all typically grounds for termination of the contract. With broadband technology more commonplace throughout the workforce, however, customers have many
Page 241
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
choices when it comes to software delivered under the SaaS model. Customers can research SaaS vendors thoroughly, and request current references, to help avoid any nondelivery issues. Plex Online Company utilizes SaaS model to deliver fullblown manufacturing ERP software functions in an ondemand mode. Plex online is a web based ERP solution, requiring only a web browser to access all of the analysis and management functions of a full-featured manufacturing ERP system, designed to meet the requirements of all departments within a quality-driven manufacturer, shop floor to top floor. Web 2.0 and SaaS software have captured a significant portion of the market led by legacy ERP systems. Even large enterprises (which only showered their love on ERP’s) are now accepting the value of SaaS application over legacy ERP’s. A survey about SaaS conducted in August 2009 by Business Week Research Service found that four out of five managers and senior executives in North America are either interested in, or in the process of, adopting the SaaS approach to IT. In fact roughly a third of the 326 respondent’s companies have already fully or partially adopted the SaaS approach for at least the one application. The key features of SaaS applications that are usually missing in traditional ERP software – usability, intuitive Graphic User Interface, collaboration, global accessibility, configurable, higher security levels and zero IT hassles for clients.
D. Cloud Computing Represents the next evolutionary step toward elastic IT. It will transform the way your IT infrastructure is constituted and managed, through consumable services for infrastructure, platform and applications. This will convert your IT infrastructure from a “factory” into a “supply chain”. Cloud computing is a general term for anything that involves delivering hosted services over the Internet. These services are broadly divided into three categories: Infrastructure-as-a-Service (IaaS), Platform-as-a-Service (PaaS) and Software-as-a-Service (SaaS). The name cloud computing was inspired by the cloud symbol that's often used to represent the Internet in flowcharts and diagrams. Cloud Computing SaaS is the next generation ERP and optimized for B2B/SMB’s with built-in CRM and SFA to cut costs and improve productivity.
Velammal College of Engineering and Technology, Madurai
The highest level of Cloud Computing is SaaS, BScaler Cloud computing SaaS is a leader in SaaS, using Web 2.0, Web Services, SOA, AJAX, XML, Java and SaaS 2.0 technologies to provide users with the highest security and data access performance via internet browsers. With BScaler ERM(Enterprise Resource Manager) SaaS, no upfront investment in hardware or software is required, customers do not have to operate their applications or IT infrastructure and maintenance and updates are no longer necessary! Furthermore, wherever you are in the world – at any time – your business will be with you. All you need is browser! This software also lets users focus on the businesses instead of an over-loaded, quickly aging data centers. E. Green ERP Scientists and intellectuals have become increasingly concerned with the human impact on our environment. Governments are now regulating businesses with strict environmental restrictions. Software systems such as Green ERP are gaining popularity as companies transform their practices. This concept is relatively new, so very few companies offer ERP systems that integrate environmental goals into the system’s structure. The best option is tailored software. Tailored ERP software utilizes the best features of both generic (turn-key) and custom software while avoiding all of the disadvantages. The concept, tailoring software was started to allow flexibility in a software system while avoiding the pit falls of a custom system such a high cost, having to rebuild the system to upgrade and long development time. The best tailored or configured software begins with an existing software structure. Key individuals at the company help to model the software’s functionality, but unlike custom coding, the lengthy process of developing the software is circumvented by high-end programming tools. Since the basic architecture’s not altered, upgrades as simple as with a generic, out-of-the-box software system. F. Wireless technology in ER: Wireless technology has helped enterprise operations in many ways. Firstly it has facilitated the stakeholders in getting upto-date information on enterprise operations as and when required through the use of modern communication devices like mobile phones, laptops (all connected to the internet).
Page 242
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1) Proper coordination with the organizational network: Wis ERP will give the best results only if it falls in line with proper communication channels. The communication channels should be improved in the organization to make it (Wireless) ERP friendly. The obsolete computers should be replaced with the latest phones. Mobile and telecommunication facilities should also be at par with the industry standards. It happens many a times that companies resort to wireless ERP without improving the communication facilities. This does not serve the purpose no matter whatsoever the money spent in establishing ERP or the resource persons employed to procure the same. In addition there is another major advantage in improving communication facilities. Apart from dissemination of information and inflating profits and improving productivity it also helps the companies to rise to professional standards in the market. This will also motivate the companies to improve all other facilities that directly or indirectly contribute to the working of ERP and make use of facilities like image Enabled ERP system and ERP data capture. 2) Privacy issues: Privacy is a burning issue that occupies significance whenever there is a technological explosion and ERP is no exception to this principle. These calls for more attention due to the (further) improvement namely wireless ERP (from ERP). Privacy always becomes a subject matter of conflict whenever things are brought under public domain. The company is poised to lose valuable information or prospective businesses when things are exposed unnecessarily .In those cases wireless ERP will be a disaster to the company (and not a boon). Wireless ERP should be provided with maximum security. The scrutiny should prevent third party access without prior permission and approval from the company. This might look impossible practically but it is strongly advocated to have them in force in order to safeguard the interests of the company. The security system will be successful if a department in the company is able to have a track of the details communicated. The viability of the company is also to be taken into account in this regard. Bigger companies can go for such system whereas the others should restrict depending on their capacity and needs. If they are required unconditionally the companies should resort to alternative arrangements .It should be kept in mind that none of these arrangements should interfere with the functioning of ERP. Image Enabled ERP system and ERP data capture will help to achieve this.
improvement. This will automatically help wireless ERP to succeed. 4) Multiple applications and connectivity speed: Wireless ERP should not depend on just one or two devices like laptop or mobile phone but make use of maximum appliances. This will help the companies to gain technological edge. In additions alternatives will be of great help especially in times of emergency and when a particular system fails. G. What makes dot net compatible to ERP? What are the probable reasons for preferring Dot net platform? They are supposed to be cheap when compared with J2EE. Since dot net is from Microsoft Corporation it goes without saying that they will be designed on application basis. Besides there are some other features that make dot net compatible to ERP. Some of the issues deserving attention are as follows: 1) Effectively checks for mistakes and procedural errors: Be it enterprise resource planning or software programs the chances of errors are often very high. These errors not only undermine the working of the software but also disturb the whole process in the company which even makes them to sigh " If only computers were not invited". No platform (including dot net and J2EE) can be guaranteed for controlling errors. However the unique codes set in dot net helps in knowing as and when errors occur and that too at the early stage. This is an important issue especially for ERP the functioning of which has to be constantly monitored to ensure accountability and performance. Since the platform immediately points out what is going wrong it is easier for the company to freeze the whole process rectify the errors and set back things to motion properly. 2) Flexibility to function simultaneously: Any platform or application poses a major challenge to the company. The more they are different and the larger are they in numbers the greater is the trouble for the company. The reason is that the simultaneous function leads to lot of collusions and confusions. Even though they work separately the functioning of one will clash with that of the other. This case could be worse if one of them is dependent on the other for one or more supportive function. Dot net has been designed to overcome this drawback. Dot net will neither have interference nor an influence with other applications/platforms unless they are desired in the course of work. Even in such cases dot net will only act as an inducer. This explains the reason for preferring dot net platform.
3) Message persistency: The frequency of message transfer is referred as message persistency. This rate is important in deciding the success of the communication systems and ERP functions. The organization should ensure that there is no slag in the systems or process or any other procedures that is likely to affect this rate. Companies should constantly monitor and suggest areas for
3) Safety and lesser risk: Computers and software programs become vulnerable to hacking and all sort of mishaps and no amount of technological advancements have helped to resolve this trouble. The security features of dot net platform are not only stringent but also a tough thing for the user himself. Unlike from other software programs which award
Velammal College of Engineering and Technology, Madurai
Page 243
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
licenses directly to the user the special facility is that they can have the liberty to award the license. 4) Communication, costs and other services: Dot Net software offers these facilities at an unbelievable cost that is predicted as low as 5% when compared with J2EE platforms. This is good news to companies and especially those who are unable to make use of the latest IT facilities due to budgetary constraints. This would have otherwise been a handicap to companies for not being able to serve their own requirements. Dot net also facilitates faster communication and renders many additional services to companies availing them. Since they are one among the latest the vendors take pains in ensuring that the companies understand the entire system and still continue preferring dot net platform.
Companies are now focusing on optimizing processes within those functions. Aligning employee roles and responsibilities with a process, as well as function, can reap big benefits by giving employees visibility to the process from end to end. Having a common picture of how a process works "stimulates tremendous innovation and improvement,". "It's a new way of looking at your business." ERP has thrown open opportunities for many companies to trade with foreign counter parts in the name of outsourcing, implementation and deployment of the existing ones. It has contributed lot to the economy. Academics also boast its own share of ERP relations. It has promoted lot of employment and educational opportunities. India happens to be a key beneficiary in this aspect.
IV FUTURE MARKET TRENDS OF ENTERPRISE RESOURCE PLANNING.
V LEADING ERP GIANTS:
The dream of enterprise resource planning systems is that a single application can track and monitor all of a business’s functions. In a perfect world, a manager opens a single ERP app to find data about any aspect of the business, from financials to HR to distribution schedules. Also, we’re not there yet – or at least most companies aren’t. Moreover, there are still a lot of gaps in ERP systems, particularly in industries where ERP functionality has grown up from its historic origins in manufacturing. There are even gaps in core ERP areas, Hamerman tells Datamation, “Where they just haven’t done a particularly good job, in areas like budgeting, and recruitment…where the vast majority of customer uses something other than their ERP vendor.” But despite the challenges, the movement toward a global ERP system is a key factor shaping the future of enterprise resource planning Companies looking to "shift from a functional to a business process orientation" will need a very long-term plan, and possibly a therapist on hand.
This is a niche market where only the best gets to thrive. Needless to mention the money involved and other factors ensure a tug of war between the players. Some of the leading ERP giants are as follows: 1) Microsoft: This software major holds a promising segment in the Small and medium enterprises market. They have been constantly upgrading the versions and are expert in manufacturing the products from the vendor's point of view (who knows all the practical difficulties of the stakeholders). They have been taking a very liberal stand when it comes to the market and competitors so as to be more compatible to the users. More innovations will take the company to great heights in the market around the globe. The biggest advantage of this software is that many companies will prefer them for one reason that they can be easily run in Microsoft application and platforms. 2) Oracle: They have been in the limelight ever since they have purchased People soft ERP software. Another
Velammal College of Engineering and Technology, Madurai
Page 244
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
encouraging fact for them is the mass campaign carried by UNDP (United Nations Development Program) to create an awareness of their products. They have also satisfies the customers (of both people soft and oracle) by offering many competitive features. 3) SAP: A specialized ERP product meant to solve technical and managerial issues. SAP ERP continues to be the choice of many companies around the world. It helps companies to redefine their business operations. Some of the features and advantages of this software program are at par with industry standards though some others require to be improved in order to make it compatible to the end user. 4) PeopleSoft: People soft ERP software program helps the companies to streamline their HR functions in an easy and effective manner. This software continues to be the undisputed leader in the market when it comes to HR operations anywhere in the globe. Oracles recent takeover of People soft has only been a boon to its customers who are now fable to access the e-Services of Oracle also as a result of this merger. 5) UPS: This ERP software has an important feature. They can work in combination with other applications like SAP, Oracle. This will enable them to get the best features from each software and put them in use as may be demanded. This has also hogged the limelight due to this advantage.
VI CONCLUSION The "soft" benefits of ERP provide good bottom-line savings and sustainable top-line growth, lower inventory and operating cost, more efficient production, improved quality with less waste, higher customer acquisition retention and better cash flow. The future of ERP holds an undisputed demand not only in the national level but also at the global level. If the technology can be improvised to the desired extent. ERP trends reflect positive signals for the ERP vendors and companies availing their service. It is important to remember the fact that both the vendor and the company will be able to make use of any advantage (including the modern facilities) only through proper coordination, teamwork and nurturing a cordial atmosphere. Mere IT ERP trends will not help in this aspect. Industry consultant Reed sums it up this way: "'Empower me. Give me the tools to create differentiating processes that allow me to define myself from my competitors. And make sure that it's easier for me to do, so I don't have to hire 100 programmers. Give me the building blocks to put that together quickly, so that it's just humming in the background, and leave me free to focus on what makes us better than other companies.' That's what customers are expecting now and really want”. REFERENCES www.erpwire.com www.whitepaper.techrepublic.com www.olcsoft.com www.eioupdate.com www.sdn.sap.com www.bscaler.com [7] www.plex.com [8] www.itmanagement.earthweb.com
[1] [2] [3] [4] [5] [6]
Velammal College of Engineering and Technology, Madurai
Page 245
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A New Semantic Similarity Metric for Handling all Relations in WordNet Ontology K.Saruladha#1, Dr.G.Aghila*2, Sajina Raj#3 #
Research Scholar, 3Post Graduate Student, Department of Computer Science & Engg. Pondicherry Engineering College, Pondicherry 605 014, India. 1
[email protected] 3 [email protected] *
Head of the department Department of Computer Science, Pondicherry University, Pondicherry 605 014, India. 2 [email protected] Abstract- Semantic similarity assessing methods play a central role in many research areas such as Psychology, cognitive science, information retrieval biomedicine and Artificial intelligence. This paper discuss the existing semantic similarity assessing methods and identify how these could be exploited to calculate accurately the semantic similarity of WordNet concepts. The semantic similarity approaches could broadly be classified into three different categories: Ontology based approaches (structural approach), information theoretic approaches (corpus based approach) and hybrid approaches. All of these similarity measures are expected to preferably adhere to certain basic properties of information. The survey revealed the following drawbacks The information theoretic measures are dependent on the corpus and the presence or absence of a concept in the corpus affects the information content metric. For the concepts not present in the corpus the value of information content tends to become zero or infinity and hence the semantic similarity measure calculated based on this metric do not reflect the actual information content of the concept. Hence in this paper we propose a new information content metric which provides a solution to the sparse data problem prevalent in corpus based approaches. The proposed measure is corpus independent and takes into consideration hyponymy and meronymy relations. Further the information content metric used earlier by Resnik, lin and Jiang and Cornath methods may produce better results with alternate corpora other than brown corpus. Hence the effect of corpus based information content metric on alternate corpora is also investigated. The proposed similarity function with noun pairs of R&G data set was tested for Resnik method and the correlation coefficient with human judgment were calculated. The results were promising.
Velammal College of Engineering and Technology, Madurai
Keywords-Ontology, similarity method, information retrieval, conceptual similarity, taxonomy, corpus based
I . Introduction The goal of Information retrieval process is to retrieve Information relevant to a given query request. The aim is to retrieve all the relevant information eliminating the nonrelevant information. An information retrieval system comprises of document representation, semantic similarity matching function and Query. Document representation comprises the abstract description of documents in the system. The semantic similarity matching function defines how to compare query requests to the stored descriptions in the representation. The percentage of relevant information we get mainly depends on the semantic similarity matching function we used. So far, there are several semantic similarity methods used which have certain limitations despite the advantages. No one method replaces all the semantic similarity methods. When a new information retrieval system is going to be build, several questions arises related to the semantic similarity matching function to be used. In the last few decades, the number of semantic similarity methods developed is high. This paper discusses the overall view of different existing similarity measuring methods used for ontology concept comparison. We also discuss about the pros and cons of existing similarity metrics. We have presented a new approach which is independent of the corpora for finding the semantic similarity between two concepts.
Page 246
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Section II, discusses the basic qualitative properties that a similarity measure is expected to satisfy. Section III discusses various approaches used for similarity computation and the limitations of those methods. In Section IV, comparison among different semantic similarity measures are discussed. In Section V we introduce a new semantic similarity approach and an algorithm for finding the similarity between all the relations in the WordNet taxonomy. The results based on new similarity measure is promising.
II. ONTOLOGY SIMILARITY In this section, a set of intuitive and qualitative properties that a similarity method should adhere to is discussed.[20] A. Basic Properties Any similarity measure must be compatible with the basic properties as they express the exact notion of property. o CommonalityProperty o Difference Property o Identity Property B. Retrieval Specific Properties The similarity measure cannot be symmetrical in case of ontology based information retrieval context. The similarity is directly proportional to specialization and inversely proportional to generalization. o Generalization Property C. Structure Specific Properties The distance represented by an edge should be reduced with an increasing depth. o Depth Property o Multiple Paths Property III. APPROACHES USED FOR SIMILARITY COMPUTATION In this section, we discuss about various similarity methods[20]. The similarity methods are A. B. C. D.
Path Length Approaches Depth-Relative Approaches Corpus-based Approaches Multiple-Paths Approaches
A. Path Length Approach The shortest path length and the weighted shortest path are the two taxonomy based approaches for measuring similarity through inclusion relation. 1) Shortest Path Length
Velammal College of Engineering and Technology, Madurai
A simple way to measure semantic similarity in a taxonomy is to evaluate the distance between the nodes corresponding to the items being compared. The shorter distance results in high similarity. In Rada et al. [1989][1][14], shortest path length approach is followed assuming that the number of edges between terms in a taxonomy is a measure of conceptual distance between concepts. distRada(ci; cj) = minimal number of edges in a path from ci to cj This method yields good results. since the paths are restricted to ISA relation, the path lengths corresponds to conceptual distance. Moreover, the experiment has been conducted for specific domain ensuring the hierarchical homogeneity. The drawback with this approach is that, it is compatible only with commonality and difference properties and not with identity property. 2) Weighted Shortest Path Length This is another simple edge-counting approach. In this method, weights are assigned to edges. In brief, weighted shortest path measure is a generalization of the shortest path length. Obviously it supports commonality and difference properties. - Similarity of immediate specialisation -Similarity of immediate generalisation P=(p1,…..,pn) where, Pi ISA pi+1 or Pi+1 ISA pi For each I with x=p1 and y=pn Given a path P=(p1,…..pn), set s(P) to the number of specializations and g(P) to the number of generalizations along the path P as follows: s(P)= |{i\pi ISA pi+1}| (1) g(P)=|{i|Pi+1 ISA pi}| (2) If p1,……pm are all paths connecting x and y, then the degree to which y is similar to x can be defined as follows: simWSP(x,y)=max{ s(pj) s(pj)} (3) j=1,….m The similarity between two concepts x and y, sim(x,y) WSP(weighted Shortest Path) is calculated as the maximal product of weights along the paths between x and y. Similarity can be derived as the products of weights on the paths. s(pj)
g(Pj
j
j
s(P ) and g(P ) = 0 Hence the weighted shortest path length overcomes the limitations of shortest path length wherein the measure is based on generalization property and achieves identity property. B. Depth-Relative Approaches Even though the edge counting method is simple, it limits the representation of uniform distances on the edges in the
Page 247
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
taxonomy. This approach supports structure specific property as the distance represented by an edge should be reduced with an increasing depth. This approach includes depth relative scaling method, conceptual similarity method and normalization based length method. 1) Depth Relative Scaling In his depth-relative scaling approach Sussna[1993][2] defines two edges representing inverse relations for each edge in a taxonomy. The weight attached to each relation r is a value in the range [minr; maxr]. The point in the range for a relation r from concept c1 to c2 depends on the number nr of edges of the same type, leaving c1, which is denoted as the type specifc fanout factor: W(c1→r c2)=maxr-{maxr--minr/nr(c1)} The two inverse weights are averaged and scaled by depth d of the edge in the overall taxonomy. The distance between adjacent nodes c1 and c2 are computed as: dist sussna(c1,c2)=w(c1→ r c2)+ (c1→ r’ c2)/2d (4) where r is the relation that holds between c1 and c2, and r’ is its inverse. The semantic distance between two arbitrary concepts c1 and c2 is computed as the sum of distances between the pairs of adjacent concepts along the shortest path connecting c1 and c2. 2) Conceptual Similarity Wu and Palmer [1994][3], propose a measure of semantic similarity on the semantic representation of verbs in computer systems and its impact on lexical selection problems in machine translation. Wu and Palmer define conceptual similarity between a pair of concepts c1 and c2 as: Sim wu&palmer(c1,c2)= (5) Where N1 is the number of nodes on the path from c1 to a concept c3. , denoting the least upper bound of both c1 and c2. N2 is the number of nodes on a path from c2 to c3. N3 is the number of nodes from c3 to the most general concept. 3) Normalised Path Length Leacock and Chodorow [1998][4], proposed an approach for measuring semantic similarity as the shortest path using is a hierarchies for nouns in WordNet. This measure determines the semantic similarity between two synsets (concepts) by finding the shortest path and by scaling using the depth of the taxonomy: Sim Leacock&Chaodorow(c1,c2)= -log(Np(c1,c2)/2D) (6) Np (c1,c2) denotes the shortest path between the synsets (measured in nodes), and D is the maximum depth of the taxonomy. C. Corpus-based Approach
Velammal College of Engineering and Technology, Madurai
The knowledge disclosed by the corpus analysis is used to intensify the information already present in the ontologies or taxonomies. In this method, presents three approaches viz information content method, J&C method and Lin’s universal similarity measure is discussed. All these methods incorporate corpus analysis as an additional, and qualitatively different knowledge source. 1) Information Content In this method rather than counting edges in the shortest path, they select the maximum information content of the least upper bound between two concepts. Resnik [1999] [5], argued that a widely acknowledged problem with edgecounting approaches was that they typically rely on the notion that edges represent uniform distances. According to Resnik's measure, information content, uses knowledge from a corpus about the use of senses to express nonuniform distances. Let C denote the set of concepts in a taxonomy that permits multiple inheritance and associates with each concept c 2 C, the probability p(c) of encountering an instance of concept c. For a pair of concepts c1 and c2, their similarity can be defined as: SimResnik (7) where, S(c1,c2): Set of least upper bounds in the taxonomy of c1 and c2 p(c) :Monotonically non-decreasing as one moves up in the taxonomy, p(c1) ≤ p(c2), if c1 is a c2. The similarity between the two words w1 and w2 can be computed as: wsimResnik (8) Where, s(wi): Set of possible senses for the word wi. Resnik describes an implementation based on information content using WordNet's [Miller, 1990][6], taxonomy of noun concepts [1999]. The information content of each concept is calculated using noun frequencies Freq(c) = where, words(c): Set of words whose senses are subsumed by concept c. (c)=freq(c)/N where , N: is the total number of nouns. The major drawback of the information content approach is that they fail to comply with the generalization property, due to symmetry. 2) Jiang and Conrath's Approach(Hybrid Method) Jiang and Conrath [1997][7]proposed a method to synthesize edge-counting methods and information content
Page 248
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
into a combined model by adding the information content as a corrective factor. The edge weight between a child concept cc and a parent concept cp can be calculated by considering factors such as local density in the taxonomy, node depth, and link type as, Wt(cc,cp)= (9) Where, d(cp) : Depth of the concept cp in the taxonomy, E(cp) :Number of children of cp (the local density) (¹E) : Average density in the entire taxonomy, LS(cc, cp) : Strength of the edge between cc and cp, T(cc,cp) : Edge relation/type factor The parameters 0 and ,0 1 control the in°uence of concept depth and density, respectively. strength of a link LS(cc; cp) between parent and child concepts is proportional to the conditional probability p(cc|cp) of encountering an instance of the child concept, cc, given an instance of the parent concept, cp: LS(cc, cp) = -log p(cc|cp) Resnik assigned probability to the concepts as p(cc \cp) = p(cc), because any instance of a child concept cc is also an instance of the parent concept cp. Then: p(Cc|Cp) = p(Cc|Cp) = If IC(c) denotes the information content of concept c. then: LS(Cc,Cp) = IC(Cc)-IC(Cp) Jiang and Conrath then defined the semantic distance between two nodes as the summation of edge weights along the shortest path between them [J. Jiang, 1997]: distjiang&conrath(C1,C2)= (10) Where, path(c1, c2) : the set of all nodes along the shortest path between c1 and c2 parent(c) : is the parent node of c LSuper(c1, c2) : is the lowest superordinate (least upper bound) on the path between c1 and c2.. Jiang and Conrath's approach made information content compatible with the basic properties and the depth property, but not to the generalization property. 3) Lin's Universal Similarity Measure Lin [1997; 1998][8][13] defines a measure of similarity claimed to be both universally applicable to arbitrary objects and theoretically justified. He achieved generality from a set of assumptions. Lin's information-theoretic definition of similarity builds on three basic properties, commonality, difference and identity. In addition to these properties he assumed that
Velammal College of Engineering and Technology, Madurai
the commonality between A and B is measured by the amount of information contained in the proposition that states the commonalities between them, formally: I(common(A,B)) = -log p(common(A,B) where, I(s): Negative logarithm of the probability of the proposition, as described by Shannon[1949]. The difference between A and B is measured by: I(description(A,B)) – I(common(A,B)) Where, description(A;B) : Proposition about what A and B are. Lin proved that the similarity between A and B is measured by, simLin(A,B)= (11) The ratio between the amount of information needed to state the commonality of A and B and the information needed to describe them fully. Lin’s similarity between two concepts in a taxonomy ensures that: SimLin(c1,c2)= (12) where, LUB(c1, c2): Least upper bound of c1 and c2 p(x) : Estimated based on statistics from a sense tagged corpus. This approach comply with the set of basic properties and the depth property, but would fail to comply with the generalization property as it is symmetric. D. Multiple-Paths Approaches This approach solves the problem with single path approach. Single path as a measure for the similarity, fails to truly express similarity whenever the ontologies allow multiple inheritance. In multiple-path approach measurement is made by taking into account all the semantic relations in ontologies, considering more than one path between concepts. Attributes should influence the measure of similarity, thus allowing two concepts sharing the same attribute to be considered as more similar, compared to concepts not having this particular attribute. This approach includes Hirst & St-Onge method, generalized shortest path method, shared nodes and weighted shared node methods. 1) Medium-Strong Relations Hirst and St-Onge [Hirst and St-Onge, 1998; St-Onge, 1995] [9][15], distinguishes the nouns in the Wordnet as extra-strong, strong and medium-strong relations. The extra-strong relation is only between a word and its literal repetition. A strong relation between two words exists if: 1. They have a synset in common
Page 249
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2. There is a horizontal link (antonymy, similarity) between a synset of each word 3. There is any kind of link at all between a synset of each word or if one word is a compound or phrase that includes the other word. Medium-strong exists between words when a member of a set of allowable paths connects the two words. A path is allowable if it contains no more than five edges and conforms to one of the eight patterns described in Hirst and St-Onge [1998]. 2) Generalised Weighted Shortest Path The principle of weighted path similarity can be generalized by introducing similarity factors for the semantic relations. However, there does not seem to be an obvious way to differentiate based on direction. Thus, we can generalize simply by introducing a single similarity factor and simplify to bidirectional edges. This method solves the symmetry problem by introducing weighted edges. 3) Shared Nodes This approach overcomes the limitation of single path length approach. Multiple paths are considered for measuring the similarity. The shared nodes approach with similarity function discussed above complies with all the defined properties. 4) Weighted Shared Nodes Similarity It is found that when deriving similarity using the notion of shared nodes, not all nodes are equally important. Assigning weights to edges is very important, as it generalizes the measure so that it can be make use for different domains with different semantic relations. It also allows differentiating between the key ordering relation, ISA and the other semantic relations when calculating similarity. The weighted shared nodes measure complies with all the defined properties. IV. COMPARISON OF DIFFERENT SIMILARITY MEASURES In this section we discuss about the results of comparison of the measures to human similarity judgments. The first human similarity judgment was done by Rubinstein and Goodenough [1965][11], using two groups totaling 51 subjects to perform synonymy judgments on 65 pairs of nouns and this in turn been the basis of the comparison of similarity measures.
Velammal College of Engineering and Technology, Madurai
Miller and Charles [1991][12] repeated Rubinstein and Goodenough's original experiment, they used a subset of 30 noun pairs from the original list of 65 pairs, where ten pairs were from the high level of synonymy, ten from the middle level and ten from the low level. TABLE I REPLICA OF THE RUBINSTEIN AND GOODENOUGH AND THE MILLER AND CHARLES EXPERIMENTS
Word1
Word2
Replica
R&G
M&C
Car
Automobile
3.82
3.92
Gem
Jewel
3.86
3.84
3.92 3.94
Journey
Voyage
3.58
3.54
3.58
Boy
Lad
3.10
3.76
3.84
Coast
Shore
3.38
3.70
3.60
Asylum
madhouse
2.14
3.61
3.04
Magician
Wizard
3.68
3.50
3.21
Midday
Noon
3.45
3.42
3.94
Furnace
Stove
2.60
3.11
3.11
Food
Fruit
2.87
3.08
2.69
Bird
Cock
2.62
3.05
2.63
Bird
Crane
2.08
2.97
2.63
Tool
implement
1.70
2.95
3.66
Brother
Monk
2.38
2.82
2.74
Lad
Brother
1.39
1.66
2.41
Crane
Implement
1.26
1.68
2.37
Journey
Car
1.05
1.16
1.55
Monk
Oracle
0.90
1.10
0.91
Cemetery
Woodland
0.32
0.95
1.18
Food
Rooster
1.18
0.89
Coast
Hill
1.24
0.87
1.09 1.26
Forest
Graveyard
0.41
0.84
Shore
Woodland
0.81
0.63
0.90
Monk
Slave
0.36
0.55
0.57
Coast
Forest
0.70
0.42
0.85
Lad
Wizard
0.61
0.42
0.99
Chord
Smile
0.15
0.13
0.02
Glass
Magician
0.52
0.11
0.44
Rooster
Voyage
0.02
0.08
0.04
Noon
String
0.02
0.08
0.04
1.00
Page 250
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The correlation between these two experiments is 0.97. An experiment were performed by taking the replica of the Miller and Charles experiment that included 30 concept pairs and an additional ten new compound concepts pairs and human judgment for all the 40 pairs has been performed. The correlations for the measures done by Resnik, Jiang and Conrath; Lin, Hirst and St-Onge; and Leacock and Chodorow are shown in table given below. [Budanitsky, 2001][18][19][20] TABLE II CORRELATION BETWEEN DIFFERENT SIMILARITY MEASURES & HUMAN SIMILARITY JUDGMENTS FROM THE MILLER AND CHARLES EXPERIMENT
The collection of more than 1 million words was manually tagged with about 80 parts of speech. The presently available list of WordNet concepts tagged in the brown corpus includes approximately 420000 words. There are more number of concepts in WordNet which are not tagged in Brown Corpus. We ran a small experiment on the brown corpus and found that the words soccer, fruitcake, world trade centre, CPU etc are not found in the Brown corpus. Information content of a concept in corpus based approaches is calculated using the formula IC(C)=-log p(c) For the nouns for which p(c) value is zero the IC value becomes zero or infinity. This issue is addressed as sparse data problem. A concept (noun) other than non informative words(stop words) is expected to have information content value>0. But the existing IC calculation method returns a zero for a concept(noun) which is unacceptable.
Approach
Correlation
Resnik
0.744
Jiang and Conrath
0.850
Lin
0.829
Hirst and St-Onge
0.744
BROWN CORPUS CATEGORIES Press : reportage (Political Sports,Society,Spot
Leacock and Chodorow
0.816
News,Financial,Cultural
TABLE IV
Press: Editorial (Institutional Daily, Personal, Letters to the
The table given below shows the correlations between the replica and the two previous experiments. The correlations between the replica experiment and the previous experiments are fairly good[20]. TABLE III CORRELATION BETWEEN THE THREE HUMAN SIMILARITY JUDGMENT EXPERIMENTS Correlation Rubinstein Goodenough Rubinstein & Goodenough Miller & Charles
Editor) Press: Reviews (Theatre, Books, Music, Dance) Religion (Books, Periodicals, Tracts) Skill and Hobbies (Books, Periodicals) Popular Lore (Books, Periodicals) Belles-Lettres (Books, Periodicals)
Miller & Charles
0.97
Replica
0.93
Replica
0.95
Miscellaneous: US Government & House Organs (Government Documents, Foundation Reports, College Catalog, Industry House organ) Learned (Natural Sciences, Medicine, Mathematics, Social and Behavioral Sciences, Political Science, Law, Education,
A. Sparse Data Problem in Corpus based Approach Following the standard definition from Shannon and Weaver’s information theory [1949], the information content of c is −log p(c). Information content in the context of WordNet is drawn from the brown university standard corpus. This corpus refers to a collection of documents of widely varying genres collected in 1961 which was updated in 1971and 1979 to reflect new literatures shown in table IV.
Humanities, Technology and Engineering) Fiction: General (Novels, Short Stories) Fiction: Mystery and Detective Fiction (Novels, Short Stories) Fiction: Adventure and Western (Novels, Short Stories) Fiction: Romance and Love Story (Novels, Short Stories) Humor (Novels, Essays)
B. Significance of the work We have proposed an algorithm based on new information content which takes into consideration the meronymy
Velammal College of Engineering and Technology, Madurai
Page 251
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
relations. The Information content metric proposed by Pierro [2009] takes into concern the holonym relations that a concept has with other concepts. All of the concepts of the WordNet will not have all of the relation types. Some of the concepts will have holonym relation and few with meronym relations etc. The information content metric which considers holonymy relation alone, if it ignores the other type of relations and if it produces a zero value, it means that the concept has no information which is not correct. Hence we decided to consider both the meronym and holonym relations. If a concept do not have holonymy relation then the information imparted by meronomy is taken into consideration and the information content will become zero when the concept has no relations with other concepts which is a very rare case.
2. To investigate the existing IC metric on corpora other than brown corpus. TABLE V WORDNET MERONYMY/HOLONYMY RELATION Relation
Example
Meronym(Part-of)
Engine is a Meronym of Car
Holonym(Has-part) Car is a Holonym of Engine
Meronym(Has-Member)
Team has member player
Holonym Player is the member of Team (Member-of)
V. PROPOSED WORK The semantic similarity measures are mostly based on the information content. Most of the corpus based similarity methods like Lin[8], Jiang Cornath[7] and Resnik[5] are IC based and the IC calculation is done using Brown corpus. All concepts of WordNet are not present in the Brown Corpus. The noun such as autograph, serf and slave are not present in the Brown Corpus. Similarity measures that rely on information content can produce a zero value for even the most intuitive pairs because the majority of WordNet concepts occur with a frequency of zero. This makes the Lin method and Jiang cornath method to return zero or infinity in the continuous domain and hence the similarity measure is not true or reliable. Hence the computation of Information content should be computed in a different way so that the similarity measure becomes reliable.
A. Proposed Corpora Independent Measure to Solve Sparse Data Problem
Similarity
The objective of this work is twofold. 1. To design a new information metric which solves sparse data problem which is corpus independent and is based on semantic information available in well known information source like WordNet.
Velammal College of Engineering and Technology, Madurai
The Existing Similarity methods do not consider the Holonymy/Meronymy relationships defined in Wordnet. Hence we propose to devise a new similarity measure which considers these relationships and experimentally evaluate and compare it with the existing similarity measures using R&G data set and extended data set. The fig.1 shows architecture. Google based corpus Reuters Corpus Brown Corpus
the
semantic
Compute Frequency of Wordnet Concepts Semantic Similarity computation Resnik Jiang & Conrath Lin Pirro&Seco
similarity
system
Human Judgments for Extended R&G Data set
Compute Correlation Coefficient
WordNet Taxonomy Rubenstein & Goodenough Extended DataSet
Performance Analysis
Fig. 1 Semantic Similarity System Architecture
Page 252
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
B. Proposed IC Calculation Involving Meronym
IC (LCS (c1,c2))--hyper
The intrinsic IC for a concept c is defined as: IC(c) = 1- log(hypo(c)+mero(c)+1
---(16)
Log(maxcon) where, Hypo(c)-Function returns the number of hyponyms Mero(c)-Function returns the number of meronyms Maxcon- is a constant that indicates the total number of concepts in the considered taxonomy Note: Assume hypo(c), mero(c)>=0 & maxcon>0 The function hypo and mero returns the number of hyponyms and meronyms of a given concept c. Note that concepts representing leaves in the taxonomy will have an IC of one, since they do not have hyponyms. The value of one states that a concept is maximally expressed and cannot be further differentiated. Moreover maxcon is a constant that indicates the total number of concepts in the considered taxonomy.
(c1 c2)
----(17)
where, the function LCS finds the lowest common subsumer of the two concepts c1 and c2 and the function hyper finds all the hypernyms of c1 and c2 upto the LCS node. The Proposed IC formula will be used in existing semantic similarity methods such as Resnik, Lin and Jiang and cornath for computation of similarity measure using R&G and M&C data sets. The influence of meronymy and hyponymy relations in calculation of similarity measure will be studied. We will test the existing Resnik, Lin and Jiang cornath method using alternate corpora like reuters and Google Based Corpora. Some researches have taken this issue into consideration and have proposed corpus independent information content calculation. But those calculation takes into consideration only the holonymy relations [21] excluding other semantic relations that exists between concepts.
C. Proposed Similarity Function Based On Proposed IC According to our new formula, the similarity between the two concepts c1 and c2 can be defined as, Simext(c1,c2) =
Velammal College of Engineering and Technology, Madurai
Page 253
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Input (noun1, noun2), Output (similarity score) Step 1: For each sense c1 and c2 in noun1 and noun2 If c1 and c2 are hyponyms Calculate IC (c1) for hyponyms i.e
VI. Conclusion This paper has discussed the various approaches that could be used for finding similar concepts in an ontology and between ontologies. We have done a survey to exploit the similarity methods for ontology based query expansion to aid better retrieval effectiveness of Information retrieval models. The experiments conducted by early researches provide better correlation values which gives promising direction of using them in Ontology based retrieval models. A new semantic similarity metric has been introduced which overcomes the shortcomings of existing semantic similarity methods mainly sparse data problem. We are working with the new similarity function which will combine the advantages of the similarity methods discussed in this paper and we will test it with ontologies of particular domain. Since we are considering all the relationship present in the WordNet taxonomy, the new metric gives accurate results.
IC(c) = 1- log(hypo(c)+1) log(maxcon) go to step 3 Else if c1 and c2 are meronyms Calculate IC(c1) for meronyms i.e IC(c) = 1- log(mero(c)+1) log(maxcon) go to step 3 Step 2: For IC value for both hyponyms and meronyms using the proposed IC formula,
VII. REFERENCES IC(c) = 1- log(hypo(c)+mero(c)+1) [1] Roy Rada, H. Mili, Ellen Bicknell, and M. Blettner. “Development and application of a metric on semantic nets” IEEE Transactions on Systems, Man, and Cybernetics, 19(1), 17{30}, January 1989. [2] Michael Sussna: “Word sense disambiguation for tree-text indexing using a massive semantic network” In Bharat Bhargava [3]
Zhibiao Wu and Martha Palmer. “Verbs semantics and lexical
selection”, In Proceedings of the 32nd annual meeting on Association for Computational Linguistics”, Association for Computational Linguistics, 1994. pages 133{138, Morristown, NJ, USA [4] Claudia Leacock and Martin Chodorow: “Combining local context and wordnet similarity for word sense identification”. In Christiane Fellbaum, editor, WordNet: An Electronic Lexical Database. MIT Press, 1998. [5] Philip Resnik.: “Semantic similarity in a taxonomy: An information-based measure and its application to problems of ambiguity in natural language”, 1999. [6] George A. Miller, Richard Beckwith, Christiane Fellbaum, Derek Gross, and K.J. Miller.[Resnik, 1995] Philip Resnik. “Using information content to evaluate semantic similarity in a taxonomy”. In IJCAI, pages 448{453, 1995. [7] D. Conrath J. Jiang.: “Semantic similarity based on corpusstatistics and lexical taxonomy”. In Proceedings on International Conference on Research in Computational Taiwan, pages 19{33, 1997.
Linguistics,
Velammal College of Engineering and Technology, Madurai
log(maxcon) go to step 3 Step 3: Call the existing semantic similarity function Simres(c1,c2), SimLin(c1,c2), SimJ&C(c1,c2) And then go to step 4. Step 4: Call the proposed semantic similarity function for the given concepts c1 & c2 Simext(c1,c2) = IC(LCS(c1,c2))hyper(c1 c2) Step 5: Collect human judgments and save it as a separate table for the R&G and M&C data sets Step 6: Calculate the correlation coefficients between results of the similarity measures and human judgments Step 7: Compare the similarity measures for R&G data set using the proposed IC and proposed similarity existing similarity measures.
Page 254
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[8]
Dekang
Lin.:
“An
information-theoretic
definition
of
A
similarity”. In Jude W. Shavlik, editor, ICML, pages 296{304. Morgan Kaufmann, 1998. ISBN 1-55860-556-8. [9]
Graeme
Hirst and
David St-Onge: “Lexical chains as detection and correction
representation of context for the
malapropisms”In Christiane Fellbaum, editor, WordNet: An Electronic Lexical Database. MIT Press, 1998. [10] Troels Andreasen, Rasmus Knappe, and Henrik Bulskov: ” Domain-specifiic similarity and retrieval”. In Yingming Liu, GuoqingChen, and Mingsheng Ying, editors, 11th International FuzzySystems Association World Congress, volume 1, pages 496{502,Beijing,China, 2005. IFSA 2005, Tsinghua University Press. [11] Rubinstein and Goodenough, H. Rubinstein and J. B. a Goodenough: “Contextual correlates of synonymy”. Communications of the ACM, 8(10), 1965. [12] George A. Miller and W. G. Charles: “Contextual correlates of semantic similarity”. Language and Cognitive Processes, 6(1):1{28, 1991. [13] Dekang Lin: “Using syntactic dependency as local context to resolve word sense ambiguity”. In ACL, pages 64{71, 1997. [14] Roy Rada and Ellen Bicknell: “Ranking documents with a thesaurus”. JASIS, 40(5):304{310,1989.Timothy Finin, and Yelena Yesha, editors, Proceedings of the 2nd International Conference on Information and Knowledge Management, pages 67{74,New York, NY, USA, November 1993. ACM Press. [15] Alexander Budanitsky: “University of Toronto Graeme Hirst, University of Toronto Evaluating WordNet-based Measus Of Lexical Semantic Relatedness”, Association for Computational Linguistics,2006 [16] Philip Resnik: “Using Information Content to Evaluate Semantic Similarity in a Taxonomy”, Sun Microsystems Laboratories, 1995 [17]
Henrik Bulskov Styltsvig: “Ontology-based Information Retrieval Computer Science Section Roskilde University”, 1996
[18] A. Budanitsky. “Lexical semantic relatedness and its application in natural language processing”, 1999. [19] A. Budanitsky. Semantic distance in wordnet: An experimental, application-oriented evaluation of ¯ve measures, 2001. [20] H. Bulskov: “Ontology-based Information Retrieval”, PhD Thesis Andreasen, T., Bulskov, H., & Knappe, R. (2006). [21] N. Seco, T. Veale, J. Hayes: “ An intrinsic information content metric for semantic similarity in WordNet, in : Proceedings of ECAI, 2004, pp. 1089–1090. [22] Giuseppe Pirró: “A semantic similarity metric combining features and intrinsic information content”, Data & Knowledge Engineering 68(2009) 1289–1308 2009.
Velammal College of Engineering and Technology, Madurai
Page 255
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fault Prediction Using Conceptual Cohesion in Object Oriented System V.Lakshmi#1,P.V.Eswaripriya*2 ,C.Kiruthika*3 and M.Shanmugapriya*4 #1
Faculty of Kamaraj College of Engg and Technology, Virudhunagar, Tamilnadu. 1
*2
[email protected]
. Final Year CSE,Kamaraj College of Engg and Technology, Virudhunagar Tamilnadu. 2
*3
[email protected]
. Final Year CSE,Kamaraj College of Engg and Technology, Virudhunagar Tamilnadu *4 . Final Year CSE,Kamaraj College of Engg and Technology, Virudhunagar Tamilnadu Abstract-High cohesion is desirable property in software systems to achieve reusability and maintainability. In this project we are measures for cohesion in Object-Oriented (OO) software reflect particular interpretations of cohesion and capture different aspects of it. In existing approaches the cohesion is calculate from the structural information for example method attributes and references. In conceptual cohesion of classes, i.e. in our project we are calculating the unstructured information from the source code such as comments and identifiers. Unstructured information is embedded in the source code. To retrieve the unstructured information from the source code Latent Semantic Indexing is used. A large case study on three open source software systems is presented which compares the new measure with an extensive set of existing metrics and uses them to construct models that predict software faults. In our project we are achieving the high cohesion and we are predicting the fault in Object Oriented Systems. Keywords- Software cohesion, textual coherence, prediction, fault proneness,, Latent Semantic indexing.
fault
I.INTRODUCTION Software modularization, Object oriented decomposition is an approach for improving the organization and comprehension of source code. In order to understand OO software, software engineers need to create a well connected representation of the classes that make up the system. Each class must be understood individually and, then, relationships among classes as well. One of the goals of the OO analysis and design is to create a system where classes have high cohesion and there is low coupling among them. These class properties facilitate comprehension, testing, reusability, maintainability, etc. A.Software Cohesion Software cohesion can be defined as a measure of the degree to which elements of a module belong together. Cohesion is also regarded from a conceptual point of view. In this view, a cohesive module is a crisp abstraction of a concept or feature from the problem domain, usually described in the requirements or specifications. Such
Velammal College of Engineering and Technology, Madurai
definitions, although very intuitive, are quite vague and make cohesion measurement a difficult task, leaving too much room for interpretation. In OO software systems, cohesion is usually measured at the class level and many different OO cohesion metrics have been proposed which try capturing different aspects of cohesion or reflect a particular interpretation of cohesion. Proposals of measures and metrics for cohesion abound in the literature as software cohesion metrics proved to be useful in different tasks, including the assessment of design quality, productivity, design, and reuse effort, prediction of software quality, fault prediction modularization of software and identification of reusable of components. B.Approaches for Cohesion Most approaches to cohesion measurement have automation as one of their goals as it is impractical to manually measure the cohesion of classes in large systems. The tradeoff is that such measures deal with information that can be automatically extracted from software and analyzed by automated tools and ignore less structured but rich information from the software (for example, textual information). Cohesion is usually measured on structural information extracted solely from the source code (for example, attribute references in methods and method calls) that captures the degree to which the elements of a class belong together from a structural point of view. These measures give information about the way a class is built and how its instances work together to address the goals of their design. The principle behind this class of metrics is to measure the coupling between the methods of a class. Thus, they give no clues as to whether the class is cohesive from a conceptual point of view (for example, whether a class implements one or more domain concepts) nor do they give an indication about the readability and comprehensibility of the source code. Although other types of metrics were proposed by researchers to capture
Page 256
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
different aspects of cohesion, only a few metrics address the conceptual and textual aspects of cohesion. We propose a new measure for class cohesion, named the Conceptual Cohesion of Classes (C3), which captures the conceptual aspects of class cohesion, as it measures how strongly the methods of a class relate to each other conceptually. The conceptual relation between methods is based on the principle of textual coherence. We interpret the implementation of methods as elements of discourse. There are many aspects of a discourse that contribute to coherence, including co reference, causal relationships, connectives, and signals. The source code is far from a natural language and many aspects of natural language discourse do not exist in the source code or need to be redefined. The rules of discourse are also different from the natural language.
• Lack of cohesion implies classes should probably be split into two or more sub/classes. • Any measure of disparateness of methods helps identify flaws in the design of classes. • Low cohesion increases complexity, thereby increasing the likelihood of errors during the development process.
C3 is based on the analysis of textual information in the source code, expressed in comments and identifiers. Once again, this part of the source code, although closer to natural language, is still different from it. Thus, using classic natural language processing methods, such as propositional analysis, is impractical or unfeasible. Hence, we use an Information Retrieval (IR) technique, namely, Latent Semantic Indexing (LSI), to extract, represent, and analyze the textual information from the source code. Our measure of cohesion can be interpreted as a measure of the textual coherence of a class within the context of the entire system. Cohesion ultimately affects the comprehensibility of source code. For the source code to be easy to understand, it has to have a clear implementation logic (that is, design) and it has to be easy to read (that is, good language use). These two properties are captured by the structural and conceptual cohesion metrics, respectively.
III.PROPOSED C3
II.RELATED WORKS A.STRUCTURAL METRICS We study the cohesion metric LCOM B.Henderson-Sellers [2] of object oriented programs. The LCOM metric is based on the assumption, that if methods and instance variables of a class are interconnected, then the class is cohesive. If a method x uses instance variable y, then x and y are interconnected. Henderson-Sellers defined the metric LCOM* that is similar to LCOM, but has a fixed scale. Briand.et al observed that the scale of LCOM* is from 0 to 2, and gave a refined version of this metric with scale from 0 to 1. Other similar cohesion metrics are LCC and TCC, and CBMC. Good surveys of cohesion metrics have been made by Briand et al and Chae et al . • Cohesiveness of methods within a class is desirable, since it promotes encapsulation.
Velammal College of Engineering and Technology, Madurai
B.Formula for Lcom5 lcom=((((1/a)*Mu)-m)/deno); Mu-count for fields, m-methods length, a-fields Length, deno=1-m; lcom51=lcom*lcom; lcom5=Math.sqrt (lcom51);
LSI [4] is a corpus-based statistical method for inducing and representing aspects of the meanings of words and passages (of the natural language) reflective of their usage in large bodies of text. LSI is based on a Vector Space Model (VSM) as it generates a real-valued vector description for documents of text. Results have shown that LSI captures significant portions of the meaning not only of individual words but also of whole passages, such as sentences, paragraphs, and short essays. The central concept of LSI is that the information about the contexts in which a particular word appears or does not appear provides a set of mutual constraints that determines the similarity of meaning of sets of words to each other. LSI was originally developed in the context of IR as a way of overcoming problems with polysemy and synonymy that occurred with VSM approaches. Some words appear in the same contexts and an important part of word usage patterns is blurred by accidental and inessential information. The method used by LSI to capture the essential semantic information is dimension reduction, selecting the most important dimensions from a co-occurrence matrix (words by context) decomposed using singular value decomposition (SVD) . As a result, LSI offers a way of assessing semantic similarity between any two samples of text in an automatic unsupervised way. LSI relies on an SVD of a matrix (word _ context) derived from a corpus of natural text that pertains to knowledge in the particular domain of interest. According to the mathematical formulation of LSI, the term combinations that occur less frequently in the given document collection tend to be precluded from the LSI subspace. LSI does “noise reduction,” as less frequently co-occurring terms are less mutually related and, therefore, less sensible. Similarly, the most frequent terms are also eliminated from the analysis. The formalism behind SVD is rather complex and too lengthy to be presented here.
Page 257
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Once the documents are represented in the LSI subspace, the user can compute similarity measures between documents by the cosine between their corresponding vectors or by their length. These measures can be used for clustering similar documents together to identify “concepts” and “topics” in the corpus. This type of usage is typical for text analysis tasks.
This metric provide as an accurate calculation of cohesion of classes by using c3.Figure 4.2 reveals the measurement of cohesion using C3.
A.Advantages 1. We can predict the Cohesion. 2. We can predict the particular system is cohesive or not. B.Formula for C3 For every class C, the conceptual similarity between the methods mk and mj. C3=Mk*Mj /[Mk]2*[Mj]2 K=1, 2, 3……n; n: up to n comments j=2, 3…..n; n: up to n comments IV.COMPARISION Structural metrics are calculated from the source code such as references and data sharing between methods of a class belong together for cohesion. It define and measure relationships among the methods of a class based on the number of pairs of methods that share instance or class variables one way or another for cohesion. This metric bring only approximate value of cohesion .Figure 4.1 reveals the measurement of cohesion using Lcom5 .
Fig.4.1 PROCESS OF LCOM MEASURE
In proposed System unstructural information is retrieved from the source code like comments and identifiers. Information is retrieved from the source code using Latent Semantic Indexing.
Velammal College of Engineering and Technology, Madurai
FIG 4.2 PROCESS OF C3 MEASURE
. Here the cohesion is calculated for structured information using LCOM5 method and for unstructured information using Conceptual Cohesion of Classes .Then the results are compared and C3 method gives the accurate cohesion value. With the help of C3 and existing metrics we are achieving the high cohesion and low coupling V.CONCLUSION Classes in object-oriented systems, written in different Programming languages contain identifiers and comments which reflect concepts from the domain of the software system. This information can be used to measure the cohesion of software. To extract this information for Cohesion measurement, Latent Semantic Indexing can be used in a manner similar to measuring the coherence of natural language texts. This paper defines the conceptual cohesion of classes, which captures new and complementary dimensions of cohesion compared to a host of existing structural metrics. Principal component analysis of measurement results on three open source software systems statistically supports this fact. In addition, the combination of structural and conceptual cohesion metrics defines better models for the prediction of faults in classes than combinations of structural metrics alone. Highly cohesive classes need to have a design that ensures a strong coupling among its methods and a coherent internal description. ACKNOWLEDGMENT The authors are grateful to the management of Kamaraj College of Engineering and Technology, Virudhunagar, India for granting permission to undertake this work. Our thanks are due to the Head of the Department of Computer Science and Engineering of Kamaraj College of Engineering and Technology for allowing us the use of the laboratories and computing facilities.
Page 258
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VI.REFERENCES [1] Andrian Marcus, Denys Poshyvanyk, and Rudolf Ferenc,”Using Conceptual of Classes for Fault prediction in object oriented system”.IEEE TRANSACTIONS ON SOFTWARE ENGINEERING, VOL. 34, NO. 2, MARCH/APRIL 2008. [2] B. Henderson-Sellers, Software Metrics. Prentice Hall, 1996. [3] E.B. Allen, T.M. Khoshgoftaar, and Y. Chen, “Measuring Coupling and Cohesion of Software Modules: An Information-Theory Approach,” Proc. Seventh IEEE Int’l Software Metrics Symp.,pp. 124-134, Apr. 2001. [4] A. Marcus, A. De Lucia, J. Huffman Hayes, and D. Poshyvanyk, “Working Session: Information-Retrieval-Based Approaches in Software Evolution,” Proc. 22nd IEEE Int’l Conf. Software Maintenance, pp. 197-199, Sept. 2006. [5]P.W. Foltz, W. Kintsch, and T.K. Landauer, “The Measurement of Textual Coherence with Latent Semantic Analysis,” Discourse Processes, vol. 25, no. 2, pp. 285-307, 1998.
Velammal College of Engineering and Technology, Madurai
Page 259
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
On the Investigations of Design,Implementation, Performance and Evaluation issues of a Novel BDSIIT Stateless IPv4/IPv6 Translator 1
Hanumanthappa.J.1,Manjaiah.D.H.2,Aravinda.C.V.3 Teacher Fellow, Dos in CS, University of Mysore, Manasagangothri, Mysore, INDIA. 2 Reader &Chairman, Mangalore University, Mangalagangothri, Mangalore, INDIA. 3 M.Tech. KSOU. Manasagangotri, Mysore, INDIA. [email protected] [email protected] [email protected]
Abstract. Today, the Internet consists of native IPv4 (IPv4-only), native IPv6 and both IPv4/IPv6 dual networks. Currently both IPv4 and IPv6 existing protocols are considered as incompatible protocols. Unfortunately IPv4 and IPv6 are incompatible protocols when both the versions are available and the users of internet want to connect without any restrictions, a transition mechanism is required. Since a huge amount of resources have been invested on current IPv4 based Internet, how to smoothly transit the Internet from IPv4 to IPv6 is also a great tremendous and interesting research topic. During the time of migration from IPv4 to IPv6 networks, a number of transition mechanisms have been proposed by IETF to ensure smooth, stepwise and independent changeover. The development of Internet Protocol Version 6(IPv6), in addition to being a fundamental step to support growth of Internet is at the base of the increase IP functionality and Performance. It will enable the deployment of new applications over the Internet, opening a broad scope of technological development.BD-SIIT is one of the transitional technique which is mentioned in the IETF draft to perform the transition from IPv4/IPv6.We implement the BD-SIIT transition algorithm. The Stateless Internet Protocol/Internet Control Messaging Protocol Translation (SIIT) [RFC2765] is an IPv6 transition mechanism that allows IPv6-only hosts to talk to IPv4-only hosts. The mechanism involves a stateless mapping or bidirectional translation algorithm between IPv4 and IPv6 packet headers as well as between Internet Control Messaging Protocol version 4(ICMPv4) and ICMPv6 messages. SIIT is a stateless IP/ICMP translation, which means that the translator is able to process each conversion individually without any reference to previously translated packets. Most IP header field translations are relatively simple however; there is one issue, namely, how to translate the IP addresses between IPv4 and IPv6 packets. The NS-2 simulator is used to implement the BD-SIITmechanism
Velammal College of Engineering and Technology, Madurai
I.
INTRODUCTION.
1 .Importance of IPv6 based 4G Networks and On-going work. The IPv4/IPv6 transition process always occurs in deploying IPv6 based services crosses the IPv4 Internet. The IETF Next Generation Transition Working Group (NGtrans) has proposed many transition mechanisms to enable the seamless integration of IPv6 facilities into current networks. It also addresses the performances of various tunneling transition mechanisms used in different networks. The effect of these mechanisms on the performance of end-to-end applications is explored using metrics such as transmission latency, throughput, CPU utilization, and packet loss. This synopsis also looks into usage of Internet protocol version 6(IPv6) in one of the network architecture defined by Moby Dick research project. The Moby Dick has more straight forward approach. It takes Physical and Data link layer transmission functionalities of different access network as given exchange all existing all higher-level tasks by fully IPbased mechanisms. The Moby Dick was a three year European Union (EU) information Society Technologies (IST) project finished in December 2003.Project name Moby Dick has derived from the word “Mobility and Differentiated Services in a Future IP Network”. The main aim of the Moby Dick was to study and test IPv6 based and QoS enabled mobility architecture comprising Universal Mobile Telecommunication System (UMTS), Wireless Local Area Networks (WLAN), and wired Local Area networks (LAN) based on Ethernet. The Moby Dick project presents field evaluation results of an IP-based architecture for a 4G “True-IP” network. Moby Dick demonstrated the seamless integration of three disciplines like QoS, AAA and IP Mobility over a heterogeneous network infrastructure focusing on three access technologies (WLAN, Ethernet, and TD-CDMA).The Moby dick architecture incorporates Mobile IPv6, Fast handovers, AAA- control (Authentication, Authorization and Accounting).The migration from circuit-switched to IP
Page 260
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
based technologies and the growing role of mobility pave the way to a next generation Integrated network. The Importance of IP-based communication has already been identified in UMTS as well as in EDGE/IMT-2000 which proves an IP packet service using tunneling mechanisms but still employing access mechanisms of second generation networks.4G networks scenarios IP is used to glue all different link-level technologies, deploying technology un-aware protocols for mobility or quality of service (QoS).4G architectures should be able, thus, to embrace almost any wireless access technology available. This paper presents the field results obtained in the MobyDick architecture, an IP-based 4th Generation(4G)architecture for heterogeneous environments covering UMTS–like TD-CDMA wireless access technology, wireless and Ethernet LAN’s.This is one of the first implemented leader approaches to a 4G network thus a key contribution to this research work. The MIND (Mobile IP based network developments) project which was the follow-up of the BRAIN (Broadband radio access over IP networks) project was focused in mobility aspects, as well as ad-hoc, self-organizing, and meshed networks. The LONG project [ ] on IPv6 transition and deployment issues and Moby dick profited from some of the outcomes of that project. Nevertheless LONG didn’t aim at deploying a native IPv6 4G system, as Moby Dick did. 2. The IPv6 Transition. An IPv6 transition mechanism is a method to connect the hosts/networks using the same or different IP protocols under some specific IPv6 transition environment. The Internet protocol was started in early 1980.In the early s1990’s the Internet was growing some temporary solutions were offered in order to cover the fast growth in number of Internet user’s such as NAT, CIDR.At the Same time IETF began working to develop a new Internet Protocol namely IPv6 which was designed to be a Successor to the IPv4 Protocol. The main reason for designing this new Internet protocol(IPv6) was the need to increase the number of addresses(address spaces).The IPv6 address was designed with a 128-bit address scheme instead of 32-bit scheme in IPv4.So,the number of possible addresses in IPv6 is 3.4X1038 unique addresses.IPv6 will have enough to uniquely address every device (example Tele phone, Cell phone,mp3 player,hosts,routers,bridges etc) on the surface of earth with full end-to-end connectivity(about 32 addresses per square inch of dry land).In addition IPv6 is designed to support IPSec,Scalability,Multimedia transmissions,Security,Routing, Mobility, Real time applications like audio,video,Cryptography techniques like encryption and Decryption, Large address space, Better support for QoS,Stateless and Stateful address configuration, Enhanced support for Mobile IP and Mobile computing devices, Increased number of multicast addresses and improved support for multicast, New
Velammal College of Engineering and Technology, Madurai
protocol for neighboring node interaction, and New header format that is designed to keep header overhead to a minimum etc.Over all IPv6 was carefully thought out designed and was designed with future applications in mind. But the question now is whether it is possible for these two IPv4, IPv6 to work together smoothly and in an easy way? The answer is yes because the IETF IPng has proposed several mechanisms to be suitable for both Internet protocols to coexist. Some of the transition mechanisms work by encapsulating IPv6 packets in IPv4 packets then transporting them via an IPv4 network infrastructure and others work by operating dual IPv4/IPv6 stacks on the hosts or edge routers to allow the two versions of IPv4 and IPv6 to work together. Researchers designing new transition mechanisms in order to develop a proper mechanism, to be used in the transition from IPv4 from IPv6 and vice versa which overcomes some of the shortcomings that may appear in some proposed mechanisms. Some of the transition mechanisms work by encapsulating IPv6 packets in IPv4 packets then transporting them via an IPv4 network infrastructure and others work by operating dual IPv4/IPv6 stacks on the hosts or edge routers to allow the two versions of IPv4 and IPv6 to work together. Researchers designing new transition mechanisms in order to develop a proper mechanism, to be used in the transition from IPv4 from IPv6 and vice versa which overcomes some of the shortcomings that may appear in some proposed mechanisms. III. Related Work. 3.1. How the BD-SIIT (Bidirectional Stateless Internet Protocol/Internet control messaging Protocol Translation (SIIT) works. In this paper we proposed a new transition algorithm called BD-SIIT.This is also a new type of transition mechanism. As we Know that SIIT(Stateless Internet Protocol/Internet control messaging Protocol Translation(SIIT) is an IPv6 transition mechanism that allows IPv6 only hosts to talk to IPv4 only hosts. This mechanism contains a Stateless mapping or a bidirectional mapping (bidirectional translation algorithm) between IPv4,IPv6 packet headers as well as ICMPV6 and ICMPv4.Our new proposed BD-SIIT(new transition system) depends on the understanding of the received datagram, capturing the header, Identifying the header, Verification of the header, Transformation of the datagram to the destination environment, and then transmitting the datagram to the destination address.In,fact the proposed system deals with a Bidirectional operation that leads to convert the received datagram to the destination environment, depending on the address mapping ID value generated by DNS46 in order to be used in our proposed transition system between IPv4 and IPv6 protocols [29] [30] [31].BD-SIIT
Page 261
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
requires the assignment of an IPv4 address to the IPv6only host, and this IPv4 address is used by the host in forming a special IPv6 address that includes this IPv4 address. The mechanism which is intended to preserve IPv4 address rather than permanently assigning IPv4 addresses to the IPv6 only-hosts. The method of
assignment is one of the scope of BD-SIIT and also RFC 2765 also suggests that DHCP is basis for the Temporary IPv4 address assignment
Fig.1: BD-SIIT Translation process
The BD-SIIT has been proposed and designed in order to support the Bi-Directional communication sessions to be initiated by an IPv6 node which is located in the native IPv6 network and vice versa. The BD-SIIT Translation algorithm consists of the following two steps. 1. The First component explains V4-V6 Domain name System (DNS46) server that identifies the two public IPv4 and IPv6 addresses statistically or dynamically for each IPv4/IPv6 communicating system.
Stage-1: IPv6 Packet Transmission. Stage-2:IPv6-to-IPv4 mapping calculation and Address mapping. Stage-3:IPv6-in-IPv4 Header Translation Stage-4:IPv4 Packet Transmission. 2. The Second component explain the V4-V6 enabled gateway does the address mapping between an IPv4 and an IPv6 addresses as Well as header conversion between IPv4 and IPv6 packet headers
3.2. BD-SIIT Data Transmission from Source to Destination (end to end) Process:-
Fig.3: Data Packet Transmission from Source to Destination (End-toEnd). Fig.2: Translation of IPv6 header to IPv4 header.
The Data Packet transmission process for BD-SIIT is clearly defined in Fig.3.The BD-SIIT Translation mainly occurs due to the following four stages 1, 2, 3, and 4.
Velammal College of Engineering and Technology, Madurai
3.2. An Architecture of BD-SIIT. As mentioned in Fig.1.the BD-SIIT architecture consists of the DNS46(IPv4-IPv6 DNS),the IPv4-IPv6 enabled router, the IPv4 hostnames which are located in the IPv4 zone only, and the IPv6 host names which are located in the IPv6 zone only.
Page 262
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The BD-SIIT translation process can perform by using two techniques like (1) directly with in the end system. (2)Within a
Network-based device.
Fig.4: Network based BD-SIIT Translation process.
As shown in Fig.4.the BD-SIIT the upper layer protocols like TCP and UDP in transport layer can be passed through the translator relatively unscathed. For ex:-The BD-SIIT translation has been designed such that that TCP and UDP pseudo header checks sums is not affected by the translation process. While using BD-SIIT translation can cause some problems with various applications like FTP and embedded IP address in higher layer protocols and it requires the necessity of addition application specific application layer gateways in the translation process [29]. 3.3. Working Procedure of BD-SIIT Transition. As we know that BD-SIIT is a stateless IP/ICMP translation. The BD-SIIT translator is able to process each conversion individually, without any reference to previously translated packets. Although most of the IP header field translations are relatively very simple to handle, however one of the issue 0
79 80-Zero bits
Sl.No 1 2 3
80
related with BD-SIIT translator is how to map the IP addresses between IPv4 and IPv6 packets. A. The role of IPv6 mapped address in BD-SIIT Translator. BD-SIIT resides on an IPv6 host and converts an outgoing IPv6 headers into IPv4 headers, and incoming IPv4 headers into IPv6.To perform this operation the IPv6 host must be assigned an IPv4 address either configured on the host, or obtained via a network service left unspecified in RFC 2765.When the IPv6 hosts wants to communicate with an IPv4 host, based on DNS resolution to an IPv4 address, the BD-SIIT algorithm recognizes IPv6 address as an IPv4 mapped-address as shown in Fig.3.The one of the mechanism to translate the resolved IPv4 address into an IPv4 mapped address is provided by Bump-in-the stack(BIS)or Bump-inthe API(BIA) techniques[30].
95 FFFF(16 bits)
96
127 32 bits(IPv4 Address)
Fig.5:IPv4-mapped-IPv6 address. Table.1: Address mapping IPv6/IPv4. IPv6 Address IPv4 Address Address mapping value ABC2::4321 195.18.231.17 1 ABC2::4321 195.18.231.17 2 ABC2::4321 223.15.1.3 37
1 2`
Table.2:DNS46 Corresponding to IPv4 and IPv6. Address IPv4 Address IPv6 Address DNS mapping value 212.17.1.5 ---B 4 223.15.1.3. 1C::DACF Y 37
Sl.No
Table.3: Address mapping IPv4/IPv6. IPv4 Address IPv6 Address Address mapping value
Sl.No
1
195.18.231.17
2
210.154.76.91
3
223.15.1.3.
ABC2::4321 ABC2::4321 ABC2::4321
3.4. The BD-SIIT A Novel Transition Address-mapping algorithm for the forward operation.
Velammal College of Engineering and Technology, Madurai
1 2 37
Based on the presence of IPv4-mapped-IPv6 address as the destination IP address the BD-SIIT algorithm performs the header translation as described in Algorithm-1 and
Page 263
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Algorithm-2 to obtain an IPv4 packet for transmission via data link and Physical layers. The Fig.4.shows the Protocol .
stack
view
of
BD-SIIT
Fig.5: Protocol Stack view of BD-SIIT.
The following steps show the forward operation during the migration from IPv4/IPv6. Algorithm-1:-IPv4->IPv6: Forward operation. 1. When Host X belongs to X-zone initiates a request (Query message) to DNS46 in order to get the IP-address of Host Y which belongs to IPv6 zone. 2. When DNS46 receives a query message as a request then its checks its (Table-2) to identify whether Host Y has an IPv6 address which is unknown for the Host X.The DNS46 knows that the whole IPv6 zone has a public IPv4 address (like NAT method) i.e. 195.18.231.17 address in the destination address field of IPv4 header then forwards it via a network. 3. Simultaneously, the DNS46 sends another message to V4-V6 enabled router in order to update table-3. 4. When the Host X receives requested address of Host Y, immediately creates the IPv4 packet, inserting 195.18.231.17 address in the destination address field of IPv4 header then forwards it via network. 5. When the IPv4 packet is arrived to V4-V6 enabled router, then the router identifies a packet and verifies that a destination Address is a public address with mapping value 2 that indeed refers ABC2::4321 IPv6 address (as shown in table3).Then the V3-V4 enabled router which updates Table-1 then creates its new IPv6 packet which is based on IPv4 packet, and forwards it to Its destination in IPv6 Zone. 6. When Host Y accepts the packet then it starts to process it successfully without getting any problem. Algorithm-2:-IPv6->IPv4: Feedback operation. The following steps show the feedback operation from IPv6-IPv4 zone one which is illustrated in Figs.1. And 5.
Velammal College of Engineering and Technology, Madurai
[Note:-Consider Host X as a Client and Host Y as a Server. If the Client A sent a HTTP get command to retrieve a web page from the server Y.].[As shown in Fig.7.steps 1- 6. 1. As a response for the received command from a client X, server Y creates packet(s) then forwards them via a network to the client X using the public IPv6 zone address(ABC2::4321) as a destination address. 2. When V4-V6 enabled router receives a packet which has been sent by Server Y, then it verifies its Table-1,Table3,depending on the addressing mapping value like 37 in our scenario, it refers to 220.12.145.10 as a Sender address in Table-2 and 223.15.1.3 as a destination address in Table-2 instead of instead of 1C:: DACF and ABC2::4321 IPv6 rely. 3. After that, the V4-V6 enabled router creates a new IPv4 packet, based on the accepted IPv6 packet then forwards it to the destination (Client X). 4. When the Client X receives the IPv4 packet, its starts to process successfully without any problem. Algorithm-3:-BD-SIIT Algorithm of IPv4 header Conversion to IPv6 header conversion. 1. When BD-SIIT allows IPv6 hosts which do not contain permanently, assigned IPv4 address 2. When the IPv6 host tries to communicate with IPv4 host 3. Then BD-SIIT handles the IP address translation between IPv4 and IPv6. 4. If IPv6 address type=0: FFFF: v4 {print “IP address as an IPv4-mapped-IPv6 address”. It is mainly used to mapping IPv4 host addresses to IPv6 addresses.} 5. If the IPv6 address type=0: FFFF: 0:v4 {Print “IP address as an IPv4 translated (compatible) address.}
Page 264
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.6: BD-SIIT, IP Forward address translation, IPv6 to IPv4.
The Fig.6.shows the IP address translation process IPv6 to IPv4 with IPv4-mapped-IPv6 address and an IPv4 translated (compatible) address. The IPv6 host has obtained a temporary address-v4temp-for use in communicating with the IPv4 host. The Fig.6.very briefly illustrates the operation of IP address translation in going from the IPv6 host, using an IPv4 translated address to the IPv4 host. The translation of the remaining field is straight forward with a couple of exceptions. If there is no IPv6 fragment header, the IPv4 header fields are set as follows [31] [32]. Algorithm-4:-BD-SIIT.Algorithm for the Forward Header conversion IPv6->IPv4. If version=4 {proceed to set Internet Header Length=5,} Else if {proceed to set type of service and precedence= by default copied from the IPv6 header traffic class field},
Else if {proceed to set total length=pay load length value from IPv6 header+IPv4 header length.}, Else if {proceed to set Identification=0}, Else if {proceed to set Flags=the more fragments flag is set=0} Else if {proceed to set Don’t Fragment flag=1}, Else if {proceed to set Fragment offset=0}, Else if {proceed to set TTL=Hop limit value copied from IPv6 header - 1.}, Else if {proceed to set Protocol=next header field copied from IPv6 header}, Else if {proceed to set Header Checksum=Compute once IPv4 header has been created. Else if Source IP address=low order 32 bits of IPv6 Source address field (IPv4 translated address field) Else Destination IP address=Low order 32 bits of IPv6 Destination address field (IPv4 mapped address field)}
Fig.7: BD-SIIT, IP Reverse address translation, IPv6ÅIPv4.
Algorithm-5: BD-SIIT.Algorithm for the Reverse Header conversion IPv4->IPv6. If version=6{proceed to set version=6} Else if {proceed to set Traffic Class= IPv4 header ToS bits}, Else if {proceed to set Flow label=0}, Else if {proceed to set Pay Load Length=IPv4 header Total length value-(IPv4 header length+IPv4 options}, Else if
Velammal College of Engineering and Technology, Madurai
{proceed to set next header =IPv4 header protocol field value} Else if {proceed to set hop limit=IPv4 TTL field value-1}, Else if {proceed to set Fragment offset=0}, Else if {proceed to set source address =0:0:0:0: FFFF: /80.Concatenated with IPv4 header source IP address}, Destination IP address=0:0:0:0:0: FFFF: /96 concatenated with IPv4 header destination address.}[31][32]
Page 265
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.8: BD-SIIT hosts connection IPv6-IPv4 operation using TCP Protocol.
IV. Implementation & Performance Metrics. We implemented our proposed BD-SIIT translation system and tested it by varying the number of packets sent as well as the transmission rate for each communication system. We have tested comparison between EED(end-enddelay)of IPv4-IPV4 and IPv6-IPv6 communication session, A Comparison between the throughput(v4-to-v4 , and v6-to-v6) communication sessions, Round trip time(RTT) for each data packet of varying sizes, DNS response time, the total transmission time with TCP and UDP Protocols. In this paper we are mainly measured two important metrics like throughput and End–to End delay for BD-SIIT. Throughput: We calculated the throughput performance metric in order to identify the rate of received and processed data at the intermediate device (router or gateway)during the simulation period. The mean throughput for a sequence of packets of specific size is calculated by using equations 1 and 2. --------(1) Where Thrj=
------------ (2)
Where Thrj is the value of the throughput when the packet j is received at intermediate device like DSTM gateway, BDSIIT Gateway, v4/v6 Router and N is the number of received packets at intermediate device, Pr is the no of packets. At intermediate device and Pg is the number of packets created by source host. 2. End-to-End Delay (EED) As we know that to calculate the performance evaluation of various novel transition mechanisms EED plays an important role. The mean EED for a sequence of packets of specific size is calculated as follows. -------(3) Where as EEDk=Tdk-Tsk-----(4)
Where EEDk is the End-to-End delay of packet “k”, Tsk is the generated time of packet “k” at source workstation and Tdk is the received time of the packet “k” at the destination workstation, Nr is the total number of received packets at the destination workstation and Mean EED. 4.1. Simulation Results and Discussions. The following table (Table-1)shows the simulation parameters, which are used to calculate the performance measurements using ns-2 simulation environment when each packet arrival follows a Poisson process with rate λ=2
.Table-4: Simulation parameters. Sl.No 1 2 3 4 5
Simulation parameters. Buffer Size Propagation delay Payload Size Very traffic Loads Queue Management Scheme
Scenario-1–The simulation results for scenario one present in the EED and throughput result when a direct link connection between either IPv4 workstations in IPv4 only network or IPv6
Velammal College of Engineering and Technology, Madurai
Value 500 Packets. 10ms 200 Bytes 6~240 Nodes. Drop tail
workstations in IPv6 network is conducted through IPv4/IPv6 router. The results are illustrated in Fig.9.andFig.10.
Page 266
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.9: The comparison between EED of v4-to-v4 and v6-to-v6 communication sessions.
source and destination hosts compared with the IPv4 that allows these processes to be performed by all the Intermediate devices, which fall between the source and destination hosts. The Fig.9.also shows that the difference between the v4-to-v4 EED and v6-to-v6 EED results is very small and the differences increases when the number of connected workstations increases as well. The v6-to-v6 EED is suddenly increases because of network congestion and when the number of connected nodes becomes more i.e. greater than 200.On the other hand the following Fig.10. Shows a Comparison between the v4-to-v4 and v6to-v6 throughput results.
From Fig.9.the simulation results shows that the v4-tov4 EED is definitely less than v6-to-v6 EED.In fact there may be so many reasons which explains why this happens. The first reason is the difference between IPv6 header (40 Bytes) and the IPv4 header (20 Bytes) and it causes more traffic overhead especially when the IP packet payload is small. The Second reason is the size of IP packet payload is fixed, in all communication sessions, this means no fragmentation process is needed, which leads to reduce the benefit from the IPv6 feature that allows the fragmentation and defragmentation processes to be performed by only
Fig.10: A Comparison between the throughput (v4-to-v4, and v6-to-v6) communication sessions.
VI. References.
4.2. DNS Response Time:The DNS Response Time (DNSRT) metric shows that the time needed to calculate, the communication session between the two end systems which are located the two heterogeneous networks. The DNS Response time can be calculated by using the equation (1)
Where Transtime = Transmission time of a Packet is the link number between the two nodes, and Proctime = Processing time, j is the node number and M is the total number of nodes. V. Conclusions. In this paper we have proposed our novel BD-SIIT Transition mechanism. The Simulation results shows the impact of the translation process, that contains performing of address mapping as well as the header translation process which are needed for each incoming and outgoing packet to and from the BD-SIIT translator. The BD-SIIT novel transition concentrated on identifying, determining, 5. translating and forwarding packets between the two different network environments (IPv6 and IPv4 zones).Our proposed new BD-SIIT reduces the size of packet compared with the 6. encapsulation method in the tunneling algorithm.
Velammal College of Engineering and Technology, Madurai
1. Hanumanthappa.J.,Manjaiah.D.H.,”IPv6 and IPv4 Threat reviews with Automatic Tunneling and Configuration Tunneling Considerations Transitional Model: A Case Study for University of Mysore Network”, International Journal of Computer Science and Information(IJCSIS)Vol.3.,No.1,July-2009,ISSN 1947-5500,Paper ID: 12060915] of IPv4 Network 2. Hanumanthappa.J.,Manjaiah.D.H.,”Transition Applications to IPv6 Applications”[TIPv4 to TIPv6],Proceedings of IEEE International Conference on emerging trends in computing(ICETiC2009),Virudhunagar,Tamilnadu 8-10,January 2009,INDIA.[Paper ID 234].
3. Hanumanthappa.J. Manjaiah.D.H., Tippeswamy.K. “An Overview of Study on Smooth Porting Process Scenario during IPv6 Transition” [TIPv6] , Proceedings of IEEE International Conference on the IEEE International Advance Computing Conference IACC-2009 on March 5-8 at Patiala, Punjab [Paper ID IEEE-APPL-1278] 1) 4. Hanumanthappa.J.,Manjaiah.D.H.,“A Study on Comparison and Contrast between IPv6 and IPv4 Feature Sets” Proceedings of International Conference on Computer Networks and Security(ICCNS2008),Pune,Sept 27-28th,2008,[Paper code CP 15]. 5.Ra’Ed AlJa’afreh, John Mellor, Mumtaz Kamala,”A Novel IPv4/IPv6 transition mechanism which support transparent connections”. 6.Behrouz A.Forouzan,Third Edition,“TCP/IP Protocol Suite” .Atul Kahate,“Cryptography and Network.Security“,Tata McGraw-
Page 267
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hill,2003,pp-8-10.Kurose.J.& Ross.K.(2005)Computer Networking:A top-down approach featuring the Internet.3rd ed,(Addison Wesley).S.Hagen:“IPv6 essentials“,Orielly,July 2002,ISBN-0-59600125-8. 7.Moby dick project: “Mobility and Differentiated services in a future IP network:” Final project report, 4.4.2004. 7. J.Wiljakka(ed.,)“Analysis on IPv6 transition in 3GPP networks”,draft-ietf-v6ops-3gpp-analysis- 04.txt,Internet draft, work in progress 8. MIND:Mobile IP based network developments, IST Project, http://www.mind-project.org. [20].NOMAD: Integrated networks for Seamless and Transparent service Discovery, IST Project, http://www.ist-moby Dick.org. 9. Moby Dick: Mobility and Differentiated services in a future IP Network, IST Project, www.ist.mobydick.org. 10. John.J.Amoss and Daniel Minoli, Handbook of IPv4 to IPv6 Transition: methodologies for institutional and Corporate Networks. Auerbach Publications. 11. S.G.Glisic Advanced Wireless Communications, 4G Technology. John Wiley Chichester,2004. 12. Juha wiljakka ,Jonne Soninnen,Managing IPv4 –to –IPv6 Transition Process in Cellular Networks and Introducing new Peer-to-Peer Services. 13. Ioan R,Sherali.Z.2003.Evaluating IPv4 to IPv6 mechanism.IEEE,West Lafayette,USA,v(1):1091–1098
Transition
14. Gilligan. & Nodmar .E. (1996) Transition Mechanisms for IPv6 Hosts and Routers. OMNeT++Discrete Event Simulation.
15. L. Toutain
16. 17.
18. 19. 20.
, H.Afifi, Dynamic Tunneling: A new method for IPv4 to IPv6 transition. Internet draft, <draft-ietf-ngtranns-dti-00.txt> E.Nordmark, Stateless IP/ICMP translation algorithm (SIIT), RFC 2765, February 2000. Ra’ed AlJa’afreh, John mellor,Mumtaz Kamala,”A Novel IPv4/IPv6 transition mechanism which support transparent connections”. TimRooney,IPv4/IPv6Transition strategies ,Director, Product management, BT Diamond IP.John.J.Amos and Daniel Minoli,Handbook of IPv4 to IPv6 Transition Methodologies for Institutional and Corporate Networks Jivika Govil,Jivesh Govil, Navkeerat Kaur, Harkeerat Kaur, An examination of IPv4 and IPv6 Networks: constraints, and various transition mechanisms. K.K.Ettikan, et al.”Application Performance Analysis in Transition mechanism from IPv4 to IPv6,”Multimedia University(MMU),Jalan Multimedia, June 2001. porject.SHISA,2006,http://www.mobileip.jp/
21. J.Bound,
Mysore-06 and currently pursuing Ph.D in Computer Science and Engineering, from Mangalore University under the supervision of Dr.Manjaiah.D.H on entitled “Design and Implementation of IPv6 Transition Technologies for University of Mysore Network (6TTUoM)”. His teaching and Research interests include Computer Networks,Wireless and Sensor Networks, Mobile Ad-Hoc Networks, Intrusion detection System, Network Security and Cryptography,Internet Protocols,Mobile and Client Server Computing,Traffic management,Quality of Service, RFID,Bluetooth,Unix internals, Linux internal, Kernel Programming,Object Oriented Analysis and Design etc.His most recent research focus is in the areas of Internet Protocols and their applications.He received his Bachelor of Engineering Degree in Computer Science and Engineering from University B.D.T College of Engineering,Davanagere,Karnataka(S),India(C),Kuve mpu University,Shimoga in the year 1998 and Master of Technology in CS&Engineering from NITK Surathkal,Karnataka(S ),India (C) in the year 2003.He has been associated as a faculty of the Department of Studies in Computer Science since 2004.He has worked as lecturer at SIR.M.V.I.T,Y.D.I.T,S.V.I.T,of Bangalore. He has guided about 250 Project thesis for BE,B.Tech,M.Tech,MCA,MSc/MS.He has Published about 15 technical articles in International ,and National Peer reviewed conferences. He is a Life member of CSI, ISTE,AMIE,IAENG,Embedded networking group of TIFAC–CORE in Network Engineering,ACM,Computer Science Teachers Association(CSTA),ISOC,IANA,IETF,IAB,IRTG,etc. He is also a BOE Member of all the Universities of Karnataka,INDIA.He has also visited Republic of China as a Visiting Faculty of HUANG HUAI University of ZHUMADIAN,Central China, to teach Computer Science Subjects like OS and System Software and Software Engineering,Object Oriented Programming With C++,Multimedia Computing for B.Tech Students. in the year 2008.He has also visited Thailand and Hong Kong as a Tourist.
Dual Stack Transition mechanism,draft-bound-dstm-exp-
1.txt,april,2004,http://www.dstm.info.
Mr.Hanumanthappa. J. is Lecturer at the DoS in CS,University of Mysore, Manasagangothri,
Velammal College of Engineering and Technology, Madurai
Dr.Manjaiah.D.H D.H. is currently Reader and Chairman of BoS in both UG/PG in the Computer Science at Dept.of Computer Science,Mangalore University, and Mangalore.He is also the BoE Member of all Universities of Karnataka and other reputed universities in India.He received Ph.D degree from University of Mangalore,
Page 268
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
M.Tech. from NITK,Surathkal and B.E.,from Mysore University.Dr.Manjaiah.D.H D.H has an extensive academic,Industry and Research experience.He has worked at many technical bodies like IAENG,WASET,ISOC,CSI,ISTE,and ACS. He has authored more than -25 research papers in international conferences and reputed journals. He is the recipient of the several talks for his area of interest in many public occasions. He is an expert committee member of an AICTE and various technical bodies. He had written Kannada text book,with an entitled, ”COMPUTER PARICHAYA” ,for the benefits of all teaching and Students Community of Karnataka.Dr.Manjaiah D.H’s areas interest are Computer Networking & Sensor Networks, Mobile Communication, Operations Research, E-commerce, Internet Technology and Web Programming
.
Aravinda.C.V.,currently pursuing M.Tech(I.T) K.S.O.U., Manasagangotri, Mysore-06.He received M.Sc ., M.Phil in Computer Science.He has worked as a Lecturer in the following institutions. 1.CIST, Manasagangotri,Mysore, 2.Vidya Vikas Institute of Engg and Technology,Mysore. 3.Govt First Grade college,Srirangapatna and Kollegal. He has published two papers in National Conference hosted by NITK,Surathkal,Mangalore.
Velammal College of Engineering and Technology, Madurai
Page 269
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The Role of IPv6 over Fiber (FIPv6): Issues, Challenges and its Impact on Hardware and Software 1
Hanumanthappa.J.1, Manjaiah.D.H.2, Aravinda.C.V.3 Teacher Fellow, Dos in CS, University of Mysore, Manasagangothri, Mysore, INDIA. 2 Reader & Chairman, Mangalore University, Mangalagangotri, Mangalore, INDIA. 3 M.Tech. KSOU. Manasagangotri, Mysore, INDIA. [email protected] [email protected] [email protected]
Abstract The Internet protocol was started in early 1980.In the earlys1990’s the Internet was growing some temporary solutions were offered in order to cover the fast growth in number of Internet user’s such as Network address translation (NAT), CIDR (Classless Inter domain Routing).At the Same time IETF began working to develop a new Internet Protocol namely IPv6 which was designed to be a Successor to the IPv4 Protocol. This paper proposes the new concept of error handling at network layer (layer-3) instead of data link layer (layer-2) in ISO/OSI reference model by adopting new capabilities and by using IPv6 over Fiber. This paper also shows how to reduce the over head In terms of header processing at data link layer by eliminating cyclic redundancy check (CRC) field by using IPv6 over Fiber. Therefore we can also prove that ISO/OSI model contains 6 layers instead of 7 layers. Key words: IPv4, IPv6, CRC, Ethernet, FCS.
I.
Introduction to IPv6
In the last 20 years, the internet undertook a huge and unexpected explosion of growth [].There was an effort to develop a protocol that can solve problems in the current Internet protocol which is in the current internet protocol which is in Internet protocol version 4(IPv4).It was soon realized that the current internet protocol the IPv4, would be inadequate to handle the internet’s continued growth. The internet Engineering task force (IETF) was started to develop a new protocol in 1990’s and it was launched IPng in 1993 which is stand for Internet protocol Next Generation. So a new generation of the Internet Protocol (IPv6) was developed [7], allowing for millions of more IP addresses. The person in charge of IPng area of the IETF recommended the idea of IPv6 in 1994 at Toronto IETF[1].But mainly due to the scarcity of unallocated IPv4 address the IPv4 protocol cannot satisfy all the requirements of the always expanding Internet because
Velammal College of Engineering and Technology, Madurai
however its 32 bit address space being rapidly exhausted[2] alternative solutions are again needed[3].It is reported that the unallocated IPv4 allocated IPv4 addresses will be used with 6 to 7 years short period[2].The Long term solution is a transition to IPv6[5]which is designed to be an evolutionary step from IPv4 where the most transport and application layer protocol need little or no modification to the work. The deployment of NAT [3] can alleviate this problem to some extent but it breaks end to end characteristic of the Internet, and it cannot resolve the problems like depletion (exhaustion) of IPv4 addresses. 1.1. Features of IPv6. The main reason for designing this new Internet protocol(IPv6) was the need to increase the number of addresses(address spaces).The IPv6 address was designed with a 128-bit address scheme instead of 32-bit scheme in IPv4.So,the number of possible addresses in IPv6 is 3.4X1038 unique addresses.IPv6 will have enough to uniquely address every device (example Telephone, Cell phone,mp3 player,hosts,routers,bridges etc) on the surface of earth with full end-to-end connectivity(about 32 addresses per square inch of dry land).In addition IPv6 is designed to support IPSec, Scalability, Multimedia transmissions, Security, Routing, Mobility, Real time applications like audio,Video,Cryptography techniques like encryption and Decryption, Large address space, Better support for QoS. 1.2. Data Link error handling function. In this paper performance can be measured in many ways, including transit time and response time. Transit time is the amount of time for a message to travel from one device to another
Page 270
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
. Fig.1:An Exchange using the OSI model.
Transit Time (TT) =Time taken from Source to Destination. [17].Response Time (RT) =Elapsed time between an inquiry and response [17].Performance is also evaluated by two networking metrics like throughput and delay
Fig.2: The Interaction between layers in the ISO/OSI model.
1.3. Layered architecture. The OSI model is composed of seven ordered layers physical(layer1),Data link(layer2),Network(layer3), Transport(layer4), Session(layer5), Presentation(layer6), Application layer(7), To reduce the design complexity, computer networks follow a layered architecture[1].Each layer clearly defines based on the previous layers and has a set of well defined functions with clear cut boundaries. Also with layered architecture the implementation details of each layer is independent of other layers.Fig.2.shows the layers involved when a message is sent from device A to device B.As a message travels from A to B it may pass through many intermediate nodes. These intermediate nodes usually involved only the first three layers of OSI model. Each layer defines a family of functions distinct
Velammal College of Engineering and Technology, Madurai
from those of the other layers. By defining and localizing functionality in this fashion, the designers created an architecture that is both comprehensive and flexible. Most importantly the ISO/OSI model allows complete interoperability between otherwise incompatible systems. Within a single machine, each layer calls upon the services of the layer just below it. The processes on each machine that communicates at a given layer are called peer-to-peer processes. The passing of the data and network information down through the layers of the sending device and back up through the layers of the receiving device is made possible by an interface between each pair of adjacent layers. Layers 1, 2 and 3–physical, data link and network are the network support layers. Layers 5, 6, and 7–session, presentation and application can be thought of as the user support layers. The upper OSI layers are almost always implemented in software; lower layers are a combination of hardware and
Page 271
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
software except for the physical layer, which is mostly hardware. In Fig.1. Which gives an overall view of the OSI layers, D7 means the data unit at layer 7.The process starts at layer 7, then moves from layer to layer in descending, sequential order.At each layer commonly a header or trailer can be added to the data unit. At each layer the packet is encapsulated with a header that contains control information to handle the data received at other side by the corresponding layer. This paper also states that how take care of error handling function very efficiently by reducing the overall packet processing time, and thus improve the transmission of IPv6 packets. This can be achieved by
utilizing the characteristics or capabilities of the communication medium used to transfer data, and by improving the existing error handling mechanisms at the lower layer. In Data communication and networking and computer networks the errors are broadly divided into single bit error and Burst errors. The term single-bit error means that only 1 bit of a given data unit (such as a byte, character or packet) from 0 to 1.The Fig.4.shows Burst error with Length 5.The most important common approaches to detect the errors are parity check(pc),Cyclic redundancy check(CRC) and Checksum.[17][19]
Fig.3: Single-bit error.
Fig.4: Burst error of Length 5.
Fig.5: Error Detection methods.
The errors that are detected by either redundancy, parity check, or cyclic redundancy check, or checksum can be corrected by with two types of mechanisms called Automatic repeat request(ARQ),and Forward error correction(FEC).This paper is organized as follows: We briefly described Introduction to IPv6 in 4G networks and its on-going work in Section 1.We described,IPv6 transition in terms of 4G in section-2.Section-3 clearly specifies related work meant for the translation of IPv4/IPv6 BD-SIIT a novel transition algorithms. Finally we concluded the whole paper in section 4.
Velammal College of Engineering and Technology, Madurai
III. The ISO/OSI Packet Transmission. The ISO/OSI and TCP/IP are the two most important popular network architectures that have been widely used. The ISO/OSI reference model has remained as a popular model for its simplicity, and clarity of its functions where the TCP/IP was a more working model that is popularly used over the Internet. The Fig.6.shows the Transport layer communication process. The data from the user on the sender side, passes through a series of layers before it is transmitted over the internet to reach the other machine(recipient side) .On the receiver side data received
Page 272
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
by the Physical layer goes up to the application layer, by neglecting the header in each layer. It can be observed from
the Fig.6.the original length of the data remains same, but the length of the header increases with each layer.
Fig.6: The Transport layer communication process.
3.1. IPv6 Packet format
. Fig.7: The IPv6 Packet format.
The IPv6 packet consists of IPv6 base header, extension headers, and upper layer protocol data unit.IPv6 base header is of fixed size, and is of 40 bytes in length. Payload length may change due to the presence of extension headers. The IPv6 packet format can also support for
multiple extension headers and the use of extension header is optional. The IPv6 based header consists of 8 fields as shown in the Fig.7.Extension headers are inserted into the packet only if options are needed. They are processed in the order in which they are present.
Fig.8: The Ethernet Frame Format.
3.2. Error handling aspects and its issues by using CRC in Data link layer.
1011000 is a code word and when cyclically left shifted then 0110001 is also a code word.
CRC is one of the most popular error detection mechanisms that are currently being used at data link layer. Cyclic codes are special linear block codes with one extra property. In a cyclic code if a code word is cyclically shifted then the result is another code word. For ex – if a
3.2.1. Cyclic redundancy check.
Velammal College of Engineering and Technology, Madurai
CRC is a cyclic code method to correct errors in networks like LAN and WANs.CRC is a popular method used simple to implement in binary hardware, and easy to analyze mathematically and are particularly good at
Page 273
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
detecting common errors caused by noise in transmission channels. In our proposed method by removing CRC from the Ethernet frame format and fix it in the IPv6 packet
format of the extension header at network layer of ISO/OSI model. The Fig.9.shows the novel IPv6 packet format [17].
IV. Network layer error checking Methodology (FIPv6EH)
Fig.9:IPv6 novel packet format with CEH (CRC Extension header).
3.3. Applications of CRC. 1.Cyclic codes performing very good performance in detecting single-bit errors, double errors, an odd number of errors, and burst errors. 2. CRC is easy to implement in hardware and software. 3. CRC is very fast to correct errors when implemented in hardware. 4. Cyclic code has made a good candidate for many networks [17].
In this, paper we considered to place error detection in the network layer to check IPv6 whole packet including header and payload. The error detection (check) method used will be same as CRC method that is currently being used with the existing systems. The CRC extension header (CRCEC) is a new IPv6 extension header to handle error detection for the entire IPv6 packet shown in Fig.9. Simulation Scenario-1: The first simulation states that when sender generates an IPv6 packet, and the corresponding FIPv6EC (CRCEC) code to be inserted to the IPv6 packet format as CRCEC.The packet with CRCEH is sent through a network with a topology as mentioned in Fig.10.The routers are mainly used to connect sender host and a recipient host will not verify the FIPv6H (CRCEH) instead it will identify the next route of the packet. When the receiver upon receiving the packet will verify the CRCEH (FIPv6EH) in its network layer to check whether the packet is error free, and then deliver to the upper layers. If suppose the received packet contains an error it will be discarded and wait for retransmission [21].
Fig.10: Novel FIPv6EH Error detection model in Network Layer.
Simulation-2: Second simulation also covers error control in data link layer, the sender generates the IPv6 packet without, any extension header inside. It is embedded in data link layer with header and trailer. The FCS in the trailer is actually CRC-32 code generated from the whole frame. The Fig.10.shows when the intermediate node-1 which is also called router-1 receives a packet, and verifies the CRC code inside, when the verification process generated no code in the packet then only the packet is processed to next router-2 of a network layer. Each time the bad packets are rejected and then the receiving station will be waiting for the retransmission of a packet. This process will continue in all the intermediate routers, which
Velammal College of Engineering and Technology, Madurai
connects the sender with receiver. We have analyzed from the simulation scenario-1 and scenario-2 there is no error control in data link layer and it takes place only at the end workstations, and will not be done at every routers. V. Performance Evaluation Metrics and Simulation Parameters. In this paper we have calculated two performance evaluation metrics like throughput, end-to-end delay. The throughput can be calculated as follows. The throughput is a measure of how fast we can actually send data through a network. Although at first glance throughput and
Page 274
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
bandwidth in bits per second seem the same, but totally there are different. The bandwidth is a total measurement of a link; the throughput is an actual measurement of how fast we can send data. We measured the throughput performance metric value in order to find out the rate of received and processed data at the router (intermediate device) during the time of simulation. The mean throughput for a sequence of packets of specific size can be calculated by using the formula
-----(1) whereas Thri=Paccept/Pcreated*100%----(2) Where Thri is the throughput value when the packet “i” is accepted at the intermediate device like router. and “n” is the total number of packets received packets at the router, and Prec is the number of received packets at router and Pcrea The below Table-1 shows the Simulation results that are mainly used to calculate, the Performance measurements .
is the number of packets created by the source hosts, and the mean throughput is the mean value for each communication. Latency (Delay): The latency or delay defines how long the entire message takes to completely arrive as at the destination from the time the first bit is sent out from the source. Therefore we can conclude latency is made up of four important components: Propagation time, Transmission time, Queuing time and Processing delay. Delay=Propagation time (Pt) +Transmission time (Tt) +Queuing time (Qt) + Processing delay (Pd). Where Pt=Distance/Propagation speed. Transmission Time=Message Size/Bandwidth. Queuing time=Time needed for each Intermediate or n devices to hold the message before it can be processed
using the NCTUNS5.0 Simulator and emulator.The Simulation parameters as shown in the Table-1
Table-1: Simulation parameters. Simulation Value Parameters 1.Buffer Size 200 packets 2.Delay 5ms 3.Pay load size 100 Bytes 4.Vary traffic load 6~150 nodes. 5.Queue mngt Drop tail Scheme V. Challenges, necessity and Impact of IPv6 over Fiber on Hardware and Software. 5.1. Challenges of IPv6 over Fiber. 1. To minimize the header overheads by handling efficiently similar or redundant functionalities among layers. 2. To maximize the performance of data transmission in terms of packet data rate, throughput, and bandwidth utilization. 3. To maximize the efficiency of operations for data transmission. 4.To propose a new new frame work by simplifying the existing network models like ISO/OSI and TCP/IP model through possible changes in data link layer and network layer.
Velammal College of Engineering and Technology, Madurai
5. By collapsing data link layer and network layer for similar functions and try to eliminate redundant functions. 6. It is also possible to enhance performance of IPv6 packet transmission in terms of frame rate, throughput and bandwidth. 7. To Simplify the ISO/OSI layered architecture by reducing the size of the data link layer header or removing it completely [21]. 5.2. Necessity of IPv6 over Fiber. 1. Ellimination of MAC addresses. 2. Discard Data Link layer. 3. No framing process. 4. Increase speed transfer. 5. Reduce Worm and Virus outbreak. 6. Efficiency can be increased and Buffering can be reduced.
Page 275
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
7. ARP currently being used is not required. 8. Data is encapsulated into many headers. 9.26 bytes of Data Link Layer header for each frame.Max data size in each frame is 1500 bytes. Some amount of frame size is allocated for the headers and it is waste of bandwidth [21].
3. Layer 3 switches will experience faster packet processing 4. Routing Table will consist of IPv6 Address and smaller routing table entry. 5. No MAC address to IP mapping is required resulting in router performing faster and more efficient [21].
5.3. Impact of IPv6 over Fiber on Hardware.
5.4. Impact of IPv6 over Fiber on Dual Stack.
The impact of IPv6 over Fiber can updates some of the changes in hardware. The following are some of the changes in IPv6 over Fiber. 1. Only drivers can be re-written for NIC.No changes required for manufacturing NIC. 2. Layer 2 switches will be obsolete or will have to run dual stack.
1. Once a new architecture is in its placeman dual stack is required to process the packets transmitted from both current architecture as well as new architecture. 2. The initiated protocol will automatically choose the appropriate stack. 3. This will be an international standard to be used by all Internet applications especially for real time applications [21].
Fig.11:IPv6 over Fiber. VI. Conclusions. In this paper we have proposed our new concept of error handling mechanism like FIPv6EH at network layer instead of at Data link layer with the help of various salient features of IPv6 protocol and the various characteristics features of high speed communication networks like Fiber. This paper not only enhances the error handling issues at network layer along with that its exploits very briefly Challenges and Impact of IPv6 over Fiber on Hardware as well as Software. The proposed FIPv6EH reduces the overhead by eliminating CRC field from its frame header and placing it in the extension header in the Network layer instead of doing header processing at data link layer. The proposed concept will enhance the performance of packet transmission. As we know that Data link layer and Network layer in ISO/OSI model are the two important lower layers where the IPv6 protocol is in between neighboring nodes. This paper also states that error checking at the packet will be more efficient than at frame level by using advanced technology and today’s faster routers. References [1]. Manjaih .D.H. Hanumanthappa.J,2008,A Study on Comparison and Contrast between IPv4 and IPv6 Feature sets.In Proceedings of ICCNS’08, 2008,Pune,297-302. [2]. Manjaih.D.H.,Hanumanthappa.J.,2008,Transition of IPv4 Network Applications to IPv6 Applications,In Proceedings of ICETiC- 09,2009,S.P.G.C. Nagar,VirudhaNagar-626 001,TamilNadu,INDIA-35-40.
Velammal College of Engineering and Technology, Madurai
[3] Manjaih.D.H.Hanumanthappa.J. 2009, IPv6 over Bluetooth: Security Aspects, Issues and its Challenges, In Proceedings of NCWNT-09, 2009, Nitte-574 110, Karnataka, INDIA–18-22. [4] Manjaih.D.H.,Hanumanthappa.J.,2009,Economical and Technical costs for the Transition of IPv4 –to-IPv6 Mechanisms [ETCTIPv4 to ETCTIPv6],In Proceedings of NCWNT-09,2009,Nitee574110,Karnataka,INDA-12-17. [5]. Manjaih.D.H. Hanumanthappa.J. 2009, Smooth porting process scenario during the IPv6 transition, in Proceedings of IACC’09, March 6-7, Patiala, Punjab, INDIA. [6] Manjaiah. D. H. and Hanumanthappa.J. IPv6 over IPv4 QoS metrics in 4G–Networks: Delay, Jitter, Packet Loss Performance, Throughput and Tunnel Discovery mechanisms. Proceedings of NCOWN-2009, RLJIT, Doddaballapur, Bangalore, Karnataka, INDIA, August-21-22-pp.122-137. [7] Hanumanthappa. J. And Manjaiah.D.H.”A Study on IPv6 in IPv4 Static Tunneling threat issues in 4G Networks using OOAD Class and Instance Diagrams” in Proceedings of the International Conference on Emerging trends in Computer science and Information technology(CSCIT-2010)organized by Dept.of.CS and Information Technology,Yeshwant ahavidyalaya,Nanded,Maharashtra,India,(2010). [8] Hanumanthappa.J. and Manjaiah.D.H.”An IPv4–to-IPv6 Threat reviews with Dual Stack Transition mechanism Considerations a Transitional threat model in 4G Wireless “in Proceedings of the International Conference on Emerging trends in Computer science and Information technology (CSCIT-2010)organized by Dept of CS and Information Technology,Yeshwant Mahavidyalaya,Nanded,Maharashtra,India,(2010). [9] Manjaiah.D.H. and Hanumanthappa.J,”IPv6 and IPv4 Threat review with Automatic Tunneling and Configuration Tunneling Considerations Transitional Model: A Case Study for University of Mysore Network”International Journal of Computer Science and Information Security (IJCSIS), Vol.3, (2009). [10] S.Deering and R. Hinden “Internet Protocol Version 6(IPv6) Specification”, RFC 2460, December 1998. [11]. J.Postel, INTERNET PROTOCOL, RFC 0791, September 1981.
Page 276
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[12] S.Tanenbaum, “Computer Networks”, Third Edition, Prentice Hall Inc., 1996, pp.686, 413-436,437-449. [13] Beerhouse A.Forouzan, Third Edition, “TCP/IP Protocol Suite”. [14] Atul Kahate, “Cryptography and Network Security”, Tata McGrawHill, 2003, pp-8-10. [15] Kurose.J. & Ross.K. (2005)Computer Networking: A top-down approach featuring the Internet.3rd Ed, (Addison Wesley). [16] S.Hagen:“IPv6 essentials“, Orielly, July 2002, ISBN-0-5960-01258. [17] Behrouz A Forouzan, Fourth edition Data communication and Networking, pp-7-8. [18] S.Shukla and N.W.Bergmann, “Single bit error correction implementation in CRC-16 on FPGA”, in proceedings of The IEEE International conference on Field-Programmable technology, pp-319-322,204. [19] J.Davies, “Understanding IPv6”, Microsoft press, Washington: Microsoft press, 2003. [20] T.Parnell, Building High Speed networks, First edition, Mc-GrawHill, 1999. [21] Raja Kumar Murugesan, Dr.Rahmat Budiarto, Prof.Sureswaran Ramadass, Enhanced performance of IPv6 packet transmission over Fibre.
Mr.Hanumanthappa. J. is Lecturer at the DoS in CS,University of Mysore, Manasagangothri, Mysore-06 and currently pursuing Ph.D in Computer Science and Engineering, from Mangalore University under the supervision of Dr.Manjaiah.D.H on entitled “Design and Implementation of IPv6 Transition Technologies for University of Mysore Network (6TTUoM)”. His teaching and Research interests include Computer Networks, Wireless and Sensor Networks, Mobile Ad-Hoc Networks, Intrusion detection System, Network Security and Cryptography, Internet Protocols, Mobile and Client Server Computing,Traffic management,Quality of Service,RFID,Bluetooth,Unix internals, Linux internal, Kernel Programming,Object Oriented Analysis and Design etc.His most recent research focus is in the areas of Internet Protocols and their applications.He received his Bachelor of Engineering Degree in Computer Science and Engineering from University B.D.T College of Engineering ,Davanagere,Karnataka(S),India(C),Kuvempu University,Shimoga in the year 1998 and Master of Technology in CS&Engineering from NITK Surathkal,Karnataka(S ),India (C) in the year 2003.He has been associated as a faculty of the Department of Studies in Computer Science since 2004.He has worked as lecturer at SIR.M.V.I.T,Y.D.I.T,S.V.I.T,of Bangalore. He has guided about 250 Project thesis for BE,B.Tech,M.Tech,MCA,MSc/MS.He has Published about 15 technical articles in International ,and National Peer reviewed conferences. He is a Life member of CSI, ISTE,AMIE,IAENG,Embedded networking group of TIFAC–CORE in Network Engineering,ACM,Computer Science Teachers
Velammal College of Engineering and Technology, Madurai
Association(CSTA),ISOC,IANA,IETF,IAB,IRTG,etc.H e is also a BOE Member of all the Universities of Karnataka,INDIA.He has also visited Republic of China as a Visiting Faculty of HUANG HUAI University of ZHUMADIAN,Central China, to teach Computer Science Subjects like OS and System Software and Software Engineering,Object Oriented Programming With C++,Multimedia Computing for B.Tech Students. in the year 2008.He has also visited Thailand and Hong Kong as a Tourist. Dr.Manjaiah.D.H D.H. is currently Reader and Chairman of BoS in both UG/PG in the Computer Science at Dept.of Computer Science,Mangalore University, and Mangalore.He is also the BoE Member of all Universities of Karnataka and other reputed universities in India.He received Ph.D degree from University of Mangalore, M.Tech. from NITK,Surathkal and B.E.,from Mysore University.Dr.Manjaiah.D.H D.H has an extensive academic,Industry and Research experience.He has worked at many technical bodies like IAENG,WASET,ISOC,CSI,ISTE,and ACS. He has authored more than -25 research papers in international conferences and reputed journals. He is the recipient of the several talks for his area of interest in many public occasions. He is an expert committee member of an AICTE and various technical bodies. He had written Kannada text book,with an entitled,”COMPUTER PARICHAYA”,for the benefits of all teaching and Students Community of Karnataka.Dr.Manjaiah D.H’s areas interest are Computer Networking & Sensor Networks, Mobile Communication, Operations Research, E-commerce, Internet Technology and Web Programming.
Aravinda.C.V.,currently pursuing M.Tech(I.T) K.S.O.U., Manasagangotri, Mysore-06.He received M.Sc ., M.Phil in Computer Science. He Worked has Lecturer at Various Institutions :1) CIST University Of Mysore Manasagangothri Mysore.2) Vidya Vikas Institute Of Engineering and Techonology Mysore.3) Govt.First Collge Srirangapatna,Mandya District .4) Govt First College Kollegal ,Chamrajanagar Distirct. 5) Worked has Technical Co-ordinator NIIT Gandhi Bazar,Bangalore. He has published two National Papers hosted by NITK Surthkal Mangalore, DK.
Page 277
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Localized CBIR for Indexing Image Databases D.Vijayalakshmi 1, P. Vijayalakshmi 2 1,2
Faculty of Kamaraj College of Engg and Technology, Virudhunagar, Tamilnadu, India. [email protected], [email protected], 3
P.Jeyapriya 3
Final B.Tech, IT, Kamaraj College of Engg and Technology, Virudhunagar.
Abstract: At present, research has been focused on image indexing techniques based on global feature distributions. However, these global distributions have limited discriminating power because they are unable to capture local image details. Localized image retrieval is required when it is preferred to index image databases based on the similarity metrics. Wavelet represents images at various resolution levels. A high wavelet coefficient at a coarse resolution corresponds to a region with high global variations and vice versa. In this paper, we present wavelet based localization for indexing image databases employing color and texture features. The feature vectors thus computed is unique and has sharp discrimination power in various image regions. The new system performs better in terms of computational time and storage requirements as compared to the global feature approaches. The potential applications include automatic image classification, video retrieval and video data management Keywords, Wavelet Decomposition, Image indexing, feature extraction, Matching Criteria, localization.
• I. INTRODUCTION Many picture libraries use keywords as their main form of retrieval – often using indexing schemes developed inhouse, which reflect the special nature of their collections. A good example of this is the system developed by Images to index their collection of stock photographs .Users should be able to explore images in a database or video clips by visual similarities. Structuring and visualizing digital images are based on their content similarities. Currently, many content based image retrieval techniques have been developed to incorporate higher-level feature extraction capabilities, but a lot of work remains to be done. Ultimately, feature-extraction techniques, combined with other techniques are expected to narrow down the gap between relatively primitive features extracted from images and high-level semantically-rich perceptions by humans. Many picture libraries use keywords as their main form of retrieval – often using indexing schemes developed in-house, which reflect the special nature of their collections. Index terms are assigned to the whole image, and the main objects. When discussing the indexing of images and videos, one needs to distinguish between systems which are geared to the formal description of the image and those concerned with subject indexing and
Velammal College of Engineering and Technology, Madurai
retrieval. The former is comparable to the bibliographical description of a book. However, there is still no one standard in use for image description, although much effort is being expended in this area by a range of organizations such as the Museum Documentation Association. Access to a desired image from a repository might thus involve a search for images depicting specific types of object or scene, evoking a particular atmosphere, or simply containing a specific texture or pattern. Potentially, images have many types of attribute which could be used for retrieval, including: • the presence of named individuals, locations, or events (e.g. the Queen greeting a crowd); • subjective emotions one might associate with the image (e.g. happiness); meta data such as who created the image, The last decade has seen the appearance of a number of commercial image data management systems. These systems normally store representations of pictorial documents (such as photographs, prints, paintings, drawings, illustrations, slides, video clips, and so on) in static archival databases, and incorporate multimedia database management systems in the storage of, and provision of wider access to, these repositories. Current image indexing techniques have much strength. Keyword indexing has high expressive power – it can be used to describe almost any aspect of image content. It is in principle easily extensible to accommodate new concepts, and can be used to describe image content at varying degrees of complexity. There is a wide range of available text retrieval software to automate the actual process of searching. But the process of manual indexing, whether by keywords or classification codes, suffers from two significant drawbacks. Firstly, it is inherently very labor-intensive. Indexing times quoted in the literature range from about 7 minutes per image for stock photographs at Getty Images, using their in-house system, to more than 40 minutes per image. Secondly, manual indexing does not appear to be particularly reliable as a means of subject retrieval of images. Markey had found that, in a review of inter-indexer consistency, there were wide disparities in the keywords that different individuals assigned
Page 278
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
to the same picture. The advantage of all these techniques is that they can describe an image at varying levels of detail (useful in natural scenes where the objects of interest may appear in a variety of guises), and avoid the need to segment the image into regions of interest before shape descriptors can be computed. Despite recent advances in techniques for image segmentation, this remains a troublesome problem. Wavelet Decomposition The continuous wavelet transform of a 1-D signal f(x) is defined as
(W a f )(b ) = ∫ ( x )Ψ
*
a, b
( x )dx
----------1 where the wavelet b , a Ψ is computed from the mother wavelet Ψ by translation and dilation,
Ψa , b( x ) =
1 Ψ (( x − a ) / b ) |a| ----------
-2 the mother wavelet Ψ satisfies the constraint of having zero mean. The eq. (1) can be discretized by restraining ’a’ and ’b’ to a discrete lattice. Typically it is imposed that the transform should be non-redundant, complete and constitutes a multi resolution representation of the original signal. In practice, the transform is computed by applying a separable filter bank to the image where denotes the convolution operator, ) 2 , 1 ↓ ( 1 , 2 ↓ denotes the down sampling along the rows (columns) and I = A0 is the original image, H and G are low pass and high pass filters, respectively. The n A is obtained by low pass filtering and is referred to as the low resolution (Approximation) image at scale n. The 3 n 2 n 1 n D , D , D are obtained by band pass filtering in a specific direction(Horizontal, Vertical and Diagonal, respectively) and thus contain directional detail information and is referred to as high resolution(Detail) images at scale n. The original image I is thus represented by a set of sub images at several scales. This decomposition is called “Pyramidal wavelet transform” decomposition or discrete wavelet decomposition (DWT). Every detail sub image contains information of a specific scale and orientation. The spatial information is retained within the sub image. In the present paper, the features are obtained using Haar Wavelet (Fig. 1.), which is given by
⎧1 0<=t <=1/2 ⎪ Ψ(t) =⎨−1 1/2<=t <=1 ⎪0 otherwise ⎩ ---------------3
Velammal College of Engineering and Technology, Madurai
2. RELATED WORK Visualization technique of images with more continuous scenes described in this article have a wide range of potential applications, for example, data mining in remote sensing images and image retrieval from film and video archives. This methodology is suitable on a sample of images with more continuous scenes, for example, video segments, so that we will be able to keep track of the impact of various feature-extraction techniques more closely [1]. A proposal is presented for constructing an efficient image sharing system, in which a user is able to interactively search interesting images by content-based methods in a peer-to-peer network. A distributed variant of the Tree-Structured Self- Organizing Maps (TS-SOM) by incorporating the essence of DHT is presented. Compared with the existing approaches, this method has the following advantages: many operations can be performed on the 2-D lattice instead of the highdimensional feature space, which significantly saves the computing resource; the TS-SOM-like hierarchical search speeds up the information lookup and facilitates the use of large Self-Organizing Maps; a query can involve more than one feature simultaneously. The current design is considered for the tree-structured index with fixed levels and fixed size at each level. It is not feasible for some huge other network communities during their evolving stages. Many details, for instance the image normalization and the system strategy have not been considered due to the limit of time and space [2]. In another paper, an extraction of signatures, to determine a comparison rule, including a querying scheme and definition of a similarity measure between images is proposed. Several querying schemes are available: regionbased searching, where the retrieval is based on a particular region in the image, or searching by specifying the color histogram or object shape of the images, significant semantic information is lost. Principal component analysis is used to represent and retrieve images on the basis of content which reduces the dimensionality of the search to a basis set of prototype images that best describes the images. Each image is described by its projection on the basis set; a match to a query image is determined by comparing its projection vector on the basis set with that of the images in the database, the detection performance is transform variant [3]. Here, images are normalized to line up the eyes, mouths and other features. The eigenvectors of the covariance matrix of the face image vectors are then extracted. These eigenvectors are called eigen faces. Raw features are pixel intensity values Ai. Each image has n × m features, where n, m are the dimensions of the image. Each image is encoded as a vector G of these features. Compute “mean” face in training set and detected [4]. A technique for automatic face recognition based on 2D Discrete Cosine Transform (2D-DCT) together with Principal Component Analysis (PCA) is suggested. Applying the DCT to an input sequence decomposes it into a
Page 279
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
weighted sum of basis cosine sequences. A block size of 16×16, 32×32 and 64×64 are chosen to be further transformed using PCA. The DCT transformed face images using the different block sizes are transformed into a lower order feature size using the PCA to get a feature vector of 25 features. Two other block sizes are chosen to be further transformed using PCA; namely a 32×32 and a 64×64 DCT coefficients so as to see the effect of changing the block size on the recognition rate. Also, a feature vector of 25 features is calculated in each case. The recognition rate depends on a chosen threshold based on the security level of the application. Two important aspects are investigated namely the False Accept rate (FAR) and the False Reject Rate (FRR). The FAR is the success probability of an unauthorized user to be falsely recognized as a legally registered user [5], [6].A technique for representing shapes for the purpose of image retrieval is proposed. The proposed representation is a contour based approach. Canny operator is used to detect the edge points of the image. The contour of the image is traced by scanning the edge image and re-sampling is done to avoid the discontinuities in the contour representation. After the object contour has been detected for an object, the central point of the object, the centroid of an object is computed. The resulting image is swept line by line and the neighbor of every pixel is explored to detect the number of surrounding points and to derive the shape features. In this method, when the color image database is considered for comparison 71.17% retrieval efficiency is obtained for top 20 retrievals. Shape feature requires extraction of a region or object, in itself a complicated problem. Accurate shape detection requires human intervention because methods like segmentation are very difficult to completely automate [7]. Retrieving images based on color similarity is achieved by computing a color histogram for each image that identifies the proportion of pixels within an image holding specific values. Current research is attempting to segment color proportion by region and by spatial relationship among several color regions. The main idea of appearance- based approach is to represent the images in terms of their projection onto a low-dimensional space called eigen space is presented. The projection of the images onto this space is called the eigen images. Principal components as sequence of image vectors are computed incrementally, without estimating the covariance matrix and transforming these principal components to the independent directions that maximize the non–Gaussianity of the source. This algorithm takes input image finds the non-Gaussian vector (eigenvector) which is passed as input to the ICA (independent component analysis) algorithm. The nonGaussian components will be updated using an updating rule from the previous component values in a recursive way. IPCA (incremental principal component analysis) returns the estimated eigenvectors as a matrix that represents subspaces of data and the corresponding eigen values as a
Velammal College of Engineering and Technology, Madurai
row vector, Fast ICA searches for the independent directions where the projections of the input data vectors will maximize the non-gaussianity. The nearest neighbor algorithm is used to evaluate the object recognition technique [8], [9]. Four different approaches to color texture analysis are tested on the classification of images from the VisTex database. The first method employs multi spectral approach, in which texture features are extracted from each channel of the RGB color space. The second method uses HSV color space in which texture features are extracted from the luminance channel V and color features from the chromaticity channels H and S. The third method uses YCbCr color space, in which texture features are extracted from the luminance channel Y and color features from the chromaticity channels Cb and Cr. The last one uses gray scale texture features computed for a color Image [10]. In another paper, we present a localized CBIR system, ACCIO! , that uses labeled images in conjunction with a multiple-instance learning algorithm to first identify the desired object and weight the features accordingly, and then to rank images in the database using a similarity measure that is based upon only the relevant portions of the image [11]. A number of indexing schemes use classification codes rather than keywords or subject descriptors to describe image content, as these can give a greater degree of language independence and show concept hierarchies more clearly. Local feature-based image matching was done in two steps. The first step of the feature detection is key point or interest point detection. The second step involved computing descriptors for each detected interest point. Descriptors were used for matching key points of the input image with key points in the logo database. For logo recognition the descriptor had to be distinctive and at the same time robust to changes in viewing conditions. The features were to be resilient to changes in illumination, image noise, uniform scaling, rotation, and minor changes in viewing direction. The descriptor must minimize the probability of mismatch and finally, it should also be relatively easy and fast to extract the features and compare them against a large database of local features. Based on these requirements and the reported robustness in other research projects, Scale Invariant Feature Transform (SIFT) was used in [12] 3. PROPOSED WORK Here , we have used YUV colour spaces in order to obtain one channel containing the luminance information and two others containing chrominance information. The conversion of RGB to YCbCr color model is shown in the eqn (4). Discrete wave let decomposition is employed and is shown in the Fig 3.1. Here, we have chosen Haar wavelet for wavelet decomposition. And its representation is shown in the Fig 3.2. The four quadrants coefficients of the images are obtained. The correlation between approximation and horizontal coefficients, approximation and vertical coefficients and approximation and diagonal coefficients are computed. A
Page 280
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
high wavelet coefficient at a coarse resolution corresponds to a region with high global variations and vice versa. In this paper, we present wavelet based localization for indexing image databases employing color and texture features In addition to colour and texture features and other statical features are also computed. The feature vectors have sharp discrimination power in various image regions. Texture features are computed from the luminance channel. The first order statistical features namely, mean and standard deviation, are computed from the chrominance channel. The following equations transform RGB in [0, 1] to Y Cb Cr in [0,255].
4. FEATURE EXTRACTION The co-occurrence histograms are constructed across different wavelet coefficients of an image and its complement decomposed up to 1-level. All combinations of the image quadrants of wavelet coefficients of image and its complement are considered to compute the correlation between various coefficients. From the normalized cumulative histogram, color and texture features are extracted employing equation ( 5), (6) and (7). 256
Mean :
μ
∑ nchi nch =
i =1 256
………….5
Slope : S nch = Slope of the regression line fitted
….....4
. across the sample points - --------------6 256
Mean Deviation: Dnch =
∑|nchi −μnch|
i=1
256
-----------7 • • • • • • • • • • FIG . 3.1. WAVELET DECOMPOSITION
Algorithm indexing and Retrieval Get image Apply multilevel wavelet decomposition Extract color and texture features Compute feature vector Train all images Get Test image Localize a Image segment to be indexed Compute test feature vector Apply matching criteria and search Retrieve and index all images that matches 5. ANALYSIS OF EXPERIMENTAL RESULTS
In this paper, wavelet based localization for Indexing Image databases are implemented. The paper implemented in RSI IDL 6.3 environment. The project is tested on standard images for indexing the color face images shown in the Fig 5.1. Localized portion of the face images used to index the image database. They are nose and eye images and they are shown in Fig 5.2 and 5.3. The experimental results of indexing for nose and eye image are shown in Fig 5.5 and 5.6. The objective metrics of retrieval algorithm are computed and are shown in Table 5.2. Also, success and failure rate of retrieval algorithm are computed and shown in Table in 5.1. FIG 3.2 HAAR WAVELET REPRESENTATION
Velammal College of Engineering and Technology, Madurai
Page 281
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
FIG 5.3 TRAINED AND UNTRAINED TEST EYE IMAGES
FIG 5.4 SNAPSHOT OF OUTPUT SCREEN
FIG 5.1 TRAINED FACE IMAGE DATA BASE
FIG 5.5 INDEXED OUTPUT FOR NOSE IMAGES
Fig 5.2 TRAINED AND UNTRAINED NOSE IMAGES
Velammal College of Engineering and Technology, Madurai
Page 282
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[5] Application of DCT Blocks with Principal Component Analysis for Face Recognition MAHA SHARKAS, Proceedings of the 5th WSEAS Int. Conf. on Signal, Speech And Image Processing, Corfu, Greece, August 17-19, 2005 (pp107-111). [6] Face Recognition Using the Discrete Cosine Transform, Ziad M. Hafed and Martin D. Levine, International Journal of Computer Vision 43(3), 167–188, 2001, c _ 2001 Kluwer Academic Publishers, Netherlands [7] Image Retrieval using Shape Feature, S.Arivazhagan, L.Ganesan, S.Selvani dhyananthan , International Journal Of Imaging Science And Engineering (IJISE) , ISSN: 1934-9955,Vol.1,No.3, July 2007. [8] Appearance Based 3D Object Recognition Using IPCA-ICA, V.K Ananthashayana and Asha. V, International Archives of the Photogrammetry, Remote Sensing and spatial Information Sciences. Vol. XXXVII. Part B1, Beijing 2008. [9] Statistical Appearance Models for Automatic Pose Invariant Face Recognition, M.Saquib Sarfraz and Olaf Hellwich, 978-1-4244-2154-1/08/ c2008 IEEE. [10] Localized Content Based Image Retrieval, Rouhollah Rahmani, Sally A. Goldman, Hui Zhang, Sharath R. Cholleti, and Jason E. Fritts IEEE Transactions On Pattern Analysis And Machine Intelligence, Special Issue, November 2008.
FIG 5.6 INDEXED
OUTPUT FOR
EYE IMAGES
6. CONCLUSION A novel method for indexing the image databases employing the wavelet features obtained from the coefficients of an image has been presented. The experimental results obtained is good in Y Cb Cr color spaces and provides better retrieval results. The Haar wavelets are chosen since they are more effective in texture representation compared to other wavelets. From the Table 5.1 and 5.2 , the robustness of the proposed feature set for indexing image data base is proved. The experimental results demonstrate the worth of the proposed method for indexing of image databases and the results are encouraging. It can be enhanced in future as an intelligent system by employing genetic algorithm or neural Networks. 7. REFERENCES [1] Similarity-Based Image Browsing, Chaomei Chen, George Gagaudakis, Paul Rosin Brunel University, Cardiff University, U.K, funded by the British research council EPSRC (GR/L61088 and GR/L94628). [2]
Interactive Content-based Image Retrieval in the Peer-to-peer Network Using SelfOrganizing Maps, Zhirong Yang, HUT T-110.551 Seminar on Internetworking 2005- 04-26/27 [3] Principal Component Analysis for Content-based Image Retrieval, Usha Sinha, and Hooshang Kangarloo, Radio graphics 2002; 22:12711289.), © RSNA, 2002
[11] Wavelet Based Features for Color Texture Classification with Application to Cbir P.S.Hiremath, S. Shivashankar, and Jagadeesh Pujari, IJCSNS International Journal of Computer Science and Network Security, VOL.6 No.9A, September 2008 [12] Content-based Image Retrieval, John Eakins, Margaret Graham, University of Northumbria at Newcastle
Precision rate %
Recall Rate %
CBIR
90.2
92
FGNET
85.3
89
IRDS
85
90
Database
TABLE 5.1 COMPARISON OF EXPERIMENTAL RESULTS NOSE
Database
Success
Failure rate
rate in %
in %
CBIR
90.2
9.8
FGNET
85.3
14.7
IRDS
85
15
[4].Content-Based Image Retrieval, http://www.cs.cmu.edu/afs/cs/academic/class/15385-06/lectures/.
Velammal College of Engineering and Technology, Madurai
Page 283
Ta
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Architecture Evaluation for Web Service Security Policy Joshi.vinayak.B [1]
Dr. Manjaiah D. H [2]
Hanumathappa.J [3]
Research scholar
Reader Computer Science
Lecture, computer Science
Mangalore University
Mangalore University
University of Mysore
Mangalore, India [email protected]
Mangalore, India [email protected]
Mysore, India [email protected]
Abstract Web Services can be published, discovered and invoked over the web. Web Services can be implemented in any available technology but they are accessible through a standard protocol. With web services being accepted and deployed in both research and industrial areas, the security related issues become important. In this paper, architecture evaluation for web service on negotiating a mutually acceptable security policy based on WSDL (web service description language). It allows a service consumer to discover and retrieve a service-provider’s security policy for service requests and allows a service-consumer to send its own security policy for service responses to the service-provider. The service consumer combines its own policy for service requests with that of the service provider to obtain the applied security policy for requests, which specifies the set of security operations that the consumer must perform on the request. The combining takes place in such a way that the applied security policy is consistent with both the consumer’s and provider’s security policies. The service provider also combines its own policy for responses with that of the consumer, to obtain the applied security policy for responses [1].
Nayak.Ramesh.Sunder [4] A.P, ShreeDevi Institute of Technology kenjar, Mangalore, India [email protected]
have more control for the access and the updating of information, and the reliability of the registry content. We concentrate here on one key issue, providing security in Web services architecture. In this paper, we evaluated a technique for deriving mutually acceptable quality of protection for exchanges between a service provider and a service consumer. The WSDL document of a web service would include a security policy description representing the types of security operations that are required and supported by the Web-service for its SOAP message exchanges with consumers. However, since Web-service consumers may themselves be implemented as a web service, both the consumer and provider of the service may have a security policy defined in their WSDL documents. This implies that there is a need for the web service consumers and providers to agree on the applied security policy to be used to protect the SOAP message exchange between them [6].
2. System Overview 1. Introduction
Web services are reusable Web components with their programmatic interfaces described in WSDL [4].WSDL is a XML format standard for describing the interface of a web service. The WSDL description gives information about what exactly a web service does, how to invoke its functions and where to find it. Universal Description, Discovery, and Integration (UDDI) is a registry standard, which allows organizations to publish and discover Web Services using standardised methods. The UDDI is an industry initiative to provide a platform-independent framework for creating a UDDI Business Registry [9]. There are currently several providers of UDDI registers called UDDI Operators. The UDDI specification defines a set of datastructures and an Application Programming Interface (API) for registering and finding businesses. The UDDI specification also allows organizations to create their own UDDI registries in order to
Velammal College of Engineering and Technology, Madurai
2.1 Agent Based Web Service Discovery Web service discovery can be performed based on a web service security policy using agents. It consists of a ser-vice provider, a service consumer and a UDDI to include a discovery agent and security agent and use an augmented UDDI that contains security policy information to allow secure web service discovery (as shown in Figure1). The discovery agent acts as a broker between a service consumer, a UDDI registry and a security policy that helps to discover secure web services that satisfy the consumer security requirements.
2.2 Security Agent The security agent describes the security requirement that service provider needs to be registering their WSDL into the
Page 284
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
registry. Web service security test case describe a testing methodology for web service security and outline a process that can be adopted to evaluate web service security requirements. Test case can be classified according to different categories of threat faced by web services. Security policy can be represented in the UDDI registry by a tModel, which is typically used to specify the security policy details of a web service. In the tModel, each security policy is represented by a keyed Reference, which contains the name of a security attribute as key Name, and associated method.
3. Process Model The model works with the exception that the containers hosting the consumer and provider classes emit a SOAP message, which is intercepted by the security service. The consumer and provider classes could provide the <Security Mechanisms> and <Security Services> elements to their security services, in a WSS header, with the security service module identified as the target role. Alternatively, the security service could obtain the <Security Mechanisms> and <Security Services> elements directly on its own. WSDL binding to support the publication of the security policy in the case that a provider offers a secured interface. Specifically, elements called <Security Mechanisms> and <Security Services> are associated with message definitions in the service’s WSDL instance. In addition, we specify a WSS header for conveying the consumer’s policy for service responses using the same element definitions. The <Security Mechanisms> element describes a set of security mechanism, which may be applied to one or more nodes of the SOAP document. Additionally, parameters of a security mechanism may be specified in the element [1].
Figure 1. Web service discovery using agents 2.2 Discovery Agent A discovery agent receives service requests containing specifications for functional and security requirements from the service consumer, finds the services that meet the specified criteria, and then returns a list of services to the consumer in the order of priority. Discovery should be based on web service security polices for concerned request. The list of available services will be return to the service consumer in order. This avoids the overhead of discovery mechanism to search secure web services over UDDI registry for consumers needs. This approach that allows a service consumer to discover and retrieve a service provider’s security policy for service requests and allows a service consumer to send its own security policy for service responses to the service provider. The service consumer combines its own policy for service requests with that of the service provider to obtain the applied security policy for requests, which specifies the set of security operations that the consumer must perform on the request. The combining takes place in such a way that the applied security policy is consistent with both the consumer and provider security policies [1]. Likewise, the service provider combines its own policy for responses with that of the consumer, to obtain the applied security policy for responses.
Velammal College of Engineering and Technology, Madurai
Figure 2. Model for web service security policy
Page 285
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Input: User request with specified security criteria Output: Secure match set of services from UDDI u(h): Select all the services which matches the functionality requirements of user request that exists in UDDI. Let u(h)={ws1,ws2…..wsn}wss (h): Choose the set of services which have been registered in UDDI with security specifications. Let wss(h)={ws1(s), ws2(s), ….wsi(s)} Step 1 : For each web services wsi in u(h) //find the services that match the QOS requirements Step 2: QoS based Selection=Qos_Match (u(h) , QWS Parameters); Step 3 : If wss(h) requirements specified then Step4 :{Secuirty_Search=Security_Match (QoS_Search,wss(h) specified); Step5 : If wss(h) ratings found then //find the services that matches security criteria Step6 : return output of available services in wssi in u (h) according to criteria rank} Step7 :{Else return the output of available services wsi in u (h)}
step), testing (2step), and reporting (1step). It generates a number of outputs such as: a prioritised list of quality attributes, a list of architectural decisions made, a map linking architectural decisions to quality attributes, lists of risks and non-risks, and lists of sensitivities and tradeoffs. 4.2 ATAM Process Phase 1 Step 1 - Presenting the ATAM Process ATAM stands for Architecture Tradeoff Analysis Method. It is a method that tells how well an architecture satisfies particular goals by providing insight of how quality goals interact and how they trade off. Step 2 - Present Business Drivers •Due to the increase of business-to-business communication between different organizations over internet resources, the current architecture will provide secure service connection establishment between service consumer and provider with added security policy. •Suggest the service provider to accept the service consumer requirements to add new security features to perform secure tasks. Architecture Drivers The major quality attribute are as below Pri ority 1
Quality Attribute Driver Security
The proposed architecture is evaluated by the Software Architecture Tradeoff Analysis Method (ATAM).All the scenarios corresponding to each application of the secure web service discovery and retrieval are listed and evaluated. The steps corresponding to ATAM for evaluating the web service security are described as follows
2
Interoperability
4.1 ATAM: Secure Web Service Discovery
3
Availability
We put ATAM to the test on our architecture and discuss the findings based on the outputs generated which include lists of risks, non-risks, sensitivities, and tradeoffs made. The findings show that secure web service discovery and retrieval architecture can greatly benefit from using ATAM by revealing its strengths and weaknesses before evolving the architecture further. ATM comprises of 9 steps grouped into four phases: presentation (3step), investigation and analysis (3
4
Performance
Figure 3. Service discovery algorithm 4. Evaluation of Proposed Architecture
Velammal College of Engineering and Technology, Madurai
Rationale It is a major concern to this area of the architecture because it should support authentication, encryption, Integrity and non reputation over different communication channel and platform model. The registry must be able to interact with the provider servers to be able to operate. The service should be in need to run at any time even system failure occurs over registry or service provider. Continues user request will affect the system response. we will establish the user connection based on
Page 286
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
token request.
Step 3 - Presenting the ATAM Process Step 4 - Identify Architecture Approaches
Integrity
Authentication
Non-reputation
Layering
Rationale
Trade-offs
It organizes the system in hierarchical structure that allows for easy system modification.
Security potentially reduced risk
Scenario#: 1 Attribute(s) Environment Stimulus Arch decision Sensitivity Tradeoff
Step 5 - Quality Attribute Utility Tree I=Importance D=Difficulty to achieve H, M, L = high, medium, low
Risk Quality Attribut e Performance
Attribute Refinement
Scenarios
(I, D)
Response Time
Overheads of trapping user events must be imperceptible to users. At peak load the service should be able to perform transaction with limited time. Users' information shall only be visible to
(M, L)
Throughput
Security
Confidentiality
(H, M)
(H, M)
(H, M)
Step 6 - Architecture elicitation and analysis
Important Approaches of the Secure Web Service Architectural Approach
administrative users of the system and it is encrypted before transmitting to the server. The System resists unauthorized intrusion and modification of data. This enables the user to access the service with required token It verifies the signed information from valid user
(M,M)
Non risk
Scenario#: 2 Attribute(s) Environment Stimulus
(H, L)
Velammal College of Engineering and Technology, Madurai
Response
Scenario: Separate User Interface Usability Normal operations Service Consumer can add their own required security policy to the existing security policy of service provider Reasoning The present of physical buttons and clear messages display. Complexity of the physical interface design. Due to the limited space on panel or the box, design must have to be simple and neat but understandable to the user. The label on the buttons fades out after being pressed many times. This may apply same policy to server more than once. Not apply here.
Scenario: Language neutral Security Normal operations The proposed architecture should support for implementation in any language. Changing the implementation language should not affect any components in architecture. Not apply here
Page 287
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Arch decision Sensitivity Tradeoff Risk Non risk
Reasoning It support all data format conversion between different platform Need different data format for any platform New data format cannot be supported Not apply here.
Arch decision Sensitivity Tradeoff Risk Non risk
Scenario#: 3 Attribute(s) Environment Stimulus Response Arch decision Sensitivity Tradeoff
Risk
Non risk Scenario#: 4 Attribute(s) Environment Stimulus
Response Arch decision Sensitivity Tradeoff Risk Non risk Scenario#: 5 Attribute(s) Environment Stimulus Response
Scenario: Confidentiality Security Normal operations Certificate authority has to provide security token to authenticate Intermediary has no way to read the message while establishing connection with service provider Reasoning The encryption algorithm. More computation time and resource used. Performance is the tradeoff with Security. Not apply to architecture, but the Encryption algorithm itself, if it is not complex enough, could be hacked by brute force. Not apply here. Scenario: Integrity Security Normal operations Unauthorized user without security token cannot able to access the service available in the registry Identity Certificate are required to verify the user authentication Reasoning Identity certificate Need resource to map data, Performance, but not too much. Provide certificate to user in more secret Not apply here. Scenario: Authorization Security Normal operations Service ticket has way to establish trust relationship with more than one security domain utility certificate are required to
Velammal College of Engineering and Technology, Madurai
Scenario#: 6 Attribute(s) Environment Stimulus Response Arch decision Sensitivity Tradeoff Risk Non risk
verify the user authorization Reasoning Utility certificate More computation time and resource used, Performance, but not too much. Provide certificate to user in more secret Not apply here. Scenario: Non-reputation Security Normal operations Utility has key certificate to form signed message to verify the user utility key certificate are required to verify the user sign information Reasoning Utility key certificate Need signed key information for operation response Provide certificate to user in more secret Not apply here.
Risk themes: •Heavily load on the service provider web server leads to return of result to the consumer must be unavailable. This leads to server unavailable. •Certificate Verifier should verify their security token before access the service available in the registry. While during modification of the web service by the service provider add on failure may occur. Required service may be unavailable to the consumer for particular time. Inconsistencies: •Security has no way to identify the web security policies in efficient manner. •Service provided has to provide secure interface to add new security features with the existing provided security policy for applied security policy otherwise consumer has no way to add their own security requirements. •It is risky for transferring service ticket to consumer for accessing the service that available in the registry. This may be available to other without proper criteria.
Step 7: Scenario Prioritization The following table prioritizes the Quality Scenarios for the secure web service discovery and retrieval architecture. The Scenario # refers to the scenario being referenced.
Page 288
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
5. Conclusions Universal Description Discovery and Integration has no way to identify the secure web services when multiple service providers are now providing similar functional services. With an increasing number of web services providing similar functionalities, security has become one of the most critical issues. An evaluated architecture called agent based web service discovery to automate secure web service discovery for negotiating a mutually acceptable security policy based on WSDL for both consumer and provider in dynamic nature. 6. References [13] Zahid Ahmed, Martijn de Boer,, Monica Martin, Prateek
Mishra, Dale Moberg, “Web-Services Security Quality of Protection”, Version 0.9 22 Nov 2002.
[14] Kassem Saleh and Maryam Habil, “The Security
Requirements Behavior Model for Trustworthy Software”, International MCETECH Conference on e-Technologies 2008 pp 235 - 238.
[21] Tiantian Zhang, Junzhou Luo, Weining Kong “An Agent-
based Web Service Searching Model”, 9th International Conference on Computer Supported Cooperative Work in Design Proceedings pp 390 – 295.
JOSHI VINAYAK.B is currently doing research in the computer science at Dept. of Computer Science, Mangalore University, and Mangalore. He received M.Tech. From Mysore university. and master of computer science from Karnataka university Dharwad joshi has already published 4 international conference and 5 national conference papers..He has an extensive academic, Industry and Research experience. His areas interest are Computer Networking & Sensor Networks, Mobile Communication, security and reliability, E-commerce, Internet Technology and Web Programming
[15] Janette Hicks, Madhusudhan Govindaraju, Weiyi Meng,
“Enhancing Discovery of Web Services through Optimized Algorithms” IEEE International Conference on Granular Computing 2007 pp 685 - 698.
[16] Colin Atkinson, Philipp Bostan, Oliver Hummel and
Dietmar Stoll, “A Practical Approach to Web Service Discovery and Retrieval”, IEEE International Conference on Web Services (ICWS 2007).
[17] Janette Hicks, Madhusudhan Govindaraju, Weiyi Meng,
“Search Algorithms for Discovery of Web Services”, IEEE International Conference on Web Services (ICWS 2007).
[18] Artem Vorobiev and Jun Han, “Specifying Dynamic
Security Properties of Web Service Based Systems”, IEEE Computer Society 2006.
[19] Slim Trabelsi Jean-Christphe Pazzaglia Yves Roudier,
“Secure Web Service discovery: overcoming challenges of ubiquitous computing”, Proceedings of the European Conference on Web Services (ECOWS'06).
Dr. MANJAIAH.D.H is currently Reader and Chairman, Department of Computer Science Mangalore University, Mangalore. He received PhD degree from University of Mangalore, M.Tech. From NITK, Surathkal and B.E., from Mysore University. Dr.Manjaiah D.H has an extensive academic, Industry and Research experience. He has worked at many technical bodies like IAENG, WASET, ISOC, CSI, ISTE, and ACS. He has authored more than - 45 research papers in international conferences and reputed journals. He is the recipient of the several talks for his area of interest in many public occasions and International and National conferences. He is an expert committee member of an AICTE and various technical bodies. Dr .Manjaiah. D.H’s areas interest are Computer Networking & Sensor Networks, Mobile Communication, Operations Research, E-commerce, Internet Technology and Web Programming.
[20] David Geer, “Taking Steps to Secure Web Services”,
Technology News October 2003.
Velammal College of Engineering and Technology, Madurai
Page 289
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Mr.HANUMANTHAPPA.J. is Lecturer at the DoS in CS,University of Mysore,Manasagangothri,Mysore-06 and currently pursuing Ph.D in Computer Science and Engineering, from Mangalore University under the supervision of Dr.Manjaiah.D.H on entitled “Design and Implementation of IPv4–to- IPv6 Transition Scenarios for 4G-Networks”.
Nayak Ramesh Sunder is currently working as Assistant Professor in department of Cs & E in Shreedevi Institute of technology, Kenjar, Mangalore. He received M.Tech. From Mysore University. and Bachelor of Engineering from Karnataka university Dharwad. Ramesh has already published 2 national conference papers..He has 10years of academic, and Research experience. His areas interests are Computer Networking & Mobile Communication, security and reliability, E-commerce, Internet Technology and Web Programming
Velammal College of Engineering and Technology, Madurai
Page 290
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Rule Analysis Based On Rough Set Data Mining Technique P.Ramasubramanian Professor & Head, Dept. of CSE Dr.G.U.Pope College of Engg., Sawyerpuram – 628251. Tamilnadu – India. [email protected]
V.Sureshkumar Professor, Dept. of CSE Velammal Engg. College, Chennai. Tamilnadu – India. [email protected]
P.Alli Professor & Head Dept. of CSE Velammal College of Engg & Tecchnology, Madurai. Tamilnadu – India. [email protected]
ABSTRACT Rough Set theory is a new mathematical tool to deal with representation, learning, vagueness, uncertainty and generalization of knowledge. It has been used in machine learning, knowledge discovery, decision support systems and pattern recognition. It is essential for professional colleges to improve competitive advantage and increase in placement performance. Students are a key factor for an institution’s success. In this paper, we use a ROSE system to implement the rule generation for students placements. Keywords: Data Mining, Decision Support, Human resource development, Knowledge Discovery, Rough set theory.
1. INTRODUCTION Many researchers provide important methods for human resource management. Since 80’s people have been product-oriented; most employees just need skills to produce, which is the idea of job-based human resource management. In 1973, the evaluation idea and technology was "work as the center". Today students’ do not need to pursue only their studies but also to undergo soft skills training. The objective of this research is to find out a compromising solution that satisfies both the institution and students, and finds out what kind of features and behaviors of students who can build good relationships with an institution. The results of this research will be used to guide institution as they are useful in providing a good strategy for human resource development and customer relationship management This paper is organized as follows: Section 2 reviews basic concepts of rough sets. The mathematical model employed here is briefly illustrated. Section 3 is spent to explain the problem tracked in this paper. Its results are discussed in Section 4. At the end in section 5 several remarks of this paper is given. 2. INTRODUCTION TO ROUGH SET
Velammal College of Engineering & Technology, Madurai
Z. Pawlak, a Polish mathematician, put forward Rough Sets Theory (RST) in 1980, which is a data reasoning method. In recent years, it has been rapidly developed in the field of intelligent information processing. Rough sets theory has many advantages. For instance, it provides efficient algorithms for finding hidden patterns in data, finds minimal sets of data (data reduction), evaluates significance of data, and generates minimal sets of decision rules from data. It is easy to understand and offer straightforward interpretation of results [3–7]. The rough sets theory has fundamental importance in artificial intelligence (AI) and cognitive science, especially in the areas of machine learning, knowledge acquisition, decision analysis, knowledge discovery inductive reasoning, pattern recognition in databases, expert systems, decision support systems, medical diagnosis, engineering reliability, machine diagnosis, travel demand analysis and data mining. Rough set theory provides a new different mathematical approach to analyze the uncertainty, and with rough sets we can classify imperfect data or information easily. We can discover the results in terms of decision rules. So in this research, we use rough set theory to analyze the human resource problem. Information System: An information system, denoted by IS, and defined by IS = (U,A,V,f), where U consists of set of finite objects and is named a universe and A is a finite set of attributes {a1,a2, …,an}. Each attribute a belongs to a set A, that is, a ∈ A. fa: U → Va, where Va is a set of values of attributes. It is named a domain of attribute a. In Table 1, there are 10 objects. The set of condition attributes is {English Medium, Financial Status, and Soft Skills), the set of decision attributes is {Placement}.
Page 291
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Example: Table 1: Sample Data Set
Objects
Attributes a1 English Medium
1 2 3 4 5 6 7 8 9 10
Yes No Yes Yes No Yes Yes Yes Yes No
a2 Financial status Low Low Medium High Medium Medium Low Medium Medium High
The value of the English Medium has two values namely Yes or No. The value of the Financial Status is calculated on the basis of Parents monthly income and additional income. The value of the Personality Development is calculated on the basis of student performances in Subject depth, Communication skills, Participation in seminar, Public, social activities, and Co-curricular activities. In the above table, U= {1,2, 3, 4, 5, 6, 7, 8, 9, 10} A = {a1, a2, a3} a1 = {Yes, No} a2 = {Low, Medium, High} a3 = {Low, Medium, High} D = {Low, Medium, High} Lower and Upper Approximations A method to analyze rough set is based on the two basic concepts that are lower and upper approximations. Let U be any finite universe of discourse. Let R be any equivalence relation defined on U. The pair (U,R) is called the approximation space. This collection of equivalence classes of R is called as knowledge base. Let X is a subset of elements in universe U, that is, X ⊂ U. Now, we consider a subset B in Va, i.e., B ⊆ Va. The lower approximation of P, denoted as PX , can be defined by the union of all elementary sets xi contained in X.
Velammal College of Engineering & Technology, Madurai
a3 Soft Skills
Decision D1 Placement
Medium Low High Low Low Medium High High Low Medium
High Low Medium Low Low Medium High High Medium Low
The upper approximation of P, denoted as PX , can be defined by a non-empty intersection of all elementary sets xi contained in X. The lower and upper approximations are defined respectively as follows:
{ } PX = {x i ∈ U [x i ]ind (p) ∩ X ≠ φ}
PX = x i ∈ U [x i ]ind (p ) ⊂ X B X = {X ∈ U | K B ( x) ⊆ X }
B X = {X ∈ U | K B ( x) ∩ X ≠ φ } where xi is an elementary set contained in X, 1,2,…,n. The ordered pair
i =
(PX, PX ) is called rough set. In
general, PX ⊆ X ⊆ PX. If PX = PX then X is called exact. The lower approximation of X is called the positive region of X and is denoted by POS(X) and the complement of upper approximation of X is called the negative region of X and is denoted by NEG(X). Its boundary is defined as BND(X) = PX − PX. Hence, it is trivial that if BND(X)= ϕ , then X is exact. In the theory of rough sets, the decision table of any information system is given by T=(U,A,C,D), where U is the universe of discourse, A is a set of primitive features, C and D are the subsets of A, called condition and decision features respectively.
Page 292
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
For any subset P of A, a binary relation IND(P), called the indiscernibility relation is defined as IND(P)={(x,y) ∈ U × U : a ( x) = a ( y ) for all a in P} The Rough set in A is the family of all subsets of U having the same lower and upper approximations. Usually rough set theory are explored using the decision tables i.e., information tables. Core and Reduct of Attributes: Core and reduct attribute sets are two fundamental concepts of a rough set. Reduct can minimize subset and make the object classification satisfy the full set of attributes. The core concept is commonly used in all reducts [11]. Reduct attributes can remove the superfluous attributes and give the decision maker a simple and easy information. There may be more than one reduct attributes. If the set of attributes is dependent, we are interested in finding all possible minimal subsets of attributes which have the same number of elementary sets. The reduct attribute set affects the process of decision making, and the core attribute is the most important attribute in decision-making. If the set of attributes is indispensable, the set is called the core, which is defined as
RED( P) ⊆ A COR(B) = ∩ RED(P) Decision Rules: The information system in RST is generally expressed in terms of a decision table. Data in the table are got through observation and measurement. Its rows represent objects, and its columns represent attributes divided into two sets, i.e, a condition attributes set and a decision attributes set. RST reduces the decision table, and then produces minimum decision rules sets. The general form of RST rules can be expressed by RULE : IF C THEN D where ‘C‘ represents conditional attributes and their values, ‘D’ represents a decision attributes and their values. Though the decision rules we can minimize the set of attributes, reduct the superfluous attributes and group elements into different groups. In this way we can have many decision rules, each rule has meaningful features. The stronger rule will cover more objects and the strength of each decision rule can be calculated in order to decide the appropriate rules.
Velammal College of Engineering & Technology, Madurai
3. CASE STUDY ON HRD OF PROFESSIONAL COLLEGES A questionnaire has been built based on the several analysis we provided, in an educational institution. In this research we obtained answers of 25 questions of 60 students from various disciplines. These questions are named attributes which characterize each student’s. When students have good command over English language it always helps an institution to get a placement and to earn more admissions in the next academic year. Nevertheless, it is difficult to find out whether the students have a good command over English language. This issue is an important one in every institution, but for the head of an institution/ department is hard to recognize who has good command over English and his interaction with institutions, his/her class mates and faculties. Therefore, we place a focus on the answer of the question: “Do you have a good command over English with your class mates/faculty? Let us denote an answer of student i to this question (i = 1,2,...,10) as Ai . The objective of this research is to clarify what kind of pattern of other answers to the questionnaire’s results in some value of Ai , that is, “Yes, I have a good command over English” , “no, I have not a good command over English”, “High, I have a good soft skills and Personality development, and Good financial constraints”, “Medium, I have a little bit knowledge of soft skills and financial constraints” and “Low, I have no soft skills and financial constraints”. We investigate all the answers of 10 selected students based on his/her class performance in various disciplines and find out the latent structure of the answers as head of the institution and institution can provide effective functions to motivate their students. In the analysis, we firstly processed the data that are obtained from questionnaires by ROSE [8,9]. Table 1 shows the lower and upper approximations obtained by a rough set analysis. This result has the accuracy of 1.000. This means the target set is definable on the basis of an attribute set [3,4]. Attribute Ai ( i = 1, 2, …, 10) “ Do you have a good command over English with your teachers and class mates?” is named as a decision attribute. The values that Ai (i = 1, 2, …, 10) takes are 1 (yes) and 2 (no). There are 7 YESs and 3 NOs. Figure 2: ROSE with the result of decision rule
Page 293
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Class number 1 2
Number of objects 3 7
Lower approx. 3 7
Upper approx. 3 7
accuracy 1.0000 1.0000
Table 2 : Lower and Upper Approximations Table 2 illustrates that the upper and lower approximations are equivalent. Therefore, there is no uncertainty in the classification of classes D =1 and D = 2. When decision rules are obtained, the decision rules can help a decision maker to get more information about human resource.
students’ behavior, in this way the head of an institution can differentiate features into many groups and each group of its own policies. Through the ROSE, we can obtain the results, which are shown in figure 2.
Core and Reducts In rough set analysis, the core contains the attributes that are the most important one in the information table. In this study all three condition attributes are included in the core. This indicates that any attribute that is necessary for perfect approximation of the decision classes and removal of any of them leads to the decrease of the quality of approximation. Reducts, also called minimal sets, contain no redundant information. In this study, 4 reducts, shown in Table 3, were obtained for the coded information table. The length of reducts is 2~3 and is smaller than that of the original set. They represent a reduced set of attributes that provide the same classification quality of the decision attribute as the original set of condition attributes. Table 3: Reducts for the coded information Table No. Reduct {English, Soft Skills} #1 {English, Finance} #2 {Finance, Soft skills} #3 {English, Finance, Soft Skills #4 Decision Rules induction We used LEM2 methodology in ROSE system to reduce the information table and got 7 rules. Rough set analysis provides computation intelligence to the problem of classification. Each rule has its own elements. Those elements are the features of these rules. We want to find out the typical rules that can cover most of the students and help decision makers to get the ideal picture of
Velammal College of Engineering & Technology, Madurai
There are 5 decision rules that have good placement with organizations and 2 decision rules that have bad placement with organizations. This decision table stands for students’ behaviors in the English Language communications, Financial stability of parent/Guardian, and Soft skills development in an institutions. More than half of the students in an institution, have better communication skills, soft skills and such features, when they have good placement in organizations. 4. RESULTS AND DISCUSSIONS From the rules 6 and 7 we can understand that when students have better placement opportunity in an institution in the IT corporations, they also have positive thinking and behaviors. They understand a whole system of an institution and contact with people of their teachers to learn more soft skills, personality development, and improve more communication skills. So, the head of an institution should pay attention on how to make them contribute continually and increase their motivation to make more placement opportunity for their students. In rules 1 and 2 the students with bad communication and soft skills development with their teachers, classmates and public. Adequate training is required for those students. Those who do not improve his/her skills, they never get a placement. They are advised to attend various lectures or seminars and carrier-up, and knowledge of various training etc. Those negative features mean that students do not care about their jobs and do not want to make achievements. The institution of such students must find out the reason and push them back to work. Those positive and negative decision rules let us
Page 294
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
know what kind of features and behaviors can bridge between institution and placement. Students have good relationship with staff and engage themselves to gain recognition and have activities that give the person a sense of contribution. As students try to become better, they work harder to find out their abilities. Those students always want to learn new things and know more information about their technology. Therefore, such students will play a central role in their placements. Students, who have a bad relationship with teachers, only care about themselves whether they will be employed or not. It means that, they think everything for themselves. Therefore, they do not want placement. They also do not think that they are students. Therefore, those students have no interest in doing anything out of their duty. The head of the institution has to assure placement of their students. In this way, an institution can have effective staff. The institution also has to take care of the students who have good relationship with teachers, help them do their work correctly, sometimes give them some challenges and rewards to increase their motivation. No matter whether the students have good or bad relationship with teachers, they are the human resource of an institution. So, the head of the institutions can give those two groups with different functions. But it is always not easy to divide people into groups, so the heads of the institutions can recognize their staff according to the result of rough set theory. In this research we give a clear function to make students differentiable. 5. CONCLUSION In summary, rough sets theory is a new method for analyzing, inducing, studying and discovering of data. It puts forward a method of student placement assistance based on rough set theory. We have collected data for one college corresponding to the attributes and formulated the problem of finding placements using ROSE system. This formulation can be extended to a large number of colleges. The rule analysis in ROSE system for placement of students is equivalent to the rule generation in rough set theory that was presented in[10]. REFERENCES [1]. Azibi, R., and Vanderpooten, D. “Construction of rule-based assignment models,” European Journal of Operational Research, Vol. 138, No. 2. Pages 274-293, 2002.
Velammal College of Engineering & Technology, Madurai
[2]. Cascio, W.C. Managing Human Resources: Productivity, Quality of Work Life, Profits, McGraw- Hill, New York, 1992. [3]. Pawlak, Z. “Rough sets,” International Journal of Computer and Information Science, Vol. 11, No. 5. Pages 341–356, 1982. [4]. Pawlak Z., Rough Sets, Kluwer Academic Publishers, 1991. [5]. Pawlak, Z, “Rough set and data analysis,” Proceedings of the Asian 11-14 Dec.. Pages 1 – 6, 1996. [6]. Pawlak, Z.. “Rough classification,” Int. J. Human-Computer Studies, Vol. 51, No. 15. Pages 369-383, 1999. [7]. Pawlak, Z. , “Decision networks,” In: Rough Sets and Current Trends in Computing by Shusaku Tsumoto, Roman Slowinski, Jan Komoroski, Jerzy W. Grzymala-Busse (Eds.), Lecture Notes in Artificial Intelligence (LNAI), Vol. 3066, No. 1. Pages 1–7, 2004. [8]. Predki, B., Slowinski, R., Stefanowski, J., Susmaga, R. and Wilk, Sz., “ROSE - software implementation of the rough set theory,” In: L.Polkowski, A.Skowron (eds.), Rough Sets and Current Trends in Computing, Lecture Notes in Artificial Intelligence, Vol. 1424, Springer-Verlag, Berlin. Pages 605-608, 1998. [9]. Predki, B., and Wilk, Sz., “Rough set based data exploration using ROSE system,” In: Z.W. Ras, A. Skowron (eds.). Foundations of Intelligent Systems, Lecture Notes in Artificial Intelligence, Vol. 1609. Springer-Verlag, Berlin. Pages 172-180, 1999. [10]. Ramasubramanian.P, Iyakutti.K, Thangavelu.P, “Mining Analysis of Computing Rule Confidence in Rough Set Theory”, Proceedings of National Conference on Networks, Image and Security, Organized by Noorul Islaam College of Engineering in association with The Institution of Engineering and Technology, YMS – Chennai Network, Pages .102 – 106, 2008. [11]. Shinya Imai et. Al: Rough set approach to Human Resource Development, vol. 9, No.2, May 2008. About the Authors P.Ramasubramanian, is Professor and Head in the Department of Computer Science and Engineering in Dr.G.U.Pope College of Engineering, Sawyerpuram, Tamilnadu, India. He obtained his Bachelor and Master degree in Computer Science and Engineering from M.K.University, Madurai in the year 1989 and 1996 respectively. He is doing Ph.D in Madurai Kamaraj
Page 295
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
University, Madurai. He has over 21 years of Teaching Experience and authored 15 books and 20 research papers in International, National Journals and Conferences. His current area of research includes Data Mining, Data Ware housing, Neural Networks and Fuzzy logic. He is a member of various societies like ISTE, International Association of Engineers, Computer Science Teachers Association, International association of Computer Science and Information Technology and Fellow in Institution of Engineers (India). Email: [email protected], [email protected]. V.Sureshkumar is Professor in the Department of Computer Science and Engineering in Velammal Engineering College, Chennai, Tamilnadu, India. He obtained his Bachelor degree in Elecronics and Communication Engg., from M.K.University and Master degree in Computer Science and Engineering from Manonmaniam Sundaranar University, Tirunelveli in the year 1991 and 2002 respectively. He is doing Ph.D in Madurai Kamaraj University, Madurai. He has over 18 years of Teaching Experience and published 12 research papers in International, National Journals and Conferences. His current area of research includes Fuzzy logic, Neural Networks and Data Mining. He is a life member of ISTE. Email: [email protected]
P.Alli received her Ph.D degree in Computer Science from Madurai Kamaraj University, Madurai, India. She obtained her B.E degree in Electronics and Communication Engineering from Madras University and M.S. in Software System from BITS, Pilani. She worked as Professor and Head, Department of Information Technology, Faculty of Engineering, Avinashilingam University, Coimbatore for 12 years. She is at present Professor and Head, Department of Computer Science and Engineering , Velammal College of Engineering and Technology, Madurai. Her research interests include statistical image processing, image segmentation, image enhancement, medical image analysis and novel image reconstruction algorithms. She is especially specializing in the development of adaptive filtering technique techniques for magnetic resonance imaging. Email: [email protected]
Velammal College of Engineering & Technology, Madurai
Page 296
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Robust Security metrics for the e-Healthcare Information Systems Said Jafari#1, Fredrick Mtenzi#2, Ronan Fitzpatrick#3, Brendan O’Shea#4 School of Computing, Kevin Street, Dublin Institute of Technology, Ireland. 1
[email protected] 2 [email protected] 3 [email protected] 4 [email protected] Abstract – The application of computer based and communication technologies in the healthcare domain has resulted in a proliferation of systems through healthcare. These systems whether in isolation or integrated, are generally referred to as e-healthcare information systems. Protection of healthcare information has become an increasingly complex issue with the widespread use of advanced technologies, complexity of healthcare domain and the need to improve healthcare services. The huge amount of information generated in the process of care, must be securely processed, stored and disseminated to ensure a balanced preservation of personal privacy versus its utility for overall improvement of healthcare. The general IS security protection approaches are available to address a wide range of security concerns. However, the healthcare domain has unique security concerns that have to be uniquely addressed. Understanding these specific security and privacy challenges shall enable selection, development, deployment and assessment of IS security approaches relevant for the healthcare domain. One of the primary tasks should be identifying the unique security concerns in the domain. This paper discusses among others the unique security and privacy concerns in the healthcare domain and proposes metrics that address the challenges presented. The main contribution of the paper is the demonstration of the development of the security metrics for the healthcare domain. Keywords – security metrics, security measurement, e-health security, security assessment, e-healthcare security.
VII.
I. INTRODUCTION
The application of computer based and communication technologies in the healthcare domain has resulted in a proliferation of systems through healthcare. These systems whether in isolation or integrated, are generally referred to as e-healthcare information systems. They range from clinical information systems such as laboratory, radiology, surgical, doctor consultation, critical care, to hospital administration routines such as stores management, facilities management, financial management, and human resource management. Among the key motivation that led to embrace IT in healthcare include the need to reduce medical errors, controlling escalating healthcare costs and providing
Velammal College of Engineering & Technology, Madurai
efficient healthcare services [1]-[3]. However, digital information has much higher security risks than its counterpart paper based. Once information is digitised, processing (copying, editing, deleting, searching for a particular data of interest, linking of data attributes to determine a particular meaningful pattern) becomes easier and timeless. Additionally, the physical presence of an invader which was once a significant factor is no longer a necessary condition for launching an attack. The magnitude of security risks in digital information rose faster. This is a general observation that cut across all elements of the Information Technology Functional Services Ecosystem (ITFSE) described in section II. Information systems that are 100% secure do not exist, at least to date. This is because “software” which is a life blood of any information system is ‘buggy’ [4]-[6]. The discussion on the topic of insecurity of software; reasons and consequences are well documented in the literature [7], [8], and are not the aim of this paper either. Also, there are human users in the security chain who are referred to as the weakest link in the chain [9], [10]. The general IS security approaches are available to address a wide range of security concerns. However, the healthcare domain offers unique security concerns that have to be addressed specifically. Understanding these specific security and privacy challenges shall enable development, selection, deployment and assessment of IS security approaches relevant for the healthcare domain. The “relevance” as used in this context refers to the “best possible options” that consider the security risks of the domain in order to provide reasonable, cost-effective and justifiable protection. One of the primary tasks should be identifying the unique security concerns in the domain. This paper discusses the unique security and privacy concerns in the healthcare domain and provides metrics that capture the challenges presented. The main contribution of the paper is to demonstrate the development of security metrics for the healthcare domain.
Page 297
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The rest of the paper is organised as follows: Section II discusses the holistic view of security as mapped from the proposed IT Functional Services Ecosystem. Section III presents a review of literature that addresses security and privacy issues in healthcare information systems. Section IV outlines and explains specific e-healthcare security and privacy challenges. Section V covers issues related to measurements and metrics. It includes examples of the metrics derived to address some of the presented challenges. Section VI provides conclusion and future work. VIII. IX. II. IT FUNCTIONAL SERVICES ECOSYSTEM (IT-FSE) AND SECURITY THINKING
Information Technology is the hub that connects nearly all human activities. This is illustrated in figure 1. The proliferation of automation and advancement in networking technologies has affected the way we look at implemented systems in different domains. Systems which were once considered different now have lots in common, particularly in the security context. This is because the same information infrastructures for data storage and data communication are used across them. In that way security problems that affect one entity in the IT-FSE can affect other entities as well. Based on this view, it is fair to have general security approaches that address security concerns as experienced by all entities. Tools were also developed in the same line of thinking. Firewalls, passwords, encryption algorithms, and even security protocols such as secure socket layer among others; all are designed to be application domain neutral. When thinking of technical controls, all available tools are generic. That is, as far as security is concerned, technically all systems should think alike.
In the contemporary security thinking, it is suggested that security is not merely a technical problem, so its problems cannot be solved by technical solutions alone [9], [11]. This understanding has broadened the security view to include personnel and organisational issues [9], [11]. The security field has progressed in terms of its scope and sophistication of security tools. The market is flooded with these tools for any organisation to use. However, despite this proliferation of tools, reports on security breaches increase [12], [13]. This trend suggests that a thoughtful effort should continuously be applied to address security issues particularly in application domains that involve sensitive information like healthcare domain. In the following section, a review of literature that addresses security and privacy issues is presented. The review shall lead to identification of healthcare unique security and privacy challenges. X. III. RELATED WORK Researches in healthcare security and privacy issues are considered and they are organised in four broad interrelated categories. These categories include policy models, design architectures for databases and communication networks, access control, and requirement elicitation with its modelling techniques. I. A. Policy Models In the first category, Anderson [14], [15] proposed a security policy model for clinical information systems. The model strongly considers consent propagation and enforcement to elevate privacy property. Threats and risks to the patient confidentiality, integrity and availability of the personal health information are also discussed. B. Database and Networks Architectures In the second category, Agrawal et al [16], [17] introduced the database architecture and prototype to address privacy problems for sensitive personal information. This technology is known as the Hippocratic Database, engineered in accordance to the ten designed principles proposed by author in the effort to address key issues in privacy management. Also, network architecture for healthcare are prototyped [3]. C. Access Control In the third category, Blobel [18] discussed specifications for privilege and access for healthcare. Gritzalis and Lambrinoudakis [19] propose a role-based access control security architecture to provide authentication and authorisation in web-based distributed system in the context of health care unit (HCU). In this work, security requirements for designing a secure architecture for interconnecting healthcare systems are discussed. Also, in Røstad’s PhD thesis [20] access control requirements are discussed and identify the dynamic nature of access control in healthcare as a result of the unpredictable, dynamic and individual care process.
Velammal College of Engineering & Technology, Madurai
Page 298
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
D. Requirements Engineering The fourth category is concerned with requirements elicitation. Blobel [21] presented a framework to facilitate security analysis and design for health information systems. Tools such as those described in unified modelling language (UML), object orientation and component architecture are defined to suite the healthcare environment. Also, the framework utilised generic security model and a layered security scheme to address the challenge imposed by security enhanced health information system. Cysneiros in [22], presents requirements elicitation techniques that are useful for compliance with the constraints imposed by several peculiarities intrinsic to this domain. The techniques discussed aim to assist in selecting a proper requirement elicitation approach. Matousek in [23] introduces the basic security and reliability considerations for developing distributed healthcare systems in healthcare. Based on the above review, there is large body of research specialised in healthcare domain in attempts to address various concerns. This indicates the domain has unique challenges. However, what is unique requirement other than wrapping on the “sensitivity of information” is not clearly featured in many literature reviewed. This paper includes literature [14]-[26] that in one way or another pointed and elaborated the nature of sensitivity and how to address the challenges. In the following section, a summary and discussion of these challenges are presented.
XI. IV. HEALTHCARE SECURITY AND PRIVACY CHALLENGES Healthcare information systems capture, store, process, and disseminate information across their components. This information is sensitive. Protection of this information and information systems is vital for successful healthcare delivery. Failure of protection mechanisms or inadequate
Velammal College of Engineering & Technology, Madurai
The healthcare domain in the information age, presents unique security and privacy challenges different from others [5]-[14]. These challenges as described in the following subsections, make protection of healthcare information systems a daunting task. Understanding them, provide a locus for enhancement of protection mechanisms as well as its assessment. Table 1 contains a summary of challenges and their associated security risks or threats organised into three main categories. The categories considered include information needs, nature of healthcare service delivery and laws, rules and regulations.
A. Information Needs In the first category, information need is considered. It is well known that the primary function of healthcare services is provision of care to patients and supporting public health. Patient care requires capturing of detailed information about the patient for identification, longitudinal and cross-section information for supporting evidence-based care delivery [24]. Addressing this demand becomes a significant challenge in the course of balancing protection versus accessibility of information. This could lead to information aggregation risks, a situation where more people have access to many records [14], [15]. B. Nature of Healthcare Service Delivery The second category is the nature of healthcare service delivery. The following are challenges and their associated risks for this category. The nature of work requires collaboration among multioccupations communities (e.g., physicians, nurses, technicians, and administrative staff) [25]. In some cases, duty of care is accomplished in teams (for example in operation theatres). The work is often non-routine, so it is difficult to pre-schedule events and activities [25].
Page 299
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Emergencies, exceptions and other contingencies in similar scenarios complicate standardisation of workflows and impose risks to accountability enforcement. Autonomy is another natural feature in healthcare environment. Service departments, clinicians and patients operate as autonomous entities with their own knowledge, ethics, beliefs and goals [26]. Harmonising conflicting security and privacy requirements in an autonomous culture is difficult and could be source of risks. Healthcare service utilises different technologies that generate heterogeneous data. Integrating these tools increase attack surface. Also, the sheer volume of people in healthcare environment imposes a threat to physical security [24]. C. Laws, Rules and Regulations The third category is concerned with laws, rules and regulations that healthcare organisations have to comply with. Among the sources of security requirements include governing laws, ethical rules and regulations applicable to the area of occupancy. Balancing the demand for multiple compliances is difficult and imposes significant risks to information protection controls. The discussed challenges and risks provide additional considerations during elicitation of security and privacy requirements and goals. As the focus of this paper is metrics development, the security and privacy requirements and goals are main inputs. Meaningful measurements and metrics for security are based on security goals [27]. For measurements to yield value, security requirements and goals must be properly formulated within the considerations of the unique challenges of the application domain. In the section V, security assessments and metrics for healthcare domain are considered. XII.
V. SECURITY ASSESSMENT AND METRICS
A. The Need for Metrics Healthcare stakeholders such as patients and clinicians need to have confidence in the systems they are interacting with in the duty of care. These systems have to be justifiably secure for users to have trust in them [28]. Security metrics provide insights to efforts engaged to accomplish security goals and illuminate anomalies or weaknesses in the security controls implemented [29]. They offer evidence of the security level to justify security claims and demonstrate compliances when required. The overall helpfulness of metrics covers a wide range of security and privacy decision making; particularly in the balance between security, privacy and access to information in supporting clinical operations, research, and public health, as the main concerns of healthcare domain. B. Security Metrics Clear and precise articulation of security and privacy challenges in the healthcare domain forms the basis for
Velammal College of Engineering & Technology, Madurai
understanding the security and privacy requirements. Requirements portray security goals. Good metrics are goal driven [27], [30]. Based on the unique challenges presented in table 1, security metrics for healthcare should also incorporate the unique peculiarities of the application domain. Table 2 provide examples of the metrics and criteria to be considered for measurements based on the observation made on table 1. The terms metric, base metric, derived metric and threshold as used in this paper are borrowed from the international measurement standard (ISO/IEC 27004) [31]. However, this new standard does not use the term ‘metric’, it uses ‘measure’ instead. We have taken the liberty to synonymise the terms based on our experience on usage of these terms in previous and similar standards [32], [33]. Therefore, the terms are defined as follows: 1. Metric – variable to which a value is assigned as the result of measurement. 2. Base metric – measure defined in terms of an attribute and the method for quantifying it. 3. Derived metric – measure that is defined as a function of two or more values of base measures. 4. Threshold – target, or patterns used to determine the need for action or further investigation, or to describe the level of confidence in a given result. The sample metrics and criteria presented in table 2, consider some of the challenges presented in table 1 such as those leading to information aggregation concerns and accountability. For details on how the challenges in table 1 are related to metrics in table 2, follow the superscripts cross-reference on table 1, 2 and 3. The metrics provided are those which can be captured automatically after accomplishment of care. The reasons for automating these metrics include eliminating human biases in data gathering and computations, ensuring reproducible experiments and comparison of results. These metrics (in table 2) calculate the “Information loss Factor”, a derived metric named InfoLossFactor (μ) designed to monitor suspicious use of patient information as used in the duty of care. Also, it can be used to provide evidence against claims of illegal information disclosure, or indicating weakness in security controls deployed for protecting information. To derive this metrics, the following assumptions are considered: 1. All transactions are captured in the audit trails and the integrity of the audit trails has no doubt. The system under monitoring considered is a clinical information system, in hospital settings where the number of patients attending for care is reasonably higher than clinicians. 2. Patients’ records are accessed in the duty of care involving patients whose records are accessed. This assumption may seem unrealistic; however it serves to demonstrate our proposed metrics. 3. Database maintenance functions such as update, backup are not included in the metrics proposed.
Page 300
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
To enable derivation of the formulas, terms were proposed by authors. Table 3 provides brief interpretation of terms used in table 2, source of data and technique for data extraction. The values of the thresholds provide reasonable computed value if no suspicious indication is spotted. For instance, InfoLossFactor (μ) (in this case) its threshold is 7, to represent a secure information accesses. However the computed value of μ can increase if individual thresholds of base metrics are exceeded. The increase indicates a suspicious information access may have occurred. A different explanation can be considered on the computed value of γ1 in equation (7); in a reasonable situation, its value should approach zero, indicating fewer service departments are involved in exchanging patients’ information. In the computation of μ, the value of γ1 is taken as zero (its magnitude has no significant contribution in the in the overall equation). While not addressed in this paper, setting minimum values for thresholds is of paramount. When computed values go below stated values, different insight can be spotted; either there are significant changes in work flows or the assumptions made on the formulas to compute metrics values are wrong. Also, it should be understood that not all of the challenges highlighted in table 1 can be properly addressed by technical metrics. The effort should be made to automate as many as possible. C. Metrics Evaluation Strategy Evaluation of metrics is important to any proposed metrics. The evaluation strategy for our metrics is based on the use of principles outlined. We borrowed some of these principles from Heinrich et al [34] and Reiter et al [35]. However, these principles were designed for evaluation of DataQuality metrics and authentication metrics respectively; they fit for our metrics too. A compiled list of principles is as follows: • Principle 1: Meaningful output – the metrics should strive to measure what they are intended to measure [34], [35]. • Principle 2: Feasibility – the metrics should exhibit practicality based on determinable inputs and ease of application [34], [35]. • Principle 3: Aggregation – individual values of metrics can be summed up to few values meaningfully [34]. • Principle 4: Precision – the extent that repeatable concise results can be demonstrated for several measurements taken under similar condition [36], [37]. • Principle 5: Cost effective – metrics data must be inexpensive to gather in terms of time and cost, preferably gathered automatically [29], [37]. Demonstration of these principles against our proposed metrics is beyond the scope of this paper. However, it is
Velammal College of Engineering & Technology, Madurai
evident that our metrics do comply to some of these principles straight away. As security controls, standards, laws and regulations for healthcare are developed specifically to address the challenges of this domain, the same attention should be engaged in developing assessment approaches and metrics to measure the degree of protection offered by the proposed approaches. XIII. VI. CONCLUSION Protection of healthcare information has become an increasingly complex issue with the widespread use of ICT, complexity of healthcare domain and the need to improve healthcare services. The huge amount of information generated in the process of care, must be securely stored, processed and disseminated to ensure a balanced preservation of personal privacy versus its utility for overall improvement of healthcare services. The general IS security approaches are available to address a wide range of security concerns. However, the healthcare domain offers unique security concerns that have to be addressed specifically. In this paper, we have discussed the unique challenges imposed on security and privacy of information as enhanced by an array of factors in the healthcare domain and proposed sample of the security metrics to assess the degree to which those challenges can be controlled. Understanding these specific security and privacy challenges shall enable selection, development, deployment and assessment of IS security policies, tools and approaches relevant for the healthcare domain. As part of future work, on the basis of the presented security and privacy challenges, the development of metrics and criteria to be used in the assessment of the information protection in healthcare domain will be extended. XIV. REFERENCES [1] A. Mukherjee and J. McGinnis, "E-healthcare: an analysis of key themes in research," International Journal of Pharmaceutical and Healthcare Marketing, vol. 1, pp. 349-363, 2007. [2] V. Mantzana and M. Themistocleous, "Identifying and Classifying Benefits of Integrated Healthcare Systems Using an Actor Oriented Approach," Journal of Computing and Information Technology, vol. 2, pp. 265--278, 2004. [3] M. Beyer, K. A. Kuhn, C. Meiler, S. Jablonski, and R. Lenz, "Towards a flexible, process-oriented IT architecture for an integrated healthcare network," 2004. [4] S. M. Bellovin, "On the Brittleness of Software and the Infeasibility of Security Metrics," IEEE Security \& Privacy, vol. 4, pp. 96--96, 2006. [5] J. Viega and G. McGraw, Building secure software: Addison-Wesley Boston, 2002. [6] E. Voas, F. Charron, G. McGraw, K. Miller, and M. Friedman, "Predicting how badly “good” software can behave," IEEE Software, vol. 14, pp. 73-83, 1997.
Page 301
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[7] A. C. Michael, "Who is liable for bugs and security flaws in software?" Commun. ACM, vol. 47, pp. 25-27, 2004. [8] D. Rice, Geekonomics: the real cost of insecure software: Addison-Wesley Professional, 2008. [9] M. N. Wybourne, M. F. Austin, and C. C. Palmer, "National cybersecurity research and development challenges related to economics, physical infrastructure and human behavior," Institute for Information Infrastructure Protection (I3P) 2009. [10] B. Schneier, "Secrets |& Lies," C. Long, Ed.: Wiley Publishing, Inc., 2004, pp. 255-269. [11] B. von Solms, "Information security—a multidimensional discipline," Computers & Security, vol. 20, pp. 504-508, 2001. [12] PwC, "Information security breaches survey 2006," PricewaterhouseCoopers 2006. [13] Berr, "2008 INFORMATION SECURITY BREACHES SURVEY," Department of Business Enterprises \& Regulatory Reform 2008. [14] R. J. Anderson, "A security policy model for clinical information systems," presented at IEEE Symposium on Security and Privacy, Oakland, CA, 1996. [15] R. J. Anderson, "Patient confidentiality-at risk from NHS-wide networking," Current perspectives in healthcare computing, pp. 687-692, 1996. [16] R. Agrawal, A. Kini, K. LeFevre, A. Wang, Y. Xu, and D. Zhou, "Managing healthcare data hippocratically," presented at Intl. Conf. on Management of Data, 2004. [17] R. Agrawal and C. Johnson, "Securing electronic health records without impeding the flow of information," International journal of medical informatics, vol. 76, pp. 471-479, 2007. [18] B. Blobel, "Authorisation and access control for electronic health record systems," International Journal of Medical Informatics, vol. 73, pp. 251-257, 2004. [19] D. Gritzalis and C. Lambrinoudakis, "A security architecture for interconnecting health information systems," International Journal of Medical Informatics, vol. 73, pp. 305-309, 2004. [20] L. Røstad, "Access Control in Healthcare Information Systems," in Computer and Information Science, vol. PhD. Trondheim: Norwegian University of Science and Technology, 2008, pp. 161. [21] B. Blobel, "Security requirements and solutions in distributed electronic health records," Computers and Security, vol. 16, pp. 208-209, 1997. [22] L. M. Cysneiros, "Requirements engineering in the health care domain," presented at Requirements Eng. (RE 02), 2002. [23] K. Matousek, "Security and reliability considerations for distributed healthcare systems," presented at 42nd ANNUAL 2008 IEEE INTERNATIONAL CARNAHAN CONFERENCE ON SECURITY
Velammal College of Engineering & Technology, Madurai
TECHNOLOGY (ICCST 2008), PRAGUE, CZECH REPUBLIC, 2008. [24] ISO/IEC-27799, "Health Informatics-Information security management in health using ISO/IEC 27002," International Organisation for Standardisation 2008. [25] Y. Xiao, "Artifacts and collaborative work in healthcare: methodological, theoretical, and technological implications of the tangible," Journal of biomedical Informatics, vol. 38, pp. 26-33, 2005. [26] A. C. Elliott, "Health care ethics: cultural relativity of autonomy," Journal of Transcultural Nursing, vol. 12, pp. 326-330, 2001. [27] V. R. Basili, G. Caldiera, and H. D. Rombach, "The goal question metric approach," Encyclopedia of software engineering, vol. 1, pp. 528-532, 1994. [28] J. Grimson, W. Grimson, and W. Hasselbring, "The SI challenge in health care," Communications of the ACM, vol. 43, pp. 48--55, 2000. [29] A. Jaquith, Security Metrics: Replacing Fear, Uncertainty, and Doubt: Addison-Wesley Professional, 2007. [30] J. S. Taylor, "Autonomy and informed consent: A much misunderstood relationship," The Journal of Value Inquiry, vol. 38, pp. 383-391, 2004. [31] BS-ISO/IEC-27004, "Information technology — Security techniques— Information security management measurements," International Organisation for Standardisation (ISO) 2009. [32] M. Swanson, N. Bartol, J. Sabato, J. Hash, and L. Graffo, Security metrics guide for information technology systems: National Institute of Standards and Technology, 2003. [33] E. Chew, M. Swanson, K. Stine, N. Bartol, A. Brown, and W. Robinson, "Performance Measurement Guide for Information Security (NIST Special Publication 800-55 Revision 1)," National Institute of Standards and Technology 2008. [34] B. Heinrich, M. Klier, and M. Kaiser, "A Procedure to Develop Metrics for Currency and its Application in CRM," J. Data and Information Quality, vol. 1, pp. 1-28, 2009. [35] M. K. Reiter and S. G. Stubblebine, "Toward acceptable metrics of authentication," presented at Proc. IEEE Symposium on Security and Privacy, 1997. [36] D. S. Herrmann and S. Herrmann, Complete guide to security and privacy metrics: Measuring regulatory compliance, operational resilience and ROI: CRC Press, 2007. [37] D. Rathbun and L. Homsher, "Gathering Security Metrics and Reaping the Rewards," SANs Intitute 2009.
Page 302
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Face Detection Using Wavelet Transform And Rbf Neural Network M.Madhu1, M.Moorthi2, S.Sathish Kumar3, Dr.R.Amutha4 1, 2, 3 Research Scholar, Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram, India. 4. Professor and Head of the Department Sri Venkateswara college of Engineering, Chennai, India. [email protected], [email protected]
Abstract— Face Recognition is an emerging field of research with many challenges such as physical appearance where faces does not share the same physical features, face acquisition geometry that is, how the images are obtained and what environment the face imaged, illumination, pose and occlusion and so on. The main objective of this paper is to recognize the given face image from the database using wavelets and neural networks. DWT is employed to extract the input features to build a face recognition system and neural network is used to identify the faces. In this paper, discrete wavelet transforms are used to reduce image information redundancy because only a subset of the transform coefficients are necessary to preserve the most important facial features such as hair outline, eyes and mouth. We demonstrate experimentally that when DWT coefficients are fed into a feed forward neural network for classification, a high recognition rate can be achieved by using a very small proportion of transform coefficients. This makes DWT-based face recognition much faster than other approaches. Key words: Face recognition, neural networks, feature extraction, discrete wavelet transform
Introduction Face recognition approaches on still images can be broadly grouped into geometric and template matching techniques. In the first case, geometric characteristics of faces to be matched, such as distances between different facial features, are compared. This technique provides limited results although it has been used extensively in the past. In the second case, face images represented as a two dimensional array of pixel intensity values are compared with a single or several templates representing the whole face. More successful template matching approaches use Principal Components Analysis (PCA) or Linear Discriminant Analysis (LDA) to perform dimensionality reduction achieving good performance at a reasonable computational complexity/time. Other template matching methods use neural network classification and deformable templates, such as Elastic Graph Matching (EGM). Recently, a set of approaches that use different techniques to correct perspective distortion are being proposed. These
Velammal College of Engineering & Technology, Madurai
techniques are sometimes referred to as view-tolerant. For a complete review on the topic of face recognition the reader is referred to [1] and [2]. Neural networks have been widely applied in pattern recognition for the reason that neural-networks-based classifiers can incorporate both statistical and structural information and achieve better performance than the simple minimum distance classifiers [2]. Multilayered networks (MLNs), usually employing the backpropagation (BP) algorithm, are widely used in face recognition [3]. Recently, RBF neural networks have been applied in many engineering and scientific applications including face recognition [4][5]. The paper is organized as follows. The DWT is presented in Section 2.1. Section 2.2 presents the NN classifier for face recognition. The simulation results are given in Section 3, and finally the conclusion is drawn in Section 4.
Proposed method
Wavelet Transform In this paper, we proposed to modify the DCT approach with the wavelets. The DCT has several drawbacks. Computation of the DCT takes an extremely long time and grows exponentially with signal size. In this case, we used the Haar wavelets, representing the simplest wavelet basis. We employed the non standard decomposition, which alternates between row and column processing, allowing a more efficient coefficients’ computation. The proposed algorithm uses these coefficients as inputs for the Neural network to be trained. But the DWT has a distinct advantage; The DWT, in essence, can be computed by performing a set of digital filters which can be done quickly. This allows us to apply the DWT on entire signals without taking a significant performance hit. By analyzing the entire signal the DWT captures more information than the DCT and can produce better results. The image is separated into four sub images. The bottom left, bottom-right and top-right show the high-frequency detail of the image. The top left quadrant contains the low frequency or lower detail portion of the image; we can see that most of the information is in this portion.
Page 303
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Gaussian activation function. The response of such function is non negative for all values of x. The function is defined as, f(x) = exp(-x2) The training algorithm
Figure 1.DWT sub band structure
Thus, the DWT separates the image’s high frequency components from the rest of the image, resizes the remaining parts and rearranges them to form a new ‘decomposed’ image. The coefficients of the decomposed image are given to the neural network as input for training and recognition.
Neural Networks A Neural Network is used in order to significantly reduce the computation time, false recognition rate of the system. The architecture of RBF Network consist of 3 layers they are input, hidden and output layers as shown in Fig 2.2 1
1
1
2
2
2
3
Input layer
45
Hidden layer
3
Step 1: Initialize the weights Step 2: While stopping is false do step 3 to step 10. Step 3: For each input do step 4 to step 9 Step 4: Each input unit xi= i=1----n. Receives input signals to all units in the layer above (hidden unit) Step5: Calculate the radial basis function Step 6: Choose the center for the RBF, the centers of chosen from the set of input vectors. A sufficient number of centers have to be selected in order to ensure adequate sampling of the input vector space. Step 7: The output of the unit vi(xi) in the hidden layer. vi(xi) = exp (xji ^ -xji) 2 / σi2 Where xji^ = center of the RBF unit for input variables σi2 = width of the ith RBF unit Xji = j th variable of input pattern Step 8: Initialize the weights in the output layers of network to some random value. Step 9: Calculate the output of the network Ynet= wim vi (xi)+ wo Where H= number of hidden layer Ynet= output value of mth node in output layer for the nth incoming pattern. Wim = weight between ith RBF unit and mth output node Wo = biasing term at mth output node Step 10: Calculate error and test the stopping condition. The following flow diagram shows the new approach for face detection using wavelet decomposition and Neural networks,where DWT is used for the extraction of features from facial images and recognition is done using neural network.
Output layer LOAD DATABASE
Figure 2.Architecture of RBF Neural Network
It is a multilayer feed forward network. There exists n number of input neurons and m number of output neurons with hidden layers existing between the input and output layers. The interconnection between the input and hidden layers forms hypothetical connection and between the hidden layer and output layers forms weighted connections. The training algorithm is used for the update of weights in all the interconnection.
LOAD QUERY IMAGE
WAVELET DECOMPOSITIO FEATURE TRAIN THE FACE DETECTED
The important aspect of the RBFN is the usages of activation function for computing the output. RBF uses
Velammal College of Engineering & Technology, Madurai
Page 304
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Figure 3. Flow Diagram
Conclusion
Thus for both the database and query images Haar wavelet is applied for decomposition and the wavelet coefficients are taken for calculating the features and these are taken as input to the NN for training and recognization. Finally, the recognized image is obtained.
Results The following shows the input image loaded for detection
In this paper, a new approach for face detection has been proposed by using wavelet decomposition and Neural networks,where DWT is used for the extraction of features from facial images and recognition experiments were carried out by using neural network. The ORL database images are used for conducting all the experiments. Facial features are first extracted by the DWT which greatly reduces dimensionality of the original face image as well as maintains the main facial features. Compared with the well-known DCT approach, the DWT has the advantages of data independency and fast computational speed. The presented Neural Network model for face recognition with a focus on improving the face recognition performance by reducing the input features. Simulation results on ORL face database also show that the proposed method achieves high training and recognition speed. References [1] Zhao, W.,
Figure 4. Input Image
The following shows the wavelet decomposed image of the input image
Figure 5. Wavelet decomposed face image
The following shows the simulation results of the detected face
Figure 6 Simulation result of the detected face
Velammal College of Engineering & Technology, Madurai
Chellappa R., Rosenfeld A. and Phillips P.J.: Face Recognition: A literature survey. Technical Report CART-TR-948. University of Maryland, Aug. 2002. [2] Chellappa R., Wilson, C.L., and Sirohey, S.: Human and machine recognition of faces: A survey, Proc. IEEE, 83(5): 705C740, 1995. [3] Valentin D., Abdi H., Toole, A. J. O., and Cottrell, G. W.: Connectionist models of face processing: A survey, Pattern Recognit., 27: 1209C1230, 1994. [4] Er, M. J., Wu, S., Lu, J., and Toh, H. L.: Face recognition with radial basis function (RBF) neural networks, IEEE Trans. Neural Netw., 13(3): 697C710, 2002D [5] Yang F. and Paindavoine M.: Implementation of an RBF neural network on embedded systems: Real-time face tracking and identity verification, IEEE Trans. Neural Netw., 14(5): 1162C1175, 2003. [6] Kwak., K.-C., and Pedrycz., W.: Face Recognition Using Fuzzy Integral and Wavelet Decomposition Method, IEEE Trans. on SMC-B, 34(4): 1667-1675, 2004.
Mr.M.Madhu is pursuing his Ph.D program at Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram. He completed his B.E degree in Electronics and Communication Engineering in the year 2001 and M.E Applied Electronics in the year 2006, at Arulmigu Meenakshi Amman College of Engineering, Kanchipuram, Chennai, India. He has 7 years of teaching experience and he is currently working as Assistant Professor in the department of Electronics and Communication Engineering at Rajiv Gandhi College of Engineering, Sriperumbudur, Chennai. He has published and presented many papers in National and International Conference in the area of Image processing. His research interests are Image Segmentation, Image Compression and Image detection, speech processing.
Page 305
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Mr.M.Moorthi is pursuing his Ph.D program at Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram. He completed his B.E degree at Arulmigu Meenakshi Amman College of Engineering, Kanchipuram, in Electronics and Communication Engineering in the year 2001 and M.E Medical Electronics in the year 2007 at Anna University, Guindy campus, Chennai, India. He has 9 years of teaching experience and he is currently working as Associate Professor in the department of Electronics and Communication Engineering at Prathyusha Institute of Technology and management, Chennai. He has published and presented many papers in National and International Conference in the area of Image processing. His research interests are Image Segmentation, Image Compression and Image detection.
Velammal College of Engineering & Technology, Madurai
Mr.S.Sathish Kumar is pursuing his Ph.D program at Sri Chandrasekhardendra Saraswathi Viswa Mahavidyalaya University, Kanchipuram. He completed his B.E degree in Electronics and Communication Engineering in the year 2001 and M.E Applied Electronics in the year 2006 , at Arulmigu Meenakshi Amman College of Engineering, Kanchipuram, Chennai, India. He has 9 years of teaching experience and he is currently working as Associate Professor in the department of Electronics and Communication Engineering at Prathyusha Institute of Technology and management, Chennai. He has published and presented many papers in National and International Conference in the area of Image processing. His research interests are Image Segmentation, Image Compression and Image detection, speech processing.
Page 306
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
An Clustering approach based on Functionality of Genes for Microarray data to find meaningful associations Selvanayaki M#1, Bhuvaneshwari V#2 #
School of Computer Science and Engineering, Bharathiar University Coimbatore-641 046, Tamil Nadu, India 1
[email protected] [email protected] Abstract— Bioinformatics is the science of storing, extracting, Molecular Function. A functionally meaningful cluster contains many genes that are annotated to a specific GO organizing, analyzing, interpreting, and utilizing information terms. The paper aims to cluster microarray data of yeast from biological sequences and molecules. Microarrays consist based on the functionalities of genes based on GO ontology to of large numbers of molecules (often, but not always, DNA) find meaningful associations of genes from the dataset. distributed in rows in a very small space. Clustering methods provide a useful technique for exploratory analysis of microarray data since they group gene with similar functionalities based on GO Ontology. In this paper, Gene Keywords— Bioinformatics, Data Mining, Microarray, GO Ontology is used to provide external validation for the clusters to determine if the genes in a cluster belong to a Ontology, Clustering. specific Biological Process, Cellular Component and 2
I. INTRODUCTION Data mining is the non-trivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data [3]. Data mining is a larger process known as knowledge discovery in databases (KDD). Data mining is the process of discovering meaningful, new correlation patterns and trends by shifting through large amount of data store in repositories, using patterns recognition techniques as well as statistical and mathematical techniques. Bioinformatics is the application of computer technology to the management of biological information. Computers are used to gather, store, analyze and integrate biological and genetic information which can be applied to gene-based drug discovery and development. The need for Bioinformatics capabilities has been precipitated by the explosion of publicly available genomic information resulting from the Human Genome Project [4].
numbers of genes interact with each other and how a cell's regulatory networks control vast batteries of genes simultaneously. It is also called a DNA chip or a gene chip. It is the technology used to obtain a genomic profile [15]. GO (Gene Ontology) is a controlled vocabulary used to describe the biology of a gene product in any organism. There are three independent sets of vocabularies or ontologies that describe the molecular function of a gene product, the biological process in which the gene product participates, and the cellular component where the gene product can be found. The ontology is represented as a network, a directed acyclic graph (DAG), in which terms may have multiple parents and multiple relationships to their parents. In addition, each term inherits all the relationships of its parent(s).
Bioinformatics and data mining provide exciting and challenging researches in several application areas especially in computer science. Bioinformatics is the science of managing, mining and interpreting information from biological sequences and structures [6]. Data are collected from genome analysis, protein analysis, microarray data and probes of gene function by genetic methods. Microarray is a new way of studying how large
Various data mining techniques can be used in microarray clustering analysis such as hierarchical clustering, classification, k-means algorithms, genetic algorithm, association rules and self organization map [20]. The objective of the paper is to use the data mining technique, k-means clustering for microarray data to group genes with similar functionalities based on GO Ontology. The proposed idea is to compare clustering of genes based on original microarray data and clusters based on functionality of genes derived using GO ontology measures.
Velammal College of Engineering & Technology, Madurai
Page 307
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The paper is organized as follows. Section 2 provides the literature study of the various clustering algorithms using microarray. Section 3 explores the k- means algorithm methodology for clustering. In section 4 the implemented results for the yeast dataset are analyzed and validated within the go ontology functionalities. The final section draws the conclusion of the paper. II. REVIEW OF LITERATURE This section provides the details of the literature study from various papers related to microarray cluster analysis methods and techniques. Microarray technology has become one of the indispensable tools that many biologists use to monitor genome wide expression levels of genes in a given organism [7]. The data from microarray experiments is usually in the form of large matrices of expression levels of genes (row) under different experimental conditions (columns). Different methods have been proposed for the analysis of gene expression data including hierarchical clustering, self-organizing maps, and k-means approaches. In [5], Cheremushkin E.S. et al. describe partitioned clustering techniques used to find a close optimal solution. This work was aimed at the developed and implementation of a clustering algorithm based on fuzzy c-means in association with genetic algorithm. In [2], Gruzdz A. et al. proposed association rule is used to discover mutations in colon cancer diseases. In [10], Sukjoon Yoon et al. identification of the change of gene expression in multifactorial diseases, such as breast cancer is a major goal of DNA microarray experiments. In [15], Bolshakova et al. cluster validity methods to estimate the number of clusters in cancer tumor datasets. This estimation approach support biological and biomedical knowledge discovery. In [12], John Sikorski et al. proposed microarray technology to determine the relationships between genes, those that are differentially or coordinately expressed under specific conditions. In [13], LI Guo-qi et al. had developed classifier method to predict unknown function of genes in the microarray dataset on different levels. In [15], Robert S.H. et al. describes a new clustering method based on Linear Predictive Coding to provide enhanced microarray data analysis. In this approach, spectral analysis of microarray data is performed to classify samples according to their distortion values. The technique was validated for a standard data set. In [8], Eisen et al. proposed hierarchical clustering is probably the most extensively used microarray data mining techniques, using one of several techniques to iteratively, starting with one gene, combine genes with their nearest neighbor, gradually building clusters and associations of clusters, resulting in a hierarchical tree. Distance between clusters is defined by the distance between their average
Velammal College of Engineering & Technology, Madurai
expression patterns. In [20], Ka Yee Yeung et al. cluster analysis was used to identify genes that show similar expression patterns over a wide range of experimental conditions in yeast. Such genes are typically involved in related functions and are frequently co-regulated. In [19], Taraskina et al. presented a clustering method used to indentifying biologically relevant groups of genes. This paper developed an algorithm of fuzzy c-means family, designed for clustering of microarray data and distance matrices with genetic algorithm as optimization. Genetic algorithm use to find a close optimal solution to the problem of clustering. Euclidian distance method is used to calculate distance between genes. In [18], Sterhan Symons et al. proposed machine learning techniques are applied to microarray data, for diagnostic purposes. Especially in cancer diagnostics microarray classification tools are used cancer subtype discrimination and outcome prediction. In [9], Gao Cong et al. has developed two new efficient algorithm, RERII and REPT, to explore the row enumeration space to discover frequent closed patterns. In [14], Manoranjan Dash et al. proposed novel feature selection method used for large number of genes. Feature selection is a preprocessing technique that finds an optimal subset of informative features from the original set. In [1], Ben-Dor et al. discussed approaches to identify patterns in expression data that distinguish two subclasses of tissues on the basis of a supporting set of genes that results in high classification accuracy. III. METHODOLOGY Gene ontology is a collection of controlled vocabularies that describes the biology of a gene product. It consists of approximately 20,000 terms arranged in three independent ontology: Biological Process, Cellular Component, and Molecular Function [17]. Gene ontology is used to provide external validation for the clusters to determine if the genes in a cluster belong to a specific Cellular Component, Molecular Function or Biological Process. A functionally meaningful cluster contains many genes that are annotated to a specific GO terms. The main idea of the paper is clustering microarray data based on functionalities. The following information is downloaded from the website. The Yeast Saccharomyces Cerevisiae Dataset is downloaded from National Center for Biotechnology Information Gene Expression Omnibus (NCBI GEO) website. The Yeast microarray data contains about 6400 Genes (e.g. YALO51W, YALO54C) with their corresponding Yeast values (0.1650, 0.2720). The file “yeastgenes.sgd” and GO ontology was obtained from the GO annotation website.
Page 308
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The proposed methodology as given in Fig. 1 is used for analyzing microarray data to infer association of genes based on the clustered results using k means algorithm. The framework consists of three phases. In the first phase is preprocessing of microarray data is done, in the second phase the genes are mapped based on functionality using GO ontology. The clustering is done in the third phase using k means algorithm for the proposed approaches on the original dataset and the dataset based on the functionality of genes. The clustering results are analyzed for finding the association between genes in the microarray dataset.
TABLE I. EXPRESSED GENES (3677)
The yeast microarray data gene is mapped with SGDgenes structure to find the genes that are expressed in microarray yeast data to map with the SGDgo. After mapping, 3677 genes were found to be expressed in the yeast dataset. B) Mapping genes based on functionality SGDann is a master structure of Yeast microarray data is shown in Fig 2. It contains the parameters namely SGDaspect, SGDgenes, SGDgo. The entire microarray data are mapped with this structure (SGDgenes) and the microarray data is splitted based on the functionality of gene using GO ontology. For each gene the corresponding functionality of genes are found (Biological Process, Cellular Component, and Molecular Function). All the genes with same functionality are grouped for the corresponding Yeast values from the microarray dataset. The three dataset are constructed with 1079 records for biological process, 1445 records for cellular component and 1153 records for molecular function from the original microarray data are shown in Fig 3, Fig 4 and Fig 5.
Fig 1. Framework to infer association of genes
A) Preprocessing microarray data The preprocessing phase consists of two main processes that are removal of noisy data and mapping of microarray data and SGD go. 1) Removal of noisy data The yeast microarray data contains 6400 genes, with null values and empty genes. In the preprocessing step the empty spots and null values are removed using the knnimpute method. After removal of empty genes from the dataset then there are about 6314 genes used for further analysis.
Fig 2. SGDann Structure
2) Mapping of microarray data and SGD go SGDann is the master structure of the yeast genes containing all the information about the yeast data with 37794 yeast related genes. The SGDgenes column is alone extracted from the structure to map with the microarray data proposed in the work. The snapshot of microarray genes with SGDgenes is shown in Table 1.
Velammal College of Engineering & Technology, Madurai
Page 309
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig 3. Biological process (1079)
C) Clustering using k-means algorithm 1) Vector Space Model Vector space model (or term vector model) is a standard algebraic model for representing documents (and any objects, in general) as vectors of identifiers, such as, for example, index terms. It is used in information filtering, information retrieval, indexing and relevancy rankings. Its first use was in the SMART information retrieval system. A document is represented as a vector. Each dimension corresponds to a separate term [10]. The definition of term depends on the application. Typically terms are single words, keywords, or longer phrases. Several different ways of computing these values, also known as (term) weights, have been developed. One of the best known schemes is tfidf weighting (tf-term frequency, idf-inverse document frequency). The vector space model similarity measure is computed as the Euclidean distance from the query to each document [11]. The main idea of subspace (vector space model) clustering is to find a subset of genes and a subset of conditions under which these genes exhibit a similar trend [16]. 2) Input matrix The microarray data after preprocessing has 3677 expressed genes with yeast values. The expressed genes values are taken as input for vector space model. The snapshot of input matrix is shown in Table 2. TABLE II. INPUT MATRIX
Fig 4. Cellular Component (1445)
The above matrix is used to find the similarity distance using k-means algorithm. The distance is calculated using the formula given below
Fig 5. Molecular Function (1153)
The k-means algorithm is implemented in third phase on these datasets to determine the association between genes.
Velammal College of Engineering & Technology, Madurai
In this equation distance examines the root of square difference between coordinates of a pair of objects x and y. 3) Distance matrix
Page 310
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The corresponding distance matrix generated for the given input matrix is shown in Table 3. TABLE III. DISTANCE MATRIX
The distance matrix is given as input for the k-means algorithm. The basic k-means algorithm is defined below. 4) K-means algorithm Pseudo code
Fig 6. Yeast Microarray Dataset
The entire microarray dataset is mapping based on the functionality. After mapping to get 1079 for biological process, 1445 for cellular component and 1153 for molecular function are shown in Fig 7.
Fig 7. Mapping gene based on functionality
VI. EXPERIMENTAL RESULTS A) Results The implemented results for the yeast dataset are analyzed and validated within the go ontology functionalities to find association between genes. The original microarray data has only genes. These genes are mapped with SGDgenes to get the expressed genes for the original microarray dataset. The number of genes in the dataset is shown in the Table 4. TABLE VI. DATASET
Dataset
Number of genes
SGDgenes (master structure) Yeast microarray data After removal of empty genes Expressed genes
37794 6400 6413 3677
Biological Process
1079
Cellular Component
1445
Molecular Function
1153
B) Analysis of K-Means Algorithm 1) Clustering on original microarray data The 20 cluster were generated on the microarray dataset and after preprocessing of microarray dataset using kmeans algorithm. The snapshot of the gene generated in the two different dataset is shown in Table 5. The count of genes was different in both the cluster generated for microarray data with and without preprocessing of microarray data.
The yeast microarray dataset contain 6400 genes. After removal of empty genes from the dataset then there are about 6314 genes are shown in Fig 6.
Velammal College of Engineering & Technology, Madurai
Page 311
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE V. CLUSTERING ON ORIGINAL MICROARRAY DATA
2) Clustering microarray data based on functionality
The same sets of microarray gene YAL054C, YAL034C, YBL043W were grouped in different clusters before and after preprocessing. The cluster 6 contains the above said genes grouped with 670 genes before preprocessing the cluster 14 contains the same genes grouped with 350 genes.
The genes are clustered based on the ontology functionality that is Biological Process, Cellular Component and Molecular Function. These clusters were generated based on GO Ontology using k-means algorithm as shown in Table 6.
The same set of gene (YAR068W, YGR207C) was clustered in Biological Process, Cellular Component and Molecular Function but with different count value. The gene YAR068W, YGR207C was clustered in the count value of 140 genes for Biological Process, 80 genes for Cellular Component and 130 genes for Molecular Function. Some of the genes are clustered only in two
functionality. For example, the gene YEL048C, YMR259C was clustered in the count value of 65, 120 genes for Biological Process and Cellular Component alone. The pictorial representations of Table 6 are shown in Fig 8.
Velammal College of Engineering & Technology, Madurai
Page 312
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig 8. Number of cluster based on different functionality
The clustering method proposed help to find association of genes across various functionality. For each functionality (Biological Process, Cellular Component, Molecular Function) GOid are used to find association between genes. Each functionality having more then one genes are shown in Table 7. TABLE 7. DIFFERENT FUNCTIONALITY GENES
The three functionalities show that different sets of clusters can be generated from the dataset using different set of genes. The first two functionalities (Biological Process and Molecular Function) are taken and compared with genes to find out the meaningful associations of genes from the dataset. By clustering and analyzing the clusters based on the two functionalities the following genes YAR029W, YBR074W, YBR096W, YBR138C, YBR235W, YBR239C, YBR241C, YBR255W, and YCR015C are found to be have association. The clusters generated for the two functionalities are shown in Fig 9.
Fig 9.Association of genes – Biological and Molecular Functionality
The association of genes clustered using two functionalities the biological and molecular with their goid is shown in Table 8. Velammal College of Engineering & Technology, Madurai
TABLE 8. ASSOCIATION BETWEEN GENES GOID
Page 313
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
transporter or pore. The goid 15075 is a molecular function. It catalysis of the transfer of an ion from one side of a membrance to the other. The goid 22891 is also a molecular function it enables the transfer of a specific substance or group of related substances from one side of a membrance to the other. The biologically meaningful unknown function is grouped based on functionalities. In our analysis of the clustered results we have found that the set of genes grouped based on the goid 8150, 3674 always occur together and they have a unknown biological and molecular function. The cluster approach based on fucntionalty helps to finds the meaningful clusters rather than grouping the original dataset. The genes were grouped based on the goid 8150 and 3674, where the first belongs to biological process and the second go belobgs to molecular function. The genes grouped with goid 8150 belong to unknown function of biological process. Which describes any process specifically pertinent to the functioning of integrated living units: cells, tissues, organs, and organisms. The gene clustered with goid 3674 is an unknown molecular function and it describes the actions of a gene product at the molecular level. The following goid 6810, 15075 and 22891 have a different functionality. The goid 6810 is a Biological Process; it describes the directed movement of substances (such as macromolecules, small molecules, ions) into, out of, within or between cells, or within a multicellular organism by means of some external agent such as a
V. CONCLUSION Grouping of Microarray data with similar functionalities in genes are grouped based on Go ontology and is implemented using the data mining clustering technique the K-means algorithm. In the proposed work the microarray data was extracted for functionality for Biological Process and Molecular Function and implemented and tested the clusters for analyzing meaningful associations. The work proposed helped to find the association of genes using the clustering algorithm based on GO Ontology and from the study we conclude that the clustering based on functionality of genes have meaningful associations than directly clustering the dataset.
[8]
REFERENCES [1] A. Ben-Dor, R. Shamir, and Z. Yakhini, “Clustering gene expression patterns” in J Comput Biol 6(34):281-97. [2] A.Gruzdz, A.Ihnatowicz, J.Siddiqi, and B.Akhgar, “Mining Genes Relations in Microarray Data Combined with Ontology in Colon Cancer Automated Diagnosis System”, Vol.16, Nov 2006, ISSN 13076884. [3] Arun. K. Pujari, “Data Mining Techniques”, Universities press (India) Limited 2001, ISBN-817371-3804. [4] Bradley Coe and Christine Antler, “ S po t yo ur g en es – an ov erview of th e micr oarr ay”, August 2004. [5] Cheremushkin E.S, “The Modified Fuzzy C-Means Method for Clustering Of Microarray Data”. [6] Daxin Jiang, Chun Tang, and Aidong Zhang, “Cluster Analysis for Gene Expression Data: A Survey, IEEE Transactions on knowledge AND Data Engineering”, Vol 16, No. 11, November 2004. [7] D. Jiang, Ch. Tang, A. Zhang, “Cluster Analysis for Gene Expression Data: A Survey,” IEEE, Vol. 16, no. 11, November 2004.
Velammal College of Engineering & Technology, Madurai
Eisen MB, Spellman PT, Brown PO, Botstein D: Cluster analysis and display of genome- wide expression patterns. PNAS 1998, 95(25): 1486314868. [9] Gao Cong, Kian-Lee Tan, Anthony K.H. Tung, Feng Pan, “Minning Frequent Closed Patterns in Microarray Data”. In Proc. ACM SIGKDD Int,l Conf. On Knoweldge Discovery and Data mining (KDD), 2003. [10] Georage Tsastaronis and Vicky Panagiotopoulo, “A Generalized Vector Space Model for Text Retrieval Based on SemanticRelatedness” Proceedings of the EACL 2009 Student Research Workshop, Pages 70-78, 2 April 2009 © 2009 Association for Computational Linguistics. [11] G. Salton, A. Wong, and C.S. Yang, “A Vector Space Model for Automatic Indexing”, Communications of the ACM, Vol. 18, no. 11, Pages 613-620. [12] John Sikorski, “Mining Association Rules in Microarray Data”, Bioinformatics, 19, 2003. [13] LI Guo-qi, Sheng Huan-ye,”Classification analysis of microarray data based on Ontological engineering”, journal of Zhejiang Univ Sci A 2007 8(4): 638-643. [14] Manoranjan Dash and Vivekanand Gopalkrishnan, “Distance Based Feature Selection for Clustering Microarray Data”, PP. 512-519, 2008 ©SpringerVerlag Berlin Heidelberg 2008.
Page 314
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[15] Nadia Bolshakova, Francisco Azuaje, Padraig Cunningham, “An integrated tool for Microarray data clustering and cluster validity assessment”, 2008. [16] Pankaj Chopra, Jaewoo Kang, Jiong Yang, Hyung Jun Cho, Heenam, Stancey and Min-GOO Lee, “Microarray data mining using landmark gene-guided clustering”, BMC Bioinformatics, 2008, 9:92, February 2009. [17] Pasquier. C. The : Ontology – driven analysis of microarray data. Bioinformatics 20(16), 2636-2643, 2004. (www.Wisegeek.com/what is microarray.htm.) [18] Stephan Symons and Kay Nieselt, “Data Mining Microarray Data- Comprehensive Benchmarking of Feature Selection and Classification Methods”, 2007. [19] Taraskina A.S “The Modified Fuzzy C-Means Method for Clustering Of Microarray Data”. [20] Yeung K, Medvedovic M, Bumgarner R: Clustering gene-expression data with repeated Measurements Genome Biology 2003.
Velammal College of Engineering & Technology, Madurai
Page 315
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
An Energy Efficient Adavanced Data Compression And Decompression Schemes For Wsn G.Mohanbabu#1, Dr.P.Renuga#2 Lecturer, Department of Electronics and Communication Engineering, PSNA CET, Anna University #2 Assistant Professor, Department of Electrical and Electronics Engineering, TCE, Anna University
#1
1
[email protected] 2 [email protected]
Abstract - Efficient utilization of energy has been a core area of research in wireless sensor networks. Sensor nodes deployed in a network are battery operated. As batteries cannot be recharged frequently in the field setting, energy optimization becomes paramount in prolonging the batterylife and, consequently, the network lifetime. The Communication module utilizes a major part of the energy expenditure of a sensor node. Hence data compression methods to reduce the number of bits to be transmitted by the communication module will significantly reduce the energy requirement and increase the lifetime of the sensor node. The present objective of the study contracted with the designing of efficient data compression algorithm, specifically suited to wireless sensor network. Keywords – Wireless sensor network, Data compression, Huffman algorithm.
I. INTRODUCTION Wireless Sensor Network (WSN) comprises of several autonomous sensor nodes communicating with each other to perform a common task. A wireless sensor node consists of a processor, sensor, communication module powered by a battery. Power efficiency is considered to be a major issue in WSN, because efficient use of energy extends the network lifetime. Energy is consumed by the sensor node during sensing, processing and transmission. But almost 80% of the energy is spent in the communication module for data transmission in sensor network[1]. Sensor networks have a wide range of application in temperature monitoring, surveillance, bio medical, precision agriculture. Failure of sensor node causes a partition of the WSN resulting in critical information loss. Hence there is great interest shown by the many researchers in extending the lifetime of sensor nodes by reducing the energy required for transmission. Several algorithms have been proposed for energy efficient wireless sensor network in literature. The spatio-temporal correlations among sensor observations are a significant and unique characteristic of the WSN which can be exploited to drastically increase the overall network performance. The existence of the above mentioned correlation in sensor data is exploited for the development of energy efficient communication protocols well suited to WSN. Recently there is a major interest in
Velammal College of Engineering and Technology, Madurai
the Distributed Source Compression (DSC) algorithm which utilizes the spatial correlation in a sensor network for data compression. WSN application requires dense sensor deployment[1] and as a result of this, multiple sensors record information about a single event. Therefore it is unnecessary for every sensor node to send redundant information to the sink node due to the existence of high spatial correlation. Instead a smaller number of sensor measurements might be adequate to communicate the information to the sink with certain reliability. In[2] a distributed way of continuously exploiting existing correlations in sensor data based on adaptive signal processing and distributed source coding principles is discussed. In[6,8] hardware architecture for data compression using adaptive Huffman algorithm for data compression is proposed. In[3,4,7] also the spatial correlation in sensor data is exploited for the data compression. Though research has progressed in the area of data compression, not many have worked in lossless algorithms for wireless sensor networks. Compression algorithms are mainly for data storage and therefore simple and application specific algorithms are required for resource constrained sensor nodes. In[5] a simple compression algorithm based on static Huffman coding particularly suited for memory and computational resource constrained wireless sensor node is proposed. This algorithm exploits the temporal correlation in sensor data and the algorithm computes a compressed version using a small dictionary. In[5], to design the dictionary, the statistics of the data are required, unlike in real time, statistics of the data will change depending on the event. Therefore the objective of the study is to design a simple algorithm to compress sensor data which does not require prior knowledge of the statistics of sensor data and the data compression is performed adaptively based on the temporal correlation in sensor data. II. MATERIALS AND METHODS To establish a modified algorithm for data compression in wireless sensor network, the following simple and adaptive algorithms were simulated and compared.
Page 316
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Simple Huffman algorithm In the simple Huffman algorithm proposed in[5], each sensor node measure mi is converted by an ADC to binary representation ri using R bits, where R is the resolution of the ADC. For each new measure mi, the compression algorithm computes the difference di = ri - ri-1, which is input to an entropy encoder. The encoder performs compression losslessly by encoding differences di more compactly based on their statistical characteristics. Each di is represented as a bit sequence bsi composed of two parts si and ai, where si gives the number of bits required to represent di and ai is the representation of di. Code si is a variable length code generated by using Huffman coding. The basic idea of Huffman coding is that symbols that occur frequently have a smaller representation than those that occur rarely. The ai part of the bit sequence bsi is a variable length integer code generated as follows: If di>0, ai corresponds to the ni lower-order bits of the direct representation of di If di<0, ai corresponds to the ni lower-order bits of the two’s complement representation of (di-1). If di=0, si is coded as 00 and ai is not represented. The procedure used to generate ai guarantees that all the possible values have different codes. Once bsi is generated, it is appended to the bit stream which forms the compressed version of the sequence of measures mi. Here <<si,ai>> denotes the concatenation of si and ai. Since transmission of a bit needs energy comparable to the execution of thousand instructions, just saving only a bit by compressing original data corresponds to reduce power consumption[5]. We observe that the proposed algorithm is simple but has an important drawback. In order to assign probability, the statistics of sensor data must be known already. Hence this algorithm may not be suitable for real time sensor data. III. ADAPTIVE HUFFMAN ALGORITHM Huffman coding requires prior knowledge of the probabilities of the source sequence. If this knowledge is not available, Huffman coding becomes a two pass procedure: the statistics are collected in the first pass and the source is encoded in the second pass. In the Adaptive Huffman coding procedure, neither transmitter nor receiver knows anything about the statistics of the source sequence at the start of transmission. The tree at both the transmitter and the receiver consists of a single node that corresponds to all symbols Not Yet Transmitted (NYT) and has a weight of 0. As transmission progresses, nodes corresponding to symbols transmitted will be added to the tree and the tree is reconfigured using an update procedure. Considering a simple 4 bit ADC representation for each data, then before the beginning of transmission, a fixed 4 or
Velammal College of Engineering and Technology, Madurai
5 bit code depending whether the symbol is positive or negative is agreed upon between the transmitter and receiver. The actual code consists of two parts: The prefix corresponding to the code obtained by traversing the tree and the suffix corresponding to 4 or 5 bit binary representation corresponding to positive or negative data respectively. In the process of coding, the probability for the incoming source sequence is assigned as the elements get in to the tree formed. This gives rise to a problem where the elements which come in the initial stages of tree formation having lesser probability hold smaller codes. Thereby, the compression ratio obtained is lesser. The pseudo code for adaptive Huffman algorithm is shown in Fig. 1. module adaptivehuffman_encode (diff, btree) createroot() search (diff, btree) if(node present) incrementnode->wt code = traverse(node) else prefix = traverse (nyt) if(diff>=0) suffix = diff4 if (diff<0) suffix = diff5 code =<<prefix, suffix>> nyt->lchild = createnyt() nyt->rchild = createnode (diff) nyt = nyt>lchild endif update (btree) balance (btree) endmodule Fig.1 : Pseudo code of adaptive Huffman encoding algorithm.
Using Adaptive Huffman algorithm, we derived probabilities which dynamically changed with the incoming data, through Binary tree construction. Thus the Adaptive Huffman algorithm provides effective compression by just transmitting the node position in the tree without transmitting the entire code. Unlike static Huffman algorithm the statistics of the sensor data need not be known for encoding the data. But the disadvantage in this algorithm is its complexity in constructing the binary tree which makes it unsuitable for sensor nodes. Modified adaptive algorithm In static and dynamic algorithms, the ultimate objective of compression was achieved with fixed and dynamic probabilities respectively. The main disadvantage of Static Huffman algorithm is that, we do not have the prior knowledge of the incoming source sequence. Also the statistics of sensor data may be varying. In Adaptive
Page 317
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Huffman algorithm, though the probabilities are assigned dynamically, because of the increased number of data available in the source sequence, the number of levels and hence the number of bits transmitted increases and is found to be effective only for very frequently and first occurring data. The binary tree construction is based on the order of arrival of incoming data. Hence, both static and adaptive Huffman algorithms are observed to have some drawbacks. The proposed Modified Adaptive Huffman algorithm overcomes the disadvantages of both static and dynamic Huffman algorithm by combining the advantages of the two algorithms and ultimately increasing the compression ratio. The algorithm uses a tree with leaves that represent sets of symbols with the same frequency, rather than individual symbols. The code for each symbol is therefore composed of a prefix (specifying the set, or the leaf of the tree) and a suffix (specifying the symbol within the set of same-frequency symbols). The Binary Tree is constructed with the incoming elements and the codes framed by traversing the tree as in Adaptive Huffman algorithm. Elements are grouped as nodes and codes are specifically assigned as in Static Huffman algorithm. At every update the weights at every level is checked and updated such that the higher weights will be occupying the initial stages of the tree. {0} NYT 1
12001000800600400200-
{-0.3, -0.2, 0.2, 0.3} {-0.1, 0.1}
0
This enhances the performance in two ways: It reduces the number of levels in the tree. It brings the maximum possible elements to the top level of the tree. Binary tree construction
50454035302520151050 Adaptive Huffman
Binary tree construction used in the algorithm is shown in Fig. 2, where each node consists of set of elements. The temporal correlation in sensor data only the difference of the sensor data di = ri-ri-1 is encoded. Consider for example the set {-0.2, 0.1, 0.3} is transmitted. Initially the encoder will have only the NYT node. Since -0.2 occurs for the first time, it is transmitted as 10010 and a new node
Fig.3 : Compression analysis in terms of No. of bits transmitted.
Velammal College of Engineering and Technology, Madurai
Modified Adaptive Huffman
Fig.2 : Binary Tree
Static Huffman
NYT
Modified Adaptive Huffman
1
Original Data
1
Adaptive Huffman
0
1
Static Huffman
NYT
{-0.3, -0.2, 0.2, 0.3}
which contains this data is inserted. When the next data 0.1 arrives, the tree is traversed in search of the node1 which contains the data 0.1 in the binary tree. Since the node is not available, the code corresponding to NYT node i.e., 0 is transmitted, followed by 1 corresponding to the data 0.1.. So the code transmitted is 01. For the next data is 0.3, the tree is traversed in search of the node 2 containing the data 0.3. Since node 2 is already available in the tree, the prefix corresponding to node traversal 1 is transmitted. The binary representation of the array index containing the data is transmitted as suffix i.e., 11 corresponding to 3 which is the array index containing the data 0.3 is transmitted. So the code is transmitted is 111. Initially the algorithm starts with subsets each having different probability. But this does not require to transmit the tree prior to decompression. In the proposed algorithm, once we reach a leaf while decoding, we know how many bits are there still to be read and that allows us to make a single operation to get a bit string which is readily converted to an integer.
Page 318
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 4 : Compression Ratio
Compression ratio Compression ratio is calculated as defined in[5]. The formula used for compression ratio analysis is given as follows: Compression ratio = 100(1- compressed size/original size) Compressed size is the number of bits obtained after compression and original size will be the total number of bits required without using compression algorithm. The performance of the modified adaptive Huffman, static and adaptive Huffman compression algorithm were analyzed in terms of compression in number of bits required for transmission is shown in fig. 3 and compression ratio of each algorithm is shown in fig. 4. IV.CONCLUSION
[2] Chou, J. and K. Ramachandran, 2003. A distributed and adaptive signal processing approach to reducing energy consumption inn sensor networks. Proceeding of the INFOCOM 22nd Annual Joint Conference of the IEEE Computer and Communications Societies, Mar. 30-Apr. 3, IEEE Xplore Press, USA., pp: 1054-1062. http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=1208942 [3] Tang, C., C.S. Raghavendra and K.V. Prasanna, 2003. An energy efficient adaptive distributed source coding scheme in wireless sensor networks. Proceeding of the IEEE International Conference on Communications, May 11-15, IEEE Xplore Press, USA., pp: 732-737. DOI: 10.1109/ICC.2003.1204270. [4] Ye, Q., Y. Liu and L. Zhang, 2006. An extended DISCUS scheme for distributed source coding in wireless sensor networks. Proceeding of the International Conference on Wireless Communications, Networking and Mobile Computing, Sept. 22-24, IEEE Xplore Press, Wuhan, pp: 1-4. DOI: 10.1109/WiCOM.2006.286 [5] Francesco Marcelloni, 2008. A simple algorithm for data compression in wireless sensor networks. IEEE. Commun. Lett., 12: 411413. DOI: 10.1109/LCOMM.2008.080300 [6] Lin, M.B., J.F. Lee and G.E. Jan, 2006. A lossless data compression and decompression algorithm and its hardware architecture. IEEE Trans. Very Large Scale Integrat. Syst., 14: 925-936. DOI: 10.1109/TVLSI.2006.884045 [7] Pradhan, S.S., Julius Kusuma and K. Ramachandran, 2002. Distributed compression for dense microsensor networks. IEEE. Signal Proc. Mag.,19: 15-60. DOI: 10.1109/79.985684 [8] Cesare, A., Romolo Camplani and C. Galperti, 2007. Lossless compression techniques in wireless sensor networks: Monitoring microacoustic emissions. Proceeding of the IEEE International Workshop on Robotic and Sensor Environments, Oct. 12-13, IEEE Xplore Press, Ottawa, Canada, pp: 1-5. DOI: 10.1109/ROSE.2007.4373963 [9] http://www.mass.gov/dep/water/resources/
A simple algorithm namely modified adaptive Huffman algorithm which does not require prior knowledge of the statistics of sensor data is proposed for data compression in the sensor network. The data compression is performed adaptively based on the temporal correlation in the sensor data. This modified adaptive Huffman encoding algorithm effectively combines the advantages of static and adaptive Huffman algorithms to provide effective compression by reducing the number of levels in the binary tree. The implementation of this algorithm shows better compression ratio than adaptive Huffman algorithm. Since the number of levels is restricted, this algorithm requires less computation than adaptive Huffman which is an important requirement for wireless sensor nodes. This algorithm outperforms the static and adaptive Huffman algorithm in providing better performance in terms of number of bits transmitted. Future Work The proposed work can be implemented and tested using Beagle board advanced multi core DSP processor and the hardware results can be verified and compared with the software simulation results. V.REFERENCES [1] Akylidiz, I.F., W. Su, Y. Sankarasubramaniam and E. Cayirici, 2002. Wireless sensor networks: A survey. Comput. Networks, 38: 393422. DOI: 10.1016/S1389-1286(01)00302-4
Velammal College of Engineering and Technology, Madurai
Page 319
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Active Noise Control: A Simulation Study #
Sivadasan Kottayi#1, Narayanan N.K.*2 Electrical & Electronics Engineering Technology Department, Yanbu Industrial College, Yanbu Al-Sinaiyah, Yanbu, KSA 1
[email protected]
* School of Information Science & Technology, Kannur University, Kerala, India. 2
[email protected]
AbstractIn the last decade active noise control has emerged as a viable technology for cancellation of low frequency noise. This paper explains the basic principle of active noise control and a detailed account of broad band feed forward active noise control system together with the necessity of nullifying the secondary path as well as acoustic feedback problems. A simulation study of active noise control system for a passenger train cabin, an example of broad band ANC has been carried out using the noise signal collected from an air conditioned train cabin. Though it is a multichannel ANC problem having different types of complexities, it has been simplified with certain assumptions and simulated for a single channel using different adaptive algorithms. Among the five algorithms tried, normalized LMS algorithm is found working well in this environment.
KeywordsActive noise control, Adaptive noise cancellation, Adaptive systems, Digital Signal Processing applications I. INTRODUCTION Active noise control (ANC) is an emerging research area in engineering & technology having numerous applications. It promises viability of numerous quality products. One of the most successful ANC products in the market is active noise cancelling headset which reduces undesired noise going to human ears. This increases the work efficiency of factory & aerospace employees. The advent of high speed cost effective Digital Signal Processing chips enhances the feasibility of many ANC products [1,2,3] which in turn increases the comfort of human beings while they are at home, travelling or in office. Sound frequencies audible to human ears lie in the frequency range of 20 Hz to 20 KHz. However, only a certain band of frequencies is pleasurable to human ears. The hustle and bustle of market places, noise in buses and trains, rumble of air- conditioners, hissing of fans - all make human beings exhausted. Such disturbing noises need to be eliminated. II. SCHEMES FOR ELIMINATING NOISE There are two types of noises: high frequency noise and low frequency noise. High frequency noise can be reduced through conventional passive method [4], which employs heavy barriers to block the transmission of sound and also use certain acoustic materials to absorb sound energy. However, these techniques are found less
Velammal College of Engineering and Technology, Madurai
effective at low frequency (< 500Hz) as to stop such kind of noise one need to have barriers with large width, which are expensive and often difficult to implement. Heavy and bulky barriers are prohibited in aerospace applications. In the case of mobile telephony, it is arduous to attend the incoming telephone calls in a noisy environment. Moreover, when you make a telephone call in such an environment, the surrounding noise also propagates along with the voice signal to the destination, which is also not desirable. In all these situations conventional passive techniques cannot be applied, but noise can be eliminated electronically via active noise control. III. BASIC PRINCIPLE OF ACTIVE NOISE CONTROL Active noise control (ANC) is based on destructive interference between two sound fields: one is the original or primary sound source, which has to be eliminated and the other is secondary sound source or control signal generated by the Electronic Controller, having the same frequency of primary sound but 180° out of phase, propagated through a loudspeaker to interfere with the primary sound so that both signals get cancelled by themselves. Since sound propagation is considered to be linear and the principle of superposition is applicable to it, their acoustic pressures cancel themselves. However, how much an Active Noise Control is effective depends upon the location of the sensing microphones and the secondary signal source. It also depends upon the acoustic environment as to whether the pressure wave propagates freely in air or is enclosed within a confined space. Destructive interference is most effective when two sound fields are accurately aligned in space over an acoustic wavelength. IV ACTIVE NOISE CONTROL TECHNIQUES In ANC two types of control techniques are in practice – one is Feedback ANC and the other is feed forward ANC. Both the techniques are explained below. A. Feedback ANC Feedback control approach was first tried for active noise cancelation by Olson and May [5] in the year 1953. A physical illustration of such a system in a ventilation duct is shown in Fig. 1 and its equivalent electrical block diagram is shown in Fig. 2. In Fig. 2 ‘e’ represents the signal derived from the monitor (error) microphone, which is the combined effect of the
Page 320
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
disturbance ‘d’ and the output of feedback loop. ‘W’ was a simple gain and phase inversion circuit described by Olson and May. The electrical transfer function from loudspeaker input to microphone output ‘S’ is called secondary or error path. It contains the electro – acoustic response of the loudspeaker, electro acoustic response of error microphone and the acoustic characteristics of the path between loudspeaker and error microphone. The transfer function between the disturbance or primary source and measured error is thus
Secondary source
Monitor microphone
Primary Source
Duct
Loudspeaker
Fig. 1
Electronic controller
A single channel feedback ANC System in a duct
microphone. Noise cancelling head set, which is one of the most successful ANC products in the market works on this principle as the proximity of eardrum is close to error microphone. However, most of the other applications except purely periodic disturbances, Feedback ANC is not successful because of the restriction of placing error microphone close to secondary source. This is another major disadvantage of Feedback ANC. Therefore, Feed forward ANC is widely used for active noise cancellation. B. Feed-forward ANC Feed forward ANC can be further divided into Broad-band feed forward ANC and Narrow band feed forward ANC. In this paper we concentrate on broad band feed forward ANC as the paper study about the noise inside a passenger train cabin, which is considered as broad band. The noise we hear inside an air-conditioned train cabin is a composite noise consists of engine sound, body vibration, track sound, sound of air-conditioner and the noise made by passengers. A broad band ANC system that has a single reference microphone, single secondary signal source and single error microphone is shown in Fig. 3. Signal processing time of the controller must be equal to the time taken by the acoustic signal to reach from the reference sensor to the secondary source; or else the ANC system may not be effective in reducing the noise.
d Loudspeaker
+ S
+
W
Fig. 2
Σ
e
Equivalent electrical block diagram could be increased in the feedback path without limit causing the overall transfer function of the feedback loop to become small. The feedback loop forcing ‘e’ to be small compare to ‘d’ will be to cancel acoustic pressure at the error microphone as required for active noise control. However, the frequency response of the secondary path can never be made perfectly flat and free from phase shift. The loudspeaker introduces considerable phase shift. The acoustic path from loudspeaker to microphone involves some delay due to the acoustic propagation time and this will also introduce an increasing phase shift in the secondary path. If the phase shift in the secondary path approaches 180°, the negative feedback described above becomes positive feedback and the control system become unstable is a disadvantage of Feedback ANC. If the system is digital as it is imperative for implementing adaptive control, delay in secondary path will be much more due to the process delay of A/D converter, anti-aliasing filter, D/A converter and reconstruction filter. Due to this reason feedback ANC is normally implemented using analog circuit and secondary source (loudspeaker) should be placed very close to error
Velammal College of Engineering and Technology, Madurai
Primary Source
Secondary source
Error microphone
Reference microphone Anti-aliasing filter and A/D t
D/A converter & reconstructtion filter
Anti-aliasing filter & A/D converter
Digital filter
Fig. 3
A Feed forward Adaptive ANC Algorithm
The broad band feed-forward ANC system shown in Fig. 3 is illustrated in an adaptive system identification configuration as in Fig. 4, in which an adaptive filter W(z) is used to estimate an unknown plant P(z). The primary path P(z) consists of an acoustic response from the reference sensor to the error sensor where the noise attenuation is to be realized. If the plant is dynamic, adaptive algorithm has to continuously track time variations of the plant dynamic. In Fig. 4, E(z) = 0 after adaptive filter W(z) converges. We then have W(z) = P(z) for X(z) ≠ 0. Which implies that the adaptive filter output y(n) is identical to the primary disturbance d(n). When d(n) and y(n) are acoustically combined, the residual error is e(n) = d(n) – y(n) = 0 which results in perfect
Page 321
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
cancellation of both sounds based on the principle of super position. More insight into the system reveals that we have to compensate secondary path transfer function S(z) from y(n) to e(n) which includes D/A converter, reconstruction filter, power amplifier, loudspeaker, acoustic path from loudspeaker to error microphone, error microphone, preamplifier, anti-aliasing filter and A/D converter. All these components add noise and delay in generating cancelling signal y(n). A scheme suggested by Morgan [6] x(n)
Unknown Plant P(z)
d(n)
e(n) Σ
+
-
y(n) Digital filter W(z)
Fig. 4
x(n)
System identification view point of ANC
Adaptive filter W(z) (z)
e(n)
d(n)
Unknown Plant P(z)
Σ
S(z) y(n)
LMS Algorithm Fig. 5 Block diagram of ANC system using the FXLMS algorithm. to solve this problem is to place an identical filter in the reference signal path as shown in Fig. 5 to weight update of LMS algorithm which realizes the Filtered X-LMS (FXLMS) algorithm [7]. FXLMS algorithm for active noise control has been derived by Burgess [8] as In ANC applications, S(z) is unknown and must be . Therefore, the estimated by an additional filter filtered reference signal is generated by passing the reference signal through this estimate of the secondary path as
Velammal College of Engineering and Technology, Madurai
is the estimated impulse response of the where . With in the limit of slow secondary path filter adaptation, the algorithm will converge with nearly 90° of and S(z). Of-line modeling and phase error between online modeling can be used to estimate [1] S(z). Another problem in feed-forward ANC is the radiation of cancelling signal y(n) towards the reference microphone, resulting in a corrupted reference signal x(n). The coupling of the acoustic wave from the canceling loudspeaker to the reference microphone is called acoustic feedback. The feedback component of the reference microphone signal can be cancelled electronically using a feedback neutralization filter , which model the feedback path. The models and can be estimated simultane-ously by using the offline modeling technique [1] The noise field in a reasonably large enclosure, for example, a train compartment is more complicated than a narrow duct. The sound field inside a passenger train cabin is created by standing waves rather than propagating waves and depends on the superposition of a number of acoustic modes. A mode is characterized by the number of wavelengths that fit along one dimension of the enclosure. Theoretically, each mode contains fundamental frequency plus infinite number of harmonics [9]. For perfect multichannel active noise control, infinite number of microphones are essential to sense each harmonic and also infinite number of loudspeakers are required for sending control signals, which is practically impossible to implement. However, many microphones and loudspeakers, well placed throughout the enclosure are needed to make the enclosure reasonably quiet. Situations will be worsened when the characteristics of primary signal changes with time as it is very common in ANC applications. The spectrum of the sound inside a train compartment changes when it passes through a tunnel or over a bridge. Thus a good ANC system able to tackle such situations is quite challenging. To encounter such conditions, a multichannel adaptive noise control system should be devised with suitable adaptive control algorithm, which can generate secondary signals for nullifying the noise. Primary source
P +
M
+ F
Σ
Σ
S M
J
x(n)
d(n)
K
W
y(n)
e(n)
J M
Page 322
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Block diagram of a multiple channel feed forward acoustic ANC system that includes secondary path as well as feedback path is illustrated in Fig. 6. The system has J reference sensors, K cancelling signals to drive the corresponding secondary sources and M error sensors. The wide arrows represent an array of signals (acoustic or electrical) that are symbolically expressed as vectors. primary path transfer function Matrix P represents Pmj(z) from the primary source to each error sensor output secondary path em(n). The matrix S represents transfer function Smk(z) from K secondary sources to M feedback error sensors. Also the matrix F represents path transfer function Fjk(z) from K secondary sources to J reference sensors. There are possible feed forward channels, each demanding a separate adaptive filter and adaptive filters are presented by the matrix W. these Full description of a multiple channel ANC using FXLMS algorithm is given in [1]. V. SIMULATION STUDY OF ACTIVE NOISE CONTROL SYSTEM FOR A PASSENGER TRAIN CABIN Active noise control of a closed chamber like train cabin is a multichannel system problem as explained earlier. However, for study purpose only one channel has been considered and simulated using matlab. The primary noise x(n) has been collected inside an air-conditioned train compartment. Thousand samples with sampling interval of 0.001 second are used for simulation. The primary path transfer function is modeled as low pass filter which is achieved using matlab command b=fir1(n,Wn), returns row vector b containing n+1 coefficients. This is a hamming window based linear phase filter of order n with normalized cut off frequency Wn. In simulation the order of the filter is fixed as 32. In an ANC system, secondary path introduces delay as well as noise in control signal y(n) which is due to the electronic subsystems, loud speaker and error microphone explained earlier. For simplicity, instead of estimating the secondary path transfer function S(z), one of its most important features of introducing noise in y(n) is taken into consideration and hence random noise has been added with control signal y(n). Matlab command n = 0.02*randn(1000,1) generates a scaled version of normally distributed random numbers of matrix with mean = 0 , variance = 1 and standard deviation = 1, which represents electronics noise of secondary path. The electronic noise added to control signal is approximately 1% of the train noise. The delay factor is not considered in simulation as it can be compensated while practically implementing the system by proper placement of secondary sources and error microphones. The feedback path transfer
Velammal College of Engineering and Technology, Madurai
function F(z) has been omitted in this simulation study as it is assumed that secondary source is made directing towards error microphone as well as the reference sensor is assumed as directional microphone and placed in the opposite direction of loudspeaker. These assumptions made us to eliminate feedback problem. TABLE I SIMULATION RESULTS Diff. algorithms
Resultant noise inside the train cabin for different values of µ µ=0.25
µ=0.26
µ=0.27
µ=0.28
µ=0.29
µ=0.30
-23.56 µ=1.8
-23.67 µ=1.9
-23.79 µ=2.0
-24.02 µ=2.1
-23.62 µ=2.2
-17.70 µ=2.3
-26.65 µ=0.08
-27.01 µ=0.09
-27.19 µ=0.10
-27.60 µ=0.11
-27.94 µ=0.12
-27.79 µ=0.13
-23.26 µ=0.006
-23.80 µ=0.007
-23.67 µ=0.008
-24.61 µ=0.009
-25.10 µ=0.01
-25.25 µ=0.02
-15.90 µ=0.004
-16.14 µ=0.005
-16.57 µ=0.006
-16.55 µ=0.007
-16.77 µ=0.008
-15.84 µ=0.009
-16.42
-16.30
-16.11
-15.57
-14.78
-14.73
LMS NLMS SDLMS SELMS SSLMS
The ANC system in this environment has been simulated on a digital computer using different control
Primary noise (x) 0.5 0.4 0.3 0.2 Signal value
Fig. 6 Block diagram of an adaptive multiple-channel feed forward ANC system with feedback paths.
0.1 0 -0.1 -0.2 -0.3 -0.4
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Fig 7 Original noise in side the cabin Fig. 8(a) Resultant noise when NLMS algorithm is used (-27.94 dB for µ=2.2) algorithms [1,10] namely LMS adaptive algorithm, Normalized LMS adaptive algorithm, Signed data LMS adaptive algorithm, Sign error LMS adaptive algorithm, Sign- sign LMS adaptive algorithm. In each case performance of the system has been evaluated on the basis of the minimum output power of the resultant noise (error signal) after the destructive interference between primary and secondary noise. In all cases the simulation has been done by keeping the length of the adaptive filter as 32 taps. Just for an example, the matlab command ha=adaptfilt.lms (32,mu) constructs a 32nd order FIR LMS adaptive filter. The parameter mu is the step size of adaptive filter. By adjusting the value of mu the optimum performance of the filter can be achieved. In the same manner ANC systems using other algorithms are simulated and their performance
Page 323
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 8(c) Resultant noise when LMS algorithm is used (-24.02) for µ=0.28) plot of Error signal (e) 0.6
0.5
0.4
Signal value
is tabulated in Table-1 for different values of µ. . The primary signal is shown in Fig. 7. The error signals obtained while using different algorithms mentioned above are depicted in Fig. 8(a-e). Among the five control algorithms experimented, NLMS algorithm performed well in this environment. It could reduce the train cabin noise to a minimum level of -27.94 dB
plot of Error signal (e)
0.3
0.2
0.1
0.2
0 0.15
Signal value
-0.1 0.1
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Fig. 8(d) Resultant noise when SELMS algorithm is used (-16.77 dB for µ=0.01)
0.05
plot of Error signal (e)
0
0.6 -0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
0.5
plot of Error signal (e)
0.4 Signal value
0.3
0.25
Signal value
0.2
0.3
0.2
0.1
0.15 0
0.1 -0.1
0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
0
-0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Fig. 8(b) Resultant noise when SDLMS algorithm is used (-25.25 dB for µ=0.13) plot of Error signal (e) 0.3
0.25
Signal value
0.2
0.15
Fig. 8(e) Resultant noise when SSLMS algorithm is used (-16.42 dB for µ=0.004) VI. CONCLUSION A single channel ANC system for passenger train cabin has been simulated using the real signal collected from an air-conditioned train cabin. Simulation study reveals that normalized LMS algorithm works well in this environment compare to other experimented algorithms as it could reduce more cabin noise. The study can be extended by incorporating the estimated transfer functions of secondary path as well as acoustic feedback path in the system together with problems associated in a multichannel ANC system.
0.1
0.05
0
-0.05
0
100
200
300
400 500 600 Signal samples
700
800
900
1000
Velammal College of Engineering and Technology, Madurai
Page 324
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
REFERENCE [ 1] S.M. Kuo and D.R. Morgan, Active Noise Control Systems – Algorithms and DSP implementations. New Yourk: Wiley, 1996. [ 2] S.J. Elliott and P.A. Nelson, “Active Noise Control,” IEEE Signal Processing Magazine, October 1983. [3] Colin H. Hansen, Understanding active noise cancellation, Spon Press, London, 2001 [4] L. L. Beranek and I. L. Ver, Noise and Vibration Control Engineering – Principles and applications. New York, Wiley, 1992. [5] Olson H.F & May E.G “Electronic sound absorber,” Journal of the Acoustical Society of America, 25, pp 1130-1136, 1953. [6] D.R. Morgan, “An analysis of multiple correlation cancellation loops with a filter in the auxiliary path,” IEEE Trans. Acoust., speech, Signal Processing, Vol. ASSP-28, pp. 454-467, Aug. 1980. [7] B. Widrow and S.D. Stearns, Adaptive Signal Processing. Englewood Cliffs, NJ: Prentice-Hall, 1985. [8] J.C. Burgess, “Active adaptive sound control in a duct: A Computer simulation,” Journal of Acoust. Soc. of America, vol. 70, pp. 715-726, Sept. 1981. [9] Stephen J. Elliott, “Down with Noise,” IEEE Spectrum, June 1999. [10] Monson H. Hayes, Statistical Digital Signal Processing & Modeling. John Wiley & Sons, 1996.
Velammal College of Engineering and Technology, Madurai
Page 325
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Texture Segmentation Method Based On Combinatorial Of Morphological And Statistical Operations Using Wavelets V.Vijayapriya1, Prof.K.R.Krishnamoorthy2 1 M.E, CSE, Anna university,Trichy /2 HOD, Dept. of ECE, SACS MAVMM Engineering College, Madurai. [email protected]
ABSTRACT Segmentation process for texture analysis is widely used for pattern recognition, industrial automation, biomedical image processing and remote sensing. Segmentation based on combinations of morphological and statistical operation is preferred in this project on wavelet transform images.The features like shape, size, contrast or connectivity makes attractive usage of mathematical morphology and can be considered for segmented oriented features.In order to get a better performance with wavelet transform for haar, db1, db6, coif6 and sym8 the system proposes derived equation on dilation, erosion, mean and median which results in segmentation.The process divides the wavelet combinatorial segmentation algorithm in to 3 groups based on type of operation and number of operations. The method using wavelet transform is applied on Brodatz Texture which results in good segmentation. The 500 Brodatz Texture image considered for performance Analysis which is compared with derived equations and filters in order to get a segmented results. The segmented results are used in comparative study with Peak Signal to Noise Ratio (PSNR) technique. Keywords: Discrete Wavelet transform, morphological and statistical operation, filter analysis.
I.
INTRODUCTION
The aim of the present work is to explore the improved version of wavelet based texture segmentation derived by (a) the method used for the construction of the multiresolution image with morphological operations, and (b) the combinatorial coefficient selection process designed to emphasize the information that is most important in the segmentation process.
1.1 Overview of texture analysis Texture analysis is an important issue in many areas like object recognition, image retrieval study, medical imaging, robotics, and remote sensing. Despite the development of a family of techniques over the last couple of decades, there are only a few reliable methods. Multiresolution techniques seem to be attractive for many applications. In this study, present an
Velammal College of Engineering and Technology, Madurai
approach based on the discrete wavelet transform and morphological concept. We integrate the framework of Brodatz texture images with the transform coefficients to obtain a flexible method for texture segmentation. Compared to intensity (spatial domain), the wavelet coefficients appear to be more reliable with respect to noise immunity and the ease of feature formation The wavelet transform is a multi-resolution technique, which can be implemented as a pyramid or tree structure and is similar to sub-band decomposition. There are various wavelet transforms like Haar, Daubechies, Coiflet, Symlet and etc. They differ with each other in the formation and reconstruction. The wavelet transform divides the original image into four subbands and they are denoted by LL, HL, LH and HH frequency subbands. The HH subimage represents diagonal details (high frequencies in both directions ), HL gives horizontal high frequencies (vertical edges), LH gives vertical high frequencies (horizontal edges), and the image LL corresponds to the lowest frequencies. At the subsequent scale of analysis, the image LL undergoes the decomposition using the same filters, having always the lowest frequency component located in the upper left corner of the image. Each stage of the analysis produces next 4 subimages whose size is reduced twice when compared to the previous scale. i.e. for level ‘n’ it gives a total of ‘4+(n-1)*3’ subbands. The size of the wavelet representation is the same as the size of the original. The Haar wavelet is the first known wavelet and was proposed in 1909 by Alfred Haar. Haar used these functions to give an example of a countable orthonormal system for the space of square integrable functions on the real line. The Daubechies wavelets (1992) are a family of orthogonal wavelets defining a discrete wavelet transform and characterized by a maximal number of vanishing moments for some given support. With each wavelet type of this class, there is a scaling function which generates an orthogonal multiresolution analysis.
Page 326
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
II. METHODOLOGY
Input image
The figure 1.1 consists of three blocks with arrow indicating the process between the blocks. Each block represents a separate module.
DWT(FILT ERSELECTI ON
MORPHOLOGICAL
SEGMENTATIONSYST EM
Figure 1 overview of the project 2.1 Wavelet Transform for Image Analysis (a) DWT The first module of the project is concerned with the implementation of the image decomposition using Lifting wavelet Transform, where the input images are divided in to four number of coefficients like Approximation, Horizontal, Vertical and Diagonal. (b) Filter selection Selecting the appropriate wavelet filter for achieving good segmented result where used orthogonal filters such as Haar, Daubechies, Symlet and Coiflet filters. 2.2 Morphological and Statistical Operation The second module of the project is concerned with the morphological operations using Erosion and Dilation and statistical operations using Mean, Median filters with some randomly selected coefficients as alpha, beta and gamma for getting high quality segmentation. 2.3Segmentation system The third module is concerned with comparative analysis of the Brodatz texture image segmentation using DWT for 500 set of standard images. The images used in this analysis were obtained from internet. The Eight types of segmented results are evaluated by using PSNR and Distortion Measure technique to identify the quality of segmented images among haar, daubechies, coiflet and symlet wavelet filters. III. RELATED WORK
Velammal College of Engineering and Technology, Madurai
The concept of morphology proposed many segmentation algorithms based on the orientation and type of Structuring Element (SE). But this paper advocates a new method of segmentation of images, based on combinatorial approach of mathematical morphology and primitive statistical operations. A gray level image typically consists of both bright and dark object features with respect to size or scale. The basic objective of the present segmentation algorithm is to isolate or sketch out the most optimal contours of these bright and dark features. [1] One of the most popular filter pairs is the 9/7 biorthogonal pair of Cohen, Daubechies, and Feauveau, which is adopted in the FBI finger-print compression standard. Present a technique to rationalize the coefficients of wavelet filters that will preserve biorthogonality and perfect reconstruction. Furthermore, most of the zeros at z = -1 will also be preserved. These zeros are important for achieving regularity. The rationalized coefficients filters have characteristics that are close to the original irrational coefficients filters. Three popular pairs of filters, which include the 9/7 pair, will be considered [2]. IV. PROPOSED WORK The present project advocates a new method of segmentation of images, based on combinatorial approach of mathematical morphology and primitive statistical operations. The segmentation results depend upon the combinatorial coefficients of α, β and γ. This dependency is restricted to 95% by keeping the condition α > β and α > γ. S[i,j]=α*D[i,j]-β*E[i,j]- γ *(D[i,j]-E[i,j])
….(3.1)
S[i,j]=α*D[i,j]-β*E[i,j]- γ *(D[i,j]+E[i,j]) ….(3.2) D [i, j], E [i, j] represents Dilated and Eroded values of the 3×3 mask respectively, S [i, j] represents segmented grey level value of the mask. The group two combinatorial segmentation is represented by equations (3.3) and (3.4). S[i, j] =|α *D[i, j] − β * E[i, j] −γ * Avg | …(3.3) S[i, j] =|α *D[i, j] − β * E[i, j] −γ * Med | …(3.4) Where α > β and α > γ , ‘Avg’ represents average value of the 3×3 mask and ‘Med’ represents median value of the 3×3 mask. The segmentation of group three forms from the two operations of morphology and/or statistics. The mathematical representation for group three combinatorial segmentation is given in equations (3.5), (3.6), (3.7) and (3.8). S[i, j] = [α *D[i, j]− β *Avg]
…(3.5)
Page 327
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
S[i, j] = [α *Avg − β *E[i, j]]
…(3.6)
25 Eq3
20
S[i, j] =[α *D[i, j]− β *Med]
…(3.7)
S[i, j] = [α *Med − β *E[i, j]]
…(3.8)
10
Where α ≥ β. This is needed to keep the resultant value of equations (3.5), (3.6), (3.7) and (3.8) as positive.
V. DISCUSSION
EXPERIMENTAL
RESULTS
AND
The comparative analysis of various wavelet filters using wavelet transforms has been performed for Approximated value and detailed value with the combinatorial coefficients 431 is analyzed.The PSNRvalue for one level decomposition has been performed.
5
Eq1
Eq7
Eq5
Eq6
Eq8
Approxim ation
Eq 4
Eq3
25
Eq4
20 15 10 5
Eq 1
Eq7
Eq6
Fig‐3 Approximation and detailed coefficients segmentation of coif6 filter using alpha=4, peta=3, gama=1
Eq 3
Eq1 Eq1
Detailed
Eq8
Eq 6
EQ5
Eq7 Eq3 Eq5 Eq6 Eq8 Eq2 Eq4
Eq6
Eq7 Eq8
0
EQ5 Eq 7
Eq 1 Eq 3 Eq 5 Eq7 Eq 2 Eq 4 Eq 6 Eq 8
EQ5
Eq8
Detailed
15 10
Eq4
-5
20
Eq2
30
25
Eq3
Eq1
30
5 0
Eq4
15
Approximation
0
Detailed
Approxim ation
Fig‐1 Approximation and detailed coefficients segmentation of sym8 filter using alpha=4, peta=3, gama=1
Fig‐4 Approximation and detailed coefficients segmentation of Db6 filter using alpha=4, peta=3, gama=1.
Inf
5 4
Eq3
25
3
Eq4
1.458
2
20
1.722 -1.883
1
15
10
0 EQ5 Eq1
Eq3 Eq2
Eq4
Eq6
Eq8
-1
Eq7
Eq1
Eq7
Eq5
Eq6
Eq8
-2
5
Db1 Detailed
Db6
Sym 8
Coif6
0 Approximation
Fig‐5 Segmentation2 result PSNR value comparision of Haar with Db1, Db6, sym8 and coif6 using alpha=4, peta=3,gama=1
Fig‐2 Approximation and detailed coefficients segmentation of Db1 filter using alpha=4, peta=3, gama=1
Velammal College of Engineering and Technology, Madurai
Page 328
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
conference. Acoust, Speech, Signal Processing, Volume. 6, page no. VI–513. April. 2003
VI. CONCLUSION AND FUTURE WORK The comparative analysis of various wavelet filters using wavelet transforms has been performed for Approximated value and detailed value with the combinatorial coefficient 431 is analyzed. A graph was drawn for PSNR vs filter, from the graph obtained that the db1 filter provides better performance over the remaining filter. The comparative analysis of various wavelet filters for Detailed value and approximated value with various combinatorial coefficients has to be performed with PSNR and Distortion measure techniques.
[10]
Michael Weeks,” Digital Signal Processing Using MATLAB and Wavelets”, 2007 Edition. Infinity Science Press LLC. (An imprint of Laxmi Publications Pvt. Ltd.)
[11]
JuHan, Kai-Kuang Ma, “Rotation invariant and scale-invariant Gabor features for texture image retrival”, science direct image and vision computing 25 page no. 1474-1481, 2007.
[12]
Dogu Baran Aydogan, Markus Hannula, Tuukka Arola, Jari Hyttinen. “Texture based Classification and Segmentation of Tissues using DT-CWT feature extraction methods”, 21st IEEE International Symposium on Computer-Based Medical Systems. page no. 614-619, 2008.
REFERENCES
[1]
Vijayakumar.V, Raju.U.S.N and Narasimha Rao.A. “Wavelet based texture segmentation based on combinatorial of Morphological and Statistical operations”. IJCSNS International Journal of Computer Science and Network Security , volume Page no 176-180 August 2008
[2]
Jianyu Lin and Mark J. T. Smith “New Perspectives and Improvements on the Symmetric Extension Filter Bank for Subband/Wavelet Image Compression” -IEEE Transactions on signal processing, volume. 17, NO. 2, page no 177 - 189 February 2008.
[3]
Balasingham.I and Ramstad, “Are the Wavelet T.A Transforms the Best Filter Banks for Image Compression?” research article, Hindawi Publishing Corporation EURASIP Journal on Image and Video Processing, page no 1-7, Article ID 287197, 2008.
[4]
Sweldens.W, The lifting scheme: “a coustom-design construction of biorthogonal wavelets” appl. Comut.harmon.anal.volume no. 3,page no.186-200, 1996
[5]
sweldens.w, “the lifting scheme: a construction of second generation wavelets,” SIAM J. Math.Anal,page no .511-546,1997.
[6]
Torsten Palfner and Erika Muller, “effects of symmetric periodic extension for multiwavelet filter banks on image coding”. IEEE Institute of telecommunications and information technology university of Rostock, Germany,1999.
[7]
Bryan E. Usevitch, “A tutorial on modern lossy wavelet image compression: foundations of JPEG 2000” IEEE signal processing magazine, September 2001.
[8]
[9]
Soman K.P and Ramachandran K.I,”Insight into wavelets from theory to practice, 2006 edition. Prentice hall of India private limited. Adiga A, Ramakrishnan K. R, and Adiga B. S “A Design and implementation of orthonoramal symmetric wavelet transform using PRCC Filter Banks”-in proceeding IEEE International
Velammal College of Engineering and Technology, Madurai
Page 329
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fpga Design Of Routing Algorithms For Network On Chip R.Anitha#1, Dr.P.Renuga#2 Lecturer, Department of Electronics and Communication Engineering, PSNA CET, Anna University, India #2 Assistant Professor, Department of Electrical and Electronics Engineering, TCE, Anna University, India
#1
1
[email protected] 2 [email protected]
Abstract: In this paper, we present a methodology to develop efficient and deadlock-free routing algorithms for Networkon-Chip (NoC) platforms that are specialized for an application or a set of concurrent applications. The proposed methodology, called the Application- Specific Routing Algorithm (APSRA), exploits the application-specific information regarding pairs of cores that communicate and other pairs that never communicate in the NoC platform to maximize communication adaptivity and performance. Based on the selected network topology, outgoing packets are switched either directly to the East/West router outputs or to the second crossbar for the North/South router outputs. Incoming packets that arrive through the North/South ports that are connected to the second crossbar are switched directly to the attached node element, but packets arriving through the East/West ports must be switched from the first crossbar to the second crossbar in order to reach the attached node element. In this paper we store the Routing table in compressed form in the router. Hence the number of gates required is reduced. In future also we perform the Routers data compression by using Modified Run length encoding which makes the fast data transfer.
1. INTRODUCTION Network on chip has emerged as a dominant paradigm for the synthesis of multicore SoCs they are generally viewed as the ultimate solution for the design of modular and scalable communication architectures and provide inherent support for the integration of heterogeneous core through the standardization of the network boundary. The network topology and routing algorithm used in the underlying onchip communication network are the two important aspects. A highly adaptive routing algorithm has a potential of providing high performance, fault tolerance and uniform utilization of network resources. Generally routing algorithms providing high adaptiveness guarantee freedom from deadlock by means of virtual channels (VCs).For arbitary network topologies, a relatively large number of VCs could be required to provide deadlock freedom, high adaptivity, and shortest paths. However the use of the VCs introduces some overhead in terms of both additional resources and mechanisms for their management. To limit such overhead, some proposals try to minimize the number of VCs required to obtain a high degree of adaptivity. For these reasons, we believe that reducing as much as possible
Velammal College of Engineering and Technology, Madurai
the number of VCs required to provide deadlock-free routing and, at the same, guaranteeing a high degree of adaptiveness must be considered one of the primary concerns in the NoC domain. In this paper, we present a methodology to design highly adaptive routing algorithms for NoC architectures that do not require the use of any VCs. In order to develop efficient, highly adaptive, deadlock-free and topologyagnostic routing algorithms. We call algorithms developed using this information as Application-Specific Routing Algorithms. The APSRA methodology is general, in the sense that it is not designed for a given/fixed network topology. 2. APSRA DESIGN METHODOLOGY Communicatio T
Applic ation
T Task T
T
Ta sk
T 1 T 2
T
Mappi ng Topology Graph T l
I P P 7
M
APSR
Compres sion
C C
P6
Cb
dS C
Rout i
Compre ssed
Configure/Rec
Fig. 1. APSRA Design Methodology
The APSRA design methodology is depicted in fig.1. It gets as inputs the application modeled as a task graph and the network topology modeled as a topology graph. Using Task Mapping information, APSRA generates a set of routing tables that not only guarantee both the reachability and the deadlock freeness of communication among tasks
Page 330
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
but also try to maximize routing adaptivity. A compression technique can be used to compress the generated routing tables. Finally, the compressed routing tables are uploaded to the physical chip. 2.1. ARCHITECTURAL IMPLICATIONS OF APSRA (TL,BR,Colo InReg ion Top
cfg setup dst
InReg ion Top
2.2. ROUTING TABLE COMPRESSION
hit ao
Admissible Outputs
M
Dem
dst
addr
ao
cfg setup dst
cfg
additional degree of freedom in routing algorithms as they can be reconfigured at runtime to adapt to the current traffic condition or offline to manage network topology variation due to manufacturing defects. A disadvantage is that a table can take a large space if many destination addresses should be stored. That is, in such an implementation, the cost (silicon area) of the router will be proportional to the size of the routing table.
hit
.
We proposed a method to compress the routing table to reduce its size such that the resulting routing algorithm remains deadlock free and has high adaptivity. The basic idea of the compressing method is shown in Fig.3 Let us suppose that the routing table associated to the west input port of node X is that shown in Fig. 3a.
setup ao
cfg setup
InReg ion Top
dst
hit
Enco der M:lg
Data En Cfg=(TL,BR,Col setup dat
Re TL
Data En
Re
BR color
InReg ion
ao hit
dat En
Re
Fig. 2. Block diagram of the Architectural Implications of APSRA
One way to implement the routing function is to design it in hardware logic (also known as the algorithmic approach). In this case, there can, e.g., be a state machine that takes the current node address, the destination address, and some status information stored in the router as inputs and outputs the admissible out ports. Another way to implement the routing function is to use a routing table. The index to the table, where the admissible outputs are stored, is the destination address or a function of the destination address. The values of the table are dependent on which router it is I Implemented. The algorithmic approach comes at the cost of a loss of generality with respect to the table-based approach. In fact, the algorithm is only specific to one topology and to one routing strategy on that topology. On the other hand, this mechanism turns out to be more efficient (area wise and performance wise) than table-based routing. However, routing tables introduce an
Dest.
Admissible Outputs
A
North, East
B
East
C
South, East
D
South, East
E
South, East
F
South, East
G
South
H
South
I
East
Fig. 3. (a). Routing Table before Compression
A B
A B
X
R 1
X
C D F G H I
C D E
F E G H I
R 2 Fig. 3. (b). Color based Clustering
Velammal College of Engineering and Technology, Madurai
Page 331
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
asserted, the address input is used to select the region to be configured. Configuring a
Dest.
Admissible Outputs
A
North, East
B
East
R1
South, East
R2
South
I
East
Fig. 3. (c). Compressed Routing Table
Since we are considering minimal routing, the destination node for a packet received from the west port will be in the right part of the NoC. Now, let us suppose to associate a color for each one of the five possible sets of admissible output directions. For instance, we mark with color Red (Green/Blue) all the destination nodes that, to be safely reached, needs that a packet in the west input port must be forwarded through the North (South/East) output port. Similarly, we mark with color Purple (Yellow) all the destination nodes that, to be safely reached, needs that a packet in the west input port must be forwarded through the North or East (South or East) output channels. The left side of Fig. 2b shows the set of destinations colored on the basis of the information stored in the routing table. The next step is to perform a color-based clustering of the destination nodes by means of rectangular regions. The right side of Fig. 2b shows an example of such a clustering. The basic idea of the compression technique is that it is no longer necessary to store the set of all the destinations, but only the set of Clusters, as shown in Fig. 2c. If a small loss in adaptivity is allowed, a cluster merging Procedure can be used to increase the compression ratio. For instance, in Fig. 13c, clusters R1 (yellow) and R2 (green) can be merged to form a single cluster R3 (yellow) whose destinations can be reached by using only the east output port. As reported i[55], in practical cases, using from four to eight clusters is enough to cover all practical traffic scenarios without any loss in adaptivity. Using from two to three clusters gives a small reduction in adaptivity, which translates to less than 3 percent performance degradation as compared to uncompressed routing tables. From the architectural viewpoint, the logic to manage the compressed routing table is very simple. Fig. 3 shows the block diagram of the architecture implementing the routing function that uses the compressed routing table. For a given destination, the block extracts from the compressed routing table the set of admissible outputs. The input port setup allows us to configure the router. When the setup is
Velammal College of Engineering and Technology, Madurai
region means storing attributes characterizing the region (top right, bottom left, and color) into the compressed routing table. The block In Region checks if a destination dst belongs to a region identified by its top-left corner (TL) and its bottom-right corner (BR). If this condition is satisfied, the output assumes the value of the color input, and output hit is set. 2.3. Table-Based Implementation A. Area Overhead Analysis To evaluate the overhead in silicon area due to the use of routing table with respect to routing logic, we designed in VHDL and synthesized by using Synopsys Design Compiler for a UMC 0:13-_m technology library the following blocks: Routing function: It is the block that gives the set of admissible outputs for the current node and a given destination. We designed three different blocks that implement the XY routing function, the Odd-Even routing function, and the routing-table-based routing function. The first two are pure combinatorial logic. The latter one, whose block diagram is depicted in Fig. 14, contains memory elements also. With regard to the routing-table-based routing block, we considered two different instances able to manage four regions (RT4) and eight regions (RT8), respectively. Input FIFO_5. They are the FIFO buffers at the input of each router. For a mesh-based network topology, there are five FIFO buffers in total. Crossbar: It is a general 5 _ 5 crossbar block that allows us to simultaneously route non conflicting packets. Arbiter: It is a general arbiter that manages the situation where several packets simultaneously want to use the same output. In this case, arbitration between these packets has to be performed. We used a round-robin policy. Table 3 reports the silicon area for each of the main blocks of a router for three different routers Implementing XY routing, Odd-Even routing, and table-based routing, respectively. With regard to the latter, two different implementations able to manage compressed routing tables with a budget of four (RT4) and eight (RT8) table entries have been considered. The cost overhead of a routing table implementation based on the proposed compression technique and architecture represents only a small fraction of the overall router cost.
Page 332
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
3. APSRA SIMULATED OUTPUT Simulation Results For South‐west Port
Simulation Results For East Port
4. CONCLUSION In this paper, we have proposed a topology –agnostic methodology to develop efficient APSRA for NoCs. Our methodology not only uses the information about the topology of communicating cores but also exploits information about the concurrency of communication transactions Routing algorithms generated by the APSRA method has even higher performance and adaptivity advantage over other deadlock-free routing algorithm. The APSRA methodology uses a heuristic to remove the minimum number of edges in the channel dependency graph to ensure deadlock-free routing. A straightforward table implementation may not be practically useful, since the table size grows with the number of nodes in the network. Therefore, we have also proposed a technique to compress tables such that the effect on the original routing algorithm is minimal.
[1] A. Ivanov and G.D. Micheli, “The Network-on-Chip Paradigm in Practice and Research,” IEEE Design and Test of Computers, vol. 22, no. 5, pp. 399-403, Sept.-Oct. 2005. [2] S. Kumar, A. Jantsch, J.-P. Soininen, M. Forsell, M. Millberg, J. Oberg, K. Tiensyrja, and A. Hemani, “A Network on Chip Architecture and Design Methodology,” Proc. IEEE CS Ann. Symp. VLSI, p. 117, 2002. [3] W.J. Dally and B. Towles, “Route Packets, Not Wires: On-Chip Interconnection Networks,” Proc. 38th Design Automation Conf. (DAC ’01), pp. 684-689, 2001. [4] F. Karim, A. Nguyen, and S. Dey, “An Interconnect Architecture for Networking Systems on Chips,” IEEE Micro, vol. 22, no. 5, pp. 36-45, Sept.-Oct. 2002. [5] P.P. Pande, C. Grecu, A. Ivanov, and R. Saleh, “Design of a Switch for Network on Chip Applications,” Proc. IEEE Int’l Symp. Circuits and Systems (ISCAS ’03), vol. 5, pp. 217-220, May 2003. [6] T. Bjerregaard and S. Mahadevan, “A Survey of Research and Practices of Network-on-Chip,” ACM Computing Surveys, vol. 38, no. 1, pp. 1-51, 2006. [7] P.P. Pande, C. Grecu, M. Jones, A. Ivanov, and R. Saleh, “Performance Evaluation and Design Trade-Offs for Networkon- Chip Interconnect Architectures,” IEEE Trans. Computers, vol. 54, no. 8, pp. 1025-1040, Aug. 2005. [8] D. Linder and J. Harden, “An Adaptive and FaultTolerant Wormhole Routing Strategy for k-Ary n-Cubes,” IEEE Trans. Computers, vol. 40, no. 1, pp. 2-12, Jan. 1991. [9] C.J. Glass and L.M. Ni, “The Turn Model for Adaptive Routing,” J. Assoc. for Computing Machinery, vol. 41, no. 5, pp. 874-902, Sept. 1994. [10] A.A. Chien and J.H. Kim, “Planar-Adaptive Routing: Low-Cost Adaptive Networks for Multiprocessors,” J. ACM, vol. 42, no. 1, pp. 91-123, Jan. 1995. [11] J. Upadhyay, V. Varavithya, and P. Mohapatra, “A Traffic- Balanced Adaptive Wormhole Routing Scheme for Two-Dimensional Meshes,” IEEE Trans. Computers, vol. 46, no. 2,pp. 190-197, Feb. 1997. [12] G.-M. Chiu, “The Odd-Even Turn Model for Adaptive Routing,” IEEE Trans. Parallel and Distributed Systems, vol. 11, no. 7, pp. 729-738, July 2000.
5. REFERENCES
Velammal College of Engineering and Technology, Madurai
Page 333
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Creating Actionable Knowledge within the Organization using Rough set computing Mr.R..Rameshkumar #1 , Dr. A. Arunagiri *2 , Dr.V .Khanaa #3 , Mr.C.Poornachandran
=4
# Information System Department,Bharat University, Selaiyur,Chennai – 600073, T.N,India. 1
[email protected] [email protected]
3
* EEET Department,Yanbu Industrial College, Yanbu Al- Sinaya , Yanbu, KSA. 2
[email protected]
= Department of Computer Science,Government Arts College, Nandanam,Chennai, India. 4
[email protected]
• Scalable from 2 to many experts.
Abstract For a management consultant in order to successfully assist an organization in creating new actionable knowledge (knowledge that is used to create value), the consultant must be aware of a knowledge dimension called Coalescent knowledge. The process for creating actionable knowledge in this dimension is a dialogue process. For example, product development is guided by several expert knowledge including critical process relationships which are dynamic, derived from experience and are often nonlinear. A new inductive learning intelligent technique called Rough set theory, which is to be used along with coalescent knowledge is described and proposed. The frame work of the classic management has been discussed and is modified using the Rough set theory method.
Keywords : Actionable Knowledge, knowledge, Rough Set , Critical Process
Coalescent
Morgan etal updated the knowledge creation theory documented by Nonaka and Takkeuchi, [2] and is expressed in Table 1: Changes to the Nonaka & Takkeuchi (NK) knowledge Creation theory Process/Mode
Velammal College of Engineering and Technology, Madurai
New Knowledge
Transition
Transition Form
Form Socialization
Tacit-to-Tacit
Tacit to Coalescent
Externalization
Tacit to Explicit
Coalescent to Explicit
Combination 1.0 Introduction For a management consultant to successfully assist an organization in creating new actionable knowledge (knowledge that is used to create value), the consultant must be aware of a new knowledge dimension called Coalescent knowledge (Morgan, Morabito, Merino,Reilly, 2001). Knowledge in the Coalescent dimension has the following attributes: • Created via a dialogue process involved group of experts • The knowledge is visible, expressible, shared and virtual • Can be private or public knowledge • Can be used to create a sustainable competitive advantage • Facilitates the opportunity for group of experts to act as if they have one mind (decide by rough set ) to accomplish organizational objectives
NK Knowledge
Internalization
Explicit to
Explicit to
Explicit
Coalescent
Explicit to Tacit
Coalescent to Tacit
Table 1 2.0 Rough Set Theory The concept of rough sets has been introduced by Zdzislaw Pawlak in the early 1980’s as a new concept of set used in many application and research. The Rough Set theory is an important part of soft computing. The methodology is concerned with the classificatory analysis of imprecise, uncertain or incomplete information or knowledge expressed in terms of data acquired from experience. It generates output as a list of certain and possible rules. It is
Page 334
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
formalized through the simple concepts of lower and upper approximation, which are in turn defined on the basis of set. The Rough set theory is enhancements of Neural networks, genetic algorithms, clustering ,support vector machines, regression models, etc.
Combination
Socialization
Internalization Expert No.1
Rough Set
Explicit
Coalescent . Tacit
3.0 Updated Knowledge Creation Process According to Morgan etal The first process in creating knowledge is the socialization. In this process, an individual shares his or her tacit knowledge with another individual or a group via some form of dialogue and/or observation [1], [2], In any dialogue and/or observation, each individual brings his or her tacit knowledge and references/links to explicit knowledge. For this analysis, it is assumed that the exchange of knowledge will be via dialogue. During the dialogue process, the first individual tries to define his/her tacit knowledge for the second person(s). This process requires the use of fields of interaction. The second person(s) then links their knowledge base to the knowledge being communicated. This is a repetitive action until the first and second person(s) agree on a common set of constructs, which defines the knowledge being communicated. The process has now created a shared virtual knowledge, which only exists between the individuals involved in the dialogue. This knowledge is shared and not codified, so it does not fall within the definition of tacit or explicit knowledge. The consultant can use the knowledge creation process flow that includes the interaction of people in the organization at a high level [1] to analyze the organization potential for creating knowledge in the coalescent dimension
.. 4.0 Refined Knowledge Creation Process Figure 2 shows the Refined knowledge creation process that includes the interaction of people or (experts) in the organization.
Velammal College of Engineering and Technology, Madurai
. . .
Expert No. N
Socialization
Externalization
Internalization
Refined Knowledge Creation Process Flow Figure 2
In the refined knowledge creation process[2], the creation of new tacit knowledge does not require the inclusion of explicit knowledge. As Coalescent knowledge matures it can be externalized via codification to become explicit knowledge which may be private or public. If there is only one Expert in the tacit dimension, then that expert would have both Tacit and Coalescent knowledge. If it requires to internalize some explicit knowledge, then the explicit knowledge would be converted to Coalescent knowledge. The Coalescent knowledge is shared between the expert and the creators of the explicit knowledge. Although the creators are not actively participating in the dialogue, the expert doing the internalization assigns them a virtual role. Knowledge that is shared by two or more expert’s can be considered communal knowledge. The formation of Coalescent knowledge from tacit dimension will be certain when the dimension includes minimum number of experts. Whereas the uncertainty arises in the knowledge creation process within the organization when the tacit dimension involves many number of experts. Since the Rough set deals with the uncertainties, it can be effectively used to refine the Coalescent dimension and to make the actionable knowledge more certain. The Rough set can make use of different experts tacit knowledge derived from similar organization in order to refine the uncertainty problems.
5.0 Classic Management Process The classic management process is comprised of four subprocesses: Planning, Organizing, Leading, and Controlling (PLOC)[5]. The controlling and planning sub-processes are connected through a feedback loop to ensure that the objectives of the plan are being met. The controlling function compares performance to the standards set in the planning process. If deviations are detected, the information is feed back to the planning process for changes in the plan that will cause the standards to be met. During the controlling and/or planning processes, an
Page 335
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
evaluation should be performed to determine why there was a deviation from standards.
C l a s s i c M a n a g e m e n t P r o c e s s
Planning
Leading
Organizing
Controlling
Feedback Loop (P L O C) Figure 3
6.0 Modified Classic Management process by Rough Set theory By adopting the critical reflective practice principals as defined by Van Aswegen (1998)[6] to an open group dialogue format, management consultant could make use of the Coalescent knowledge derived from the tacit dimension. Based on the formation of the uncertainty problems, the Coalescent knowledge can be refined by implementing the rough set concept in order to eliminate the uncertainty and to minimize the deviation from the standards.
Knowledge Creation Theory, Paper presented at the First International Conference on Knowledge Management, co-sponsored by Academy of Management, ISEOR, University [2] Nonaka, Ikujiro and Takeuchi, 1995, The Knowledge Creating Company, Oxford University Press, New York, NY. [3] Pawlak,Zdzislaw (1991). Rough sets: Theoretical Aspects of Reasoning About Data. Dordrecht: Kluwer Academic Publishing. ISBN 0-7923-1472-7. [4] Morabito, Joseph, Sack, Ira, and Bhate, Anilkumar, 1999, Organization Modeling – Innovate Architectures for the 21st Century. Prentice-Hall, Inc., Upper Saddle River, NJ. [5] Freeman, R.E., and Stoner, J.A.F., 1989, Management. Prentice Hall, Englewood Cliffs, NJ, 4th Edition. [6] Van Aswegen, Elsie Johanna, 1998, Critical Reflective Practice: Conceptual Exploration and Model construction (Emancipatory Learning, Naturalistic Inquiry), Dissertation, University of South Africa, South Africa [7] Komorowski, J. et al., 1999. Rough sets: A tutorial. In S.K. Pal and A. Skowron (eds.), Rough Fuzzy Hybridization: A New Trend in Decision-Making. Springer, pp. 3-98.
Modified Classic Management Process
Critical Reflective Practice Figure 4
7.0 Conclusion : This paper describes the concept of tacit knowledge dimension and the formation of Coalescent knowledge from tacit dimension. As the tacit dimension increases, the uncertainty problem is eliminated through Rough set method. The concept of Classical Management Process is explained and is modified by the refined Coalescent knowledge through Rough set in order to minimize the deviation between the reference and the actual outcome. References: [1] Morgan, Morabito, Merino, and Reilly, 2001, Defining Coalescent Knowledge: A Revision of
Velammal College of Engineering and Technology, Madurai
Page 336
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
COMPVAL – A system to mitigate SQLIA S. Fouzul Hidhaya1 , Dr. Angelina Geetha2 Department of Computer Science and Engineering, B.S. Abdur Rahman University, Vandalur, Chennai -600048, Tamil Nadu, India. 1
[email protected] 2 [email protected]
Abstract - In this Internet Era, Web Applications are a boon to the business to customer data exchange. The web applications are directly linked to the database and so the web applications are used as a interface to carry out the SQL Injection Attacks on the databases. This paper discusses a novel method to mitigate the SQL Injection Attacks. This system has two processes running in parallel. The actual process of retrieving the data from the database is done in parallel with the process of syntactically analyzing the query statement. The intended output and the actual output of the data are compared and validated and the corresponding HTTP response is framed. This system provides a double check for the SQL injection attack and from them the experimental results it has been proved effective and has imposed a average overhead of 15 milliseconds when tested on the web applications. Keywords - SQL Injection, Vulnerabilities, Web application security, Security threats, Web application vulnerability.
I. INTRODUCTION Business to customer (B2C) data exchanges like the online shopping, banking. gaming, ticket booking, travel booking etc. has been made feasible by the web applications. Web application work with a back-end database to store or retrieve data. These data are accepted from the user or retrieved from the database to be given to the user dynamically by embedding them in a HTML code. The input supplied by the user could be benign or malicious. If the data is malicious, undesired data could be extracted from the database. This type of attack is called the SQL injection attack (SQLIA). An example of SQL Injection is given here. A SQL injection attack occurs when an attacker causes the web application to generate SQL queries that are functionally different from what the user interface programmer intended. For example the application contains the following code, Query = “SELECT * FROM members WHERE login ='”+ request.getParameter(“login”)+” ‘ And password=’ “ + request.getParameter(“password”)+” ‘ “; In this code the web application retrieves user inputs from login and password, and concatenates these two user inputs into the query. This is used to authenticate the user. If the attacker enters admin into the login filed
Velammal College of Engineering and Technology, Madurai
and xyz’ OR ‘1=1 into the password field. Now the query string will be, SELECT * from members WHERE name=’admin’ AND password=’xyz’ OR ‘1=1’ In this case, the password field, which should have only one string, is replaced with sub strings, which will finally lead to the retrieval of the data of the administrator which is undesirable. An effective and easy method, for detecting and protecting existing web applications from SQL Injection Attack, is need of the moment, which will help the organizations to be secured against these attacks. Detection or prevention of SQLIA is a topic of active research in industry and academia. The most commonly injection mechanisms [13] are 1. Injection through user input. 2. Injection through cookies 3. Injection through server variables 4. Second order injection The proposed attack works on injection through user input. The rest of the paper is organized as follows. Section 2 discusses the related work carried out in this area, Section 3 illustrates the system design. Section 4 evaluates the effectiveness of the proposed approach and Section 5 concludes the paper. II. RELATED WORK Researchers have proposed a wide range of techniques to address the problem of SQL Injection. These techniques range from development best practices to fully automated frameworks for detecting and preventing SQLIAs[12]. One technique being used to mitigate SQLIA is the defensive coding practice. Since the root cause for the SQL injection vulnerabilities is insufficient input validation, the straight forward solution for eliminating these vulnerabilities is to apply suitable defensive coding practices, like Input type checking, encoding of inputs, pattern matching and identification of all input sources. But defensive coding is prone to human error and is not as rigorously and completely applied as automated techniques. While most developers do make an effort to code safety, it is difficult to apply defensive coding practices correctly to all sources of input.
Page 337
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
functionalities of the application, and stores them in the repository named the ‘Prototype document’. In the normal phase(Section 3.2), the HTTP request is interpreted by the ‘HTTP Interceptor’ and then sent to the web application which retrieves the data from the database and with this data the web application frames the HTTP response. In the validator phase (Section 3.3), the HTTP Interceptor sends the Intercepted HTTP request to the ‘Query Interceptor’. Here the query is extracted and send to the ‘parse tree generator’ and the ‘syntax analyzer’. The parse tree generator generates the parse tree for the query and sends to the syntax analyzer. Syntax analyzer compares the query from the query interceptor and the query database and extracts the user input and applies this user input to the comparator phase (Section 3.4). If the request is marked as benign the HTTP response is immediately send to the user. If the request is marked as malicious by the syntax analyzer then the comparator compares the intended output for the query and the actual output that is retrieved from the database. If they are the same then requested data is send in the HTTP response. Else a ‘data unavailable response’ is send in the HTTP response.
Researchers have proposed a range of techniques to assist developers and compensate for the shortcomings in the application of defensive coding. The black box testing methodology used in WAVES[6], which uses a web crawler to identify all points in web application that can be used to inject SQLIAs. It uses machine learning approaches to guide its testing. Static code checkers like the JDBCchecker[4] is a techniques for statically checking the type correctness of dynamically generated SQL queries. This will be able to detect only one type of SQL vulnerability caused by improper type checking of input. Combined static and dynamic analysis like the AMNESIA [5] is a model based technique that combines static analysis and runtime monitoring. AMNESIA interprets all queries before they are sent to the database and checks each query against the statically built models. SQLGuard [3] and SQLCheck [10] also check queries at runtime to see if they conform to a model of expected queries. Taint based approaches like the WebSSARI[8]detects input-validation-related errors using information flow analysis. In this approach, static analysis is used to check taint flows against preconditions for sensitive functions. Livshits and Lam [7] use information flow technique to detect when tainted input has been used to construct a SQL query. Security gateway [9] is a proxy filtering system that enforces input validation rules on the data flowing to a web application. SQLRand [2] is an approach based on instruction-set randomization. It allows developers to create queries using randomized instruction instead of normal SQL keywords. A proxy-filter intercepts queries to the database and derandomizes the keywords. III. I.
Web application
Prototype document
Figure 2: Query Extractor
SYSTEM DESIGN
System Architecture
Figure 1 : COMPVAL System Architecture
The main system architecture of COMPVAL is illustrated in Figure 1. This system could be divided into four phases, Query extractor, normal phase, validator phase and the comparator phase. The query extractor phase extracts the query from the web application code to cover all the
Velammal College of Engineering and Technology, Madurai
Query Extractor
II.
Normal Phase The normal phase is similar to the general operation of the web application without any SQLIA detecting mechanisms. The HTTP interceptor intercepts the HTTP request sent by the user and sends the request to the web application. The web application then sends the request to the database and retrieves the data. This data is then framed as a HTTP response and instead of sending it to the user the response here is send to the comparator for evaluation. C. Validator phase In this phase, the intercepted HTTP request is send to query interceptor and the query is extracted from the HTML and send to the parse tree generator and the syntax analyzer. The syntax analyzer uses XSL’s pattern matching. The syntax analyzer works as follows: 1. Accepts the query from the query interceptor and tokenizes the query.
Page 338
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2. Transforms the query into XML format. 3. Uses the XSL’s pattern matching algorithm to find the prototype model corresponding to received query. 4. From the prototype query it identifies the user input data. 5. Receives the parse tree from the parse tree generator and fits in the nonleaf nodes. 6. Validates the resulting parse tree using the validation algorithm and identifies if the query is benign or malicious. 7. Sends the prototype query and the result to the comparator. 1) XSL’s Pattern Matching : When the query is submitted by the query interceptor, the query is first analysed and tokenized as elements. The prototype document contains the query pertained to that particular application and it also has the intended output for each query model. For example, as explained in section 1, the input query is, SELECT * FROM members WHERE login=’admin’ AND password=’XYZ’ OR ‘1=1’ When this query is received this is converted into XML format using a XML schema. The resulting XML would be, <SELECT> <*> <FROM> <members> <WHERE login= ’admin’> <AND password= XYZ> <OR 1=1> </OR> </AND> </WHERE> </members> </FROM> </*> </SELECT> Using the pattern matching the elements is searched so that the nested elements is similar to query tokens. The corresponding matching XML mapping is, <SELECT> <identifier> <FROM> <identifier> <WHERE id_list= ’userip’> <AND id_list=’userip’> </AND> </WHERE>
Velammal College of Engineering and Technology, Madurai
</identifier> </FROM> </identifier> </SELECT> When the match is found, the corresponding prototype query would be, SELECT identifier FROM identifier WHERE identifier op ‘userip’ AND identifier op ‘userip’ which will be used to identify the user input data. This search is less time consuming because the search is based on text and string comparison. The time complexity is O(n). This helps in increasing the effectiveness of the program and reduces the latency time. 2) Syntax analyzer: The syntax analyzer uses the prototype query model to recognize the user inputs and the this inputs are inserted in the parse tree which is generated by the parse tree generator. Now the syntax analyzer has the parse tree and the user inputs. Figure3 represents the parse tree for the benign query in section I. Every user input string can find a non-leaf node in the parse tree, such that its sub-tree leaf nodes comprise the entire user input string. An example application of the parse tree generator is given in figure 4. This algorithm for identifying malicious output is used in SQLProb[14].This system uses the validation algorithm to validate if the request is benign or malicious. The password field is parsed into the set Uni=1 (leaf(ui)) with five leaf nodes: XYZ, OR, 1, = and 1. The analyzer algorithm is given in Table 1. Next, we do depth-first-search from these five leaf nodes. The traversed paths intersect at a non-leaf node, SQLExpression. Finally, we do breath-first-search from SQLExpression to reach all the leaf nodes of the parse tree, which is a superset of Uni=1(leaf(ui)) implying that the input string ui is malicious. The algorithm described above takes quadratic time, because step 2 and step 3 take time of n × h, where n is the number of leaf nodes parsed by ui, and h is the average number of layers from leaf nodes to nl node in the parse tree. In addition, step 4 takes time complexity for a breathfirstsearch is O(n2). Therefore, the overall time complexity is O(n2). 3) Comparator Phase: The validator phase sends the result, indicating if the request is benign or malicious depending on the analyses it carried out, to the comparator. The comparator now has the results retrieved from the database and if the analyzer had certified that the request is benign without any delay the requested data is immediately send to the user. If the request is analyzed to be malicious, the comparator extracts the intended output of the model query from the Prototype document and compares it with the actual output from the application encoded in the HTTP response is database. If the result is the same, the requested data is send to the user as HTTP response. Else HTTP response with a ‘Data not available’ reply is sent to the user.
Page 339
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
SELECT * From members WHERE login=’admin’ and password=’xyz’
SELECT
select_list
identifier
FROM
Table_list
WHERE
Where_cond
identifier
SQLExpression
SQLAndExpression *
members SQLAndExpression
SQLAndExpressio
selectItem
ID
factor
Operator
login
selectItem
ID
ID
=
factor
Operator
=
password
admin
ID
XYZ
Figure 3: Parse tree for benign query
SELECT * From members WHERE login=’admin’ and password=’xyz’
SELECT
select_list
identifier
*
FROM
Table_list
identifier
Where_cond
SQLExpression
members
SQLAndExpression
SQLAndExpression
SQLAndExpressio
selectItem factor
ID
WHERE
Operator
ID
SQLAndExpressio
factor
factor
Operator
ID
ID
ID
Operator
1 login
admin
factor
selectItem
password
=
ID
1
=
=
XYZ
OR
Figure 4: Parse tree for malicious query
Velammal College of Engineering and Technology, Madurai
Page 340
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
IV. A.
EVALUATION
Experimental Setup
The test suite used to attack is AMNESIA attack test suite [5]. This suite has a list of string patterns which are both benign and malicious. Using these attacks on the web application COMPAL achieved 100% detection rate and was able to prevent the malicious string patterns from extracting data from the database. The performance metric taken into consideration is the response time per web request. For every web request, the delay varied depending upon the query and its alignment. The response time for each web application used is shown in figure 5. TABLE I VALIDATION ALGORITHM
if Uni=1(leaf(ui)) Umk=1leaf(node)k) then Return True; else Return False; end end
0.16 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0
Actual COMPVAL
st o
re
ir ok
pd
Bo
Em
Po
rt a
l
R e s p o n s e ti m e i n S e c
6. 7. 8.
V. CONCLUSION COMPVAL, as indicated in section 1, is a novel system used to detect SQLIA’s efficiently. The proposed system does not make any changes to the source code of the application and so the source code need not be known to COMPVAL. The proposed system the detection and prevention of SQLIA is fully automated. Prototype of COMPVAL was designed to measure the performance factor and it worked successfully with 100% detection rate. In future work, the focus will be on benchmarking the system with standard algorithm and studying of alternate techniques for the detection and prevention of SQLIA using dynamic analysis. REFERENCES
Data:Parse Tree Tree(qi) and the set of user input UI(qi) Result:True if the query is an SQLIA, or False if otherwise 1. for every user input UIi,j in UI(qi) 2. do depth-first-search upward from every leaf node leaf(ui) parsed from UIi,j , according to SQL grammar G; 3. Searching stops when all the searching path intersect at a non-leaf node nl node; 4. do breath-first-search downward from nl node until reaching all m leaf nodes leaf(node)k; 5.
Figure 5: The comparative response time
Web Applications
Velammal College of Engineering and Technology, Madurai
1. C. Anley, “Advanced SQL injection In SQL Server Applications”, White paper, Next Generation Security Software Ltd., 2002. 2. S.W. Boyd and A.D. Keromytis, “SQLrand:Preventing SQL Injection Attacks”, Proc. ACNS’ 04, pp. 292-302, June 2004. 3. Gregory Buehrer, Bruce W. Weide and Paolo A. G. Sivilotti, “Using Parse Tree Validation to Prevent SQL Injection Attacks”, Proc. International Workshop on Software Engineering and Middleware, 2005. 4. C.Gould, Z.Su and P.Devanbu, JDBC Checker: A Static Analysis Tool for SQL/JDBC Application, Proc. International Conference on Software Engineering ‘04, pp.697-698, 2004. 5. W. G. Halfond and A. Orso, “AMNESIA: Analysis and Monitoring for NEutralizing SQL-Injection Attacks”, Proc. ACM International Conference on Automated Software Engineering ‘05, November 2005. 6. Y.Huang, F. Huang, T.Lin and C.Tsai, “Web Application Security Assessment by Fault Injection and Behavior Monitoring”, Proc. International World Wide Web Conference ‘03, May 2003. 7. V.B. Livshits and M.S. Lam, “Finding Security Errors in Java Programs with Static Analysis”, Proc. Usenix Security Symposium ‘05, pp. 271-286, August 2005. 8. Y.Huang, F.Yu, C. Hang,C.H. Tsai, D.T.Lee and S.Y.Kuo, “Securing Web Application Code by Static Analysis and Runtime Protection”, Proc. International World Wide Web Conference ‘04, May 2004. 9. D.Scott and R.Sharps, “Abstracting Application-level Web Security”, Proc. International Conference on the World Wide Web ‘02,pp. 396-407, 2002. 10. Zhendong Su and Gary Wassermann, “The Essence of Command Injection Attacks in Web Applications”, Proc. ACM SIGPLANSIGACT Symposium on Principles of Programming Languages ‘06, January 2006. 11. Burp proxy. [Online]. Available: http://www.portswinger.net/proxy. 12. JJTree. [online]. Available: http://javacc.dev.java.net/doc/JJTree.html. 13. W. Halfond, J. Vigeas and A.Orso, “A Classification of SQL Injection Attacks and Counter Measures”, Proc. International Symposium on Secure Software Engineering ‘06, March 2006. 14. Anyi Liu, Yi Yuan, Duminda wijesekera and Angelos Stavrou, “SQLProb: a Proxy-based Architecture Towards Preventing SQL Injection Attacks”, Proc. ACM symposium on Applied Computing ‘09, pp. 2054-2061, 2009. 15. W. Halfond, A. Orso and P. Manolios, “Using Positive Tainting and Syntax-Aware Evaluation to Counter SQL Injection Attacks”, Proc. ACM SIGSOFT Symposium on the Foundations of Software Engineering ‘06, November 2006.
Page 341
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
16. Adam Kieyzun, Philip J. Guo, Kartick Jayaraman and Micheal D. Emst, “Automatic Creation of SQL Injection and Cross-Site Scripting Attacks”, Proc. IEEE International Conference on Software Engineering ‘09, pp. 199-209, 2009. 17. Konstantinos Kemalis and Theodores Tzouramanis, “SQLIDS: A Specification-Based Approach for SQL-Injection Detection”, Proc. ACM symposium on Applied computing ‘08, pp. 2153-2158,2008. 18. Xiang Fu and Kai Quian, “SAFELI: SQL Injection Scanner using Symbolic Execution”, Proc. Workshop on Testing, Analysis and Verification of Web Services and Applications ‘08, pp. 34-39, 2008. 19. M.Johns and C. Beyerlein, “SMask: Preventing Injection Attacks in Web Application by Approximating Automatic Data/Code Separation”, Proc. ACM Symposium on Applied Computing ‘07, March 2007.
Velammal College of Engineering and Technology, Madurai
Page 342
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Integrating the Static and Dynamic Processes in Software Development V. Hepsiba Mabel#1, K. Alagarsamy*2, Justus S$3 #
Fatima College, Maryland Madurai 1
[email protected]
*
Madurai Kamaraj University Madurai
2
[email protected]
$
Velammal College of Engineering & Technology Madurai 3
[email protected]
Abstract— Software development is caught up in bunch of processes that involve project, product and people. Organizations those are into the development of software products and providing software solutions are bound by sets of processes that governs the activities of the project as well of the organizations. CMM and other standards serve as very important guidelines to companies that are in the software development. The entire processes are based on the IDEAL model, and those processes are defined static processes. Also, recent researches in Knowledge management show that knowledge based processes are deemed dynamic in nature and help the processes to improve over period of times. However, these processes are required to be integrated into one group or framework, where the practitioners will find their entire set of processes derived and defined in the CMM levels might be easy to improvement. In this work, we have integrated the static and dynamic process sets into one framework and have partially implemented them in companies who are interested in practicing the framework. Keywords— Software Processes, Process Framework, CMM.
I. INTRODUCTION The static and dynamic processes involved in the development of software products are of significant concern in research and in practice. Since most of the activities involved in software development are loosely coupled they need to be integrated to form a single framework. Since the beginning of software development and software engineering in the early 1960s, research on the activities and the processes governing the design and development of software products are active; and to till date they are active. However, it is studied from the literature [3], [7] [12] [23], [26], [30], and case studies only a limited number of publications are available for the processes that governs in software development, when compared to other areas of research. Also there is a need for a rigorous study on the process sets and key process areas that forms the capsule of
Velammal College of Engineering and Technology, Madurai
the process standards and maturity models. This notion of process and process set is a dynamic entity, because the process of software development is world wide and it crosses all continental and cultural backgrounds. Hence the process sets defined by [2] [22], are not standardized internationally. However, we need to bring the KPAs and process sets, both defined and tailor-made, into one single framework. This work deals in detail the formation of the single framework. Organizations involved in software development are concerned with delivery of quality software products within the specified timeline. Development of a software product is a time-based, distinct activity which is termed as project. A project is governed by a set of defined and tailor-made processes that controls the quality end-product and ensures that the product is delivered ontime and without overruns on the budget. To enable all the implicit processes involved in software development, project leaders and managers need to have a decision supportive intelligent system that falls into a framework for improving their software developmental processes. Motivation of the Work When in the early years of software development, there was only a limited steps and algorithms governing the developmental processes. As years and decades passed by the software development industry faced with numerous processes, processes that are anticipated, expected and unexpected, were evolved. These processes defined and regulated were not adopted by the industry for its lack in flexibility. Hence the organizations attempted to tailor-make the processes according to their organization’s need and projects’ requirements. While most of the processes defined and standardized by Software Engineering Institute [26], [30], were acceptable by software developing companies, many were deemed to be redefined and refined as years passed by. This is commended
Page 343
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
by the global, local markets, competitors, governments’ policies and other factors that arise internally and externally. Out of these changes, the software development companies and the research institutes were able to identify the two dimensions of the entire process sets. They are classified based on the nature of the processes as Static Processes and Dynamic Processes [2]. Both these process sets are formally validated by international research institutes. However the static and dynamic process sets stand alone in the practicing platform. Organizations prefer the static processes than dynamic one for their simplicity and ease of implementation and practice [23]. Although it is encouraging and result reaping, dynamic processes are recommended. This is because, dynamic processes yields good results, and provides and wider scope for process improvement. Hence we attempt in this work to integrate these two process sets and form a single framework of processes, that will help the organizations in building up their process base and leaves better opportunity for process improvement. Related Works As pointed out earlier, research works on software development processes have found a prominent position. Software Engineering Institutes [26], [30] is the major role player in proposing standards like CMM, CMMI. CMM family of standards speaks on the refinement of processes and the level of maturity with which the processes have improved. ISO9000 family is a very specialized set of standards which mainly focuses on the processes in the management of software processes and their improvement. SPICE, ISO/IEC12207, BOOTSTRAP and Rational Unified Process (RUP) are other similar standards which centers the processes and the collection of processes [12]. But organizations follow these standards just as a roadmap for their journey in the process improvement programs. Research literatures and papers on processes sets and areas cover a wide variety of topics with regard to integration and improvement [1], [2], [7], [11], [14], [24], [31]. Similar works on processes sets are still in research and study.
process becomes better defined and more consistently implemented throughout the organization. Software process capability describes the range of expected results that can be achieved by following a software process. The software process capability of an organization provides one means of predicting the most likely outcomes to be expected from the next software project the organization undertakes. Software process performance represents the actual results achieved by following a software process. Thus, software process performance focuses on the results achieved, while software process capability focuses on results expected. Software process maturity is the extent to which a specific process is explicitly defined, managed, measured, controlled, and effective. Maturity implies a potential for growth in capability and indicates both the richness of an organization's software process and the consistency with which it is applied in projects throughout the organization. As a software organization gains in software process maturity, it institutionalizes its software process via policies, standards, and organizational structures. Institutionalization entails building an infrastructure and a corporate culture that supports the methods, practices, and procedures.
A. Basics of Software Processes A software process can be defined as a set of activities, methods, practices, and transformations that people use to develop and maintain software and the associated products [23] (e.g., project plans, design documents, code, test cases, and user manuals). As an organization matures, the software
Processes and Maturity Levels The formula for success of a software development organization lies in the continual improvement of its processes and leveraging its effectiveness for business excellence [18]. As discussed previously, the learning in the organization builds knowledge and warehouses which are the help guides for process improvement. Initiating an improvement program in the organization starts from the individual; that is when the individual realizes his improvement, will he appreciate the need for the organization’s improvement. Organization’s improvement is primarily based on the improvement in the processes. Here we consider the improvement of software process. Software process is defined as a set of activities that begins with the identification of a need and concludes with the retirement of a product that satisfies the need; or more completely, as a set of activities, methods, practices, and transformations that people use to develop and maintain software and its associated products [15]. To improve the software development process, it is necessary to evaluate the success of the project and avoid repeating problems in the future. The CMM addresses these latter, less well-understood, issues in an effort to improve the software development process [10]. Each represents an improvement in the software process. An organization’s software capability can be improved by advancing through these five stages or levels. The CMM helps organizations to select improvement strategies based on current process maturity status and to identify critical issues in quality and process improvement [26]. To help small organizations, possessing only limited resources, to reach the CMM level 3 and improve certain helpful tips are formulated
Velammal College of Engineering and Technology, Madurai
Page 344
II. PROCESSES IN SOFTWARE DEVELOPMENT The process of developing software has also been under research and practice since the evolution of software. The processes and activities involved in software development is systematic and unique endeavor that use sound engineering principles in order to obtain reliable, quality software product in an economically managed timeline.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
based on the discussions given by Kathleen [15]. It is observed from surveys, the SMBs in Chennai, India, are content with what they are performing. But when they complete for global projects they fail because they have never opted for CMM standards in any of the developmental strategies of business process. If at all they attain CMM level 3, they are content in that level because of lack of action plans and further improvement initiatives. The levels of process maturity for improving organizations is given in table 1. Organizations planning for an improvement endeavor will have to identify the business and developmental process that addresses organization goals, thus identifying the key-process areas. The accepted process sets has to be tailored according
to the requirement of the project or for the implementation of activities that support specific needs, resulting in a process protocol. Then lies the real process improvement that is goal oriented and knowledge driven, leading to CMM improvement level 3. Beyond this the organization has to extend its process set by refining them, ensuing in process reusability. When the componential process sets are combined and externalized they result in the regeneration of new process sets called hybrid processes, thus leading the organization to the highest level of process optimization. All of these states of maturity are driven by the activities of knowledge management.
TABLE I LEVELS OF PROCESS MATURITY
Process Maturity Levels Process Initiation Process Development Process Improvement Process Extension Hybrid Processes
Initial State
Actions
Implementation
Static
Address Goals
Process Set
Develop Activities
Dynamic Process Improved Process Set Componential Dynamic Processes
Use dedicated resources Refine Processes and Sets Plan and apply
Engineering Practices Activities for Specific needs Use Tools for improvement Expand to other projects Derive new process sets
Result Identified KPAs Process Set CMM level improvement Process reusability Process regeneration and improvement
Static Processes The processes involved in software development are deemed for practice and improvement. Still some of the processes remain unchanged and they are termed as Static Processes. These processes mainly deal with the developmental activity of the software and the processes that govern them. There may be many factors that influence them, but whatever be the external factors these processes have to remain static throughout the software development life cycle [4]. Attempt to
change these static processes is not advised, because that may lead to unwarranted, irreversible damage to the project or product or to the organization. A process set identified in the second process maturity level is often static. This set will be the basis for all the other derived and hybrid processes that may arise during the future endeavors of the organization [8]. While this static process set is static and cannot be changed, they are deemed for improvement during maturity level 3 and the following improvement programs. Table 2 lists the set of static processes. The five set of processes identified and defined are termed as static. They are bound within the scope of the project and are subject to be refined but not changed. This set of processes is defined based on the IDEAL model [16], [23]. The IDEAL model, framed for SPI program, has led to the derivation of a knowledge driven model which focuses on the acquisition of explicit knowledge through externalization and its conversion to tacit knowledge through internalization and continuous learning [10], [17]. Though many KM tools comprise, both smaller, specific applications, and enterprise level tools, the urge to provide a compelling learning environment should be felt [19].
Velammal College of Engineering and Technology, Madurai
Page 345
III. A SINGLE PROCESS FRAMEWORK The discussion in the previous chapter on the processes and their integration with the maturity levels in software development would well be a basis for the integration of the two sets to activities, static and dynamic. The proposal and the practice of static activities by software organizations improve in their maturity level. But the proposal, practice and the promotion of dynamic activities helps the organization to improve software processes based on knowledge gained over the years of experience and the concept of sharing them [2]. This chapter deals in detail the two sets of processes and integrating them.
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE II STATIC PROCESS SET
Phases of the Process Set Development of Software Maintenance of the end product Software Processes Production Processes Management Processes
Static Processes Design and Coding Processes Finish the product on time and within budget Finished product maintenance process Expires at the end of the scope Stick on to the processes defined for development and maintenance Document preparation processes Populate Repositories Plan and Control Use Decision support tools
Hence a need-based learning process is introduced in the establishing phase of KDM through packaging, which resolves the setback of knowledge sharing. In the Acting phase, the derived knowledge is executed through effective planning, based on the earlier experiences and lessons learnt from previous similar projects. B. Dynamic Processes In continuance with the definition of static processes, there arises certain processes as a result of improvement or refinement. These newly identified processes are dynamic processes as they are subject to refinement then and there whenever required and demanded by the project or by the organization. This set of processes involves and demands the required component called interaction and sharing. Exchange of intellectual property and experience gained over the years are the major driving catalysts for the dynamic process set. Transfer of knowledge and experience through tacit or explicit or even implicit will take place in this domain of processes. This set of activities is derived based on the knowledge driven model proposed in [2]. More often these set of activities are practiced and promoted in the fourth and fifth process maturity level and keeps going for betterment and refinement. The dynamic process set consists of the following processes. Table 2 mentions them. Most or even all of the dynamic activities, mentioned in the table, are specific and centered on knowledge and its associated components. TABLE III DYNAMIC PROCESS SET
Phases of the Process Set Development of Software Maintenance of the end product Software Processes
Dynamic Processes Interaction and Sharing of resources Exchangeability of gained experience Document revision & refinement Interaction within the
Velammal College of Engineering and Technology, Madurai
Production Processes
Management Processes
development team Improvement and Generation of new processes Codify data to repositories Generate feedback to the development team Interaction with project leaders Share expertise knowledge Generate hybrid processes
While static activities are derived from CMMI or other existing conventional standards, dynamic activities are based on interactions of knowledge workers and knowledge users, and knowledge activities like codification and populating the knowledge repositories, which in later phases help the management in institutionalizing and implementing the newly found hybrid knowledge for planning and controlling the process improvement program [1], [30]. The dynamic activities are always driven by the knowledge generated by the experiences and lessons learnt from earlier and similar projects. The knowledge based components in the software development organizations lay spread in their existence and there seems to be a gap among them in the organizational chart. But functionally they stay integrated in their operational platform. Integrating the components across the nature of data repositories, cultural, continental and psychological differences of the knowledge workers and users remains a great challenge of the SPI program [31], [32]. The process activities, both static and dynamic, serve as a link in integrating the components according to the procedures that organizations adopt for the software process improvement. It is inferred that the dynamic knowledge driven activities have a significant impact on the software processes, identified for improvement. This can be verified by measuring the functionality of the knowledge components that endure integration. In a wider spectrum, there are 22 process areas [10], where all of them can be designated to be the set of activities either as static or dynamic. Conceptually, the degree of impact of the dynamic components on these processes is proportional to the estimated functional effort of the integrated components. More precisely, higher the degree of interaction among the components more the dynamic processes (activities) is involved in action, causing the entire set of processes to improve. Operation and implementation of this knowledge centered SPI program is a major challenge development companies confront with in practice. C. Integrating the Processes into Framework An attempt to integrate the two divisions of Processes is an important task and is a careful endeavor which requires a better understanding of the entire process sets. 1) Development of Software: The static processes associated with this phase is design and coding and to finish the software product in time and within the budget. While design and coding are matters concerning technical talk, the later part
Page 346
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
finishing the product within time and budget is the great challenge that is prevailing since the start of the software development. Due to these two serious issues software projects are subject to failure or partial completion [13], [31]. Finishing a successful project is the requirement and demand from the project mangers. One of the measures of project success is completing the project within timeline and without budget overrun. This problem can be solved only when the dynamic set of activities; Interaction, Sharing of resources Exchanging the gained experience The design tools and techniques, the coding methodologies and documents are the resources that are subject to be shared among the team members and also among project groups. The exchange of knowledge and experiences gained over the years by senior managers and tech leads should be shared among the team members. This experience will help the juniors and the lower-level work groups to finish their development and coding in time and also within the budget. A study in software companies show that this kind of sharing through knowledge management systems help the organization improve the productivity of the project teams and of the individuals [19], [20]. 2) Software Processes: Whenever a software project has been initiated, the project planning phase does several important things. Among them a few are defining the processes and choosing appropriate process models The processes that are associated with the development of software are identified and defined at the earlier stages of software development life cycle. Then choosing an appropriate model for the software development is an important task the project mangers are confronting with. In practice we could see that out of the 13 process models available in literature [10], the practitioners often choose just two or three and customize them according to the requirement of the nature of the project. This activity of choosing the process models is termed as static activity because this is usually just declaring and defining the process models which are usually a common and a routine practice. However this alone is not sufficient for the software product to be successful. One has to be highly dynamic and active while in the process of software development. Hence interaction among the team members and across the levels of hierarchy of people involved in the same project is required. This creates newer ideas and innovative concepts of software development. The interaction among the team members and among the members involved in the project leads to the development of newer processes, which we deem as hybrid processes [28]. The existing and the hybrid processes are subject to improvement as time moves on and the interaction increases. 3) Production Processes: Production in software development is the number of lines of code generated during the construction of code and also the amount of work done in a given time frame. This includes finished product along with
Velammal College of Engineering and Technology, Madurai
the necessary designs and the document files. Documentation is an embedded process in the process of software development. Every project under progression should posses a knowledge repository. This is an additional document and the forth coming knowledge management system for the project and for the organization. The process of populating the repositories is the requirement for any project that considers itself serious and which contributes to the future of the organization. These two processes are termed as static because they are required components for the project. The repositories are populated with information which is then analyzed, processed and codified as knowledge units in the knowledge management systems (KMS). The codification of the information to be represented as knowledge unit is a dynamic process, because information that is generated during the project progress has to be identified and should be sent for codification in the KMS. This generated knowledge units in the KMS will help the system to generate feedback to the developers tuning them to produce quality softwares and documentation processes. IV. FINDINGS FROM IMPLEMENTING THE FRAMEWORK This section shows the findings from the experiments that are conducted as implementation part of the process framework. Critical Success Factors Table 4 shows the list of CSFs cited in the literature and in CSF interviews. These factors are listed in alphabetic order in the table 4 and are associated with the numbers in the figure 1. The most frequently cited factor in the literature is staff involvement, i.e. 46%. TABLE IV CRITICAL SUCCESS FACTORS - ANALYSIS
Critical Success Factors Creating process action teams Interactions Formal Methodology Experimenting SPI Process Awareness Process Ownership Hybrid Processes Staff Involvement Tailoring processes
Occurrence in Literature (n=47) Frequency % 7 26
Occurrence in Practice (n=23) Frequency % 2 9
12 -
21 -
5 8
22 35
4
15
3
13
9
24
12
52
3
37
-
-
12
26
4
11
24
46
10
54
18
28
2
9
This suggests that in involvement of staff can play a vital role in the implementation of improvement programs. Other
Page 347
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
frequently cited factors in the literature are Process ownership (37%), and tailoring processes (28%). It shows that practitioners consider their involvement, process ownership, and tailoring imperative for the successful implementation of improvement programs. The results also show that staff time and resources and creating process action teams are also important factors. A quarter of the literature cited reviews, experienced staff, clear and relevant process goals and assigning of responsibilities.
methodology Lack of resources Organizational politics Time pressure
7
50
8
35
4
29
12
52
5
36
4
17
Fig. 2 Critical Barriers in Process Framework
Fig. 1 Critical Success Factors in Process Framework
Table 4 shows that, like the literature, the most frequently cited factors in the interviews are staff involvement and processes awareness, i.e. cited 54% and 52% respectively. Two new methods––awareness and formal methodology–– have been identified in our empirical study which has not been identified in the literature. Other frequently cited factors are resources, staff involvement and experienced staff. Other factors are less frequently cited in the interviews. Critical Barriers Our aim of identifying critical barriers is to understand the nature of issues that undermine the process framework implementation programs. Table 5 shows the list of critical barriers cited in the literature and an empirical study. The results show that most of the practitioners in literature consider lack of resources a major critical barrier for the implementation of SPI. The results also suggest that in practitioners’ opinion time pressure and inexperienced staff can undermine the success of SPI implementation programs. It shows that practitioners would prefer to avoid organizational politics during the implementation of process framework. TABLE V CRITICAL BARRIERS - ANALYSIS
Critical Barriers Inexperienced staff Lack of awareness Lack of formal
Occurrence in Literature (n=14) Frequency % 5 36
Occurrence in Practice (n=23) Frequency 4
% 17
-
-
10
43
-
-
12
52
Velammal College of Engineering and Technology, Madurai
Staff involvement is ranked highest in CSF interviews, i.e. 54%. Two new critical barriers––lack of formal methodology and lack of awareness––have been identified in our empirical study which has not been identified in the literature. The second most cited critical barrier in CSF interviews is lack of support. The critical barrier ‘‘lack of resources’’ is cited 35% in the CSF interviews. Comparison of the critical barriers provides evidence that there are some clear similarities and differences between the findings of the two sets. There are barriers in common, i.e. inexperienced staff, lack of resources, lack of support, organizational politics, and time pressure. There are also a number of differences between the findings. For example, ‘‘changing the mindset of management and technical staff’’ and ‘‘staff turnover’’ have not been cited in our empirical study but these barriers are present in the literature. Similarly, lack of awareness of SPI and lack of formal methodology are critical in our empirical study but have not been identified through the literature. This shows that practitioners, who took part in our study, are more concerned about SPI awareness activities and implementation methodology. Implications of the Process Framework In the previous sub-section we dealt with three subdivisions that commend the static and dynamic processes. This framework we have derived out of integrating the two process sets are left for discussion and validation, which is presented in the following sections. 1) Comment on the Framework Now we put the key process area in the process framework up for discussion. The first question that needs to be answered is whether a software developers’ success relies on static or dynamic processes. In the commercial cases encountered and presented in this work 50%-70% of their effort in software development rely on the existing, well-defined static processes. However to keep their production in the uprising, they need to work on the dynamic
Page 348
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
processes and help the development processes improve over time and experience. In the case of open source products, where many of the users of the product are also developers, testers, and quality assurance team members, the same premise on the process framework holds. Since framework is a process focused one, the improvement of these processes will lead to better productivity. The second thought is whether this process framework could be followed with relevant to software development. 2) Validation of the Process Framework Usability is formally defined as the “The extent to which a product can be used by specified users to achieve specified goals with effectiveness, efficiency and satisfaction in a specified context of use’’ (ISO/DIS-9241-11, 1994). The goal of a usability analysis is to pinpoint areas of confusion and ambiguity for the user which, when improved will increase the efficiency and quality of a users’ experience (ISO/DIS-9241-11, 1994). In order to measure the usability of our framework implementation maturity model we will be using the following three goals or criteria: 1) ease of learning, 2) ease of use and 3) satisfaction. In order to validate our process framework we are planning an expert panel review process. Similar process has been used by Beecham and Hall (2003) for the validation of requirements process improvement model. An expert panel questionnaire will be sent along with the description of the process framework. This feedback will be analyzed and enhanced version of the framework will be produced based on the evaluation feedback. After the expert panel review process we are also planning to conduct a case study in a software development company in order to validate and test the framework in the real world environment.
We opt to draw a roadmap by integrating the processes that are deemed for improvement and continuous monitoring. The SPI implementation framework is in its initial stage of development and needs further improvement and evaluation. A case study will be conducted in order to test and evaluate this framework and to highlight areas where this framework can be improved.
V. CONCLUSIONS AND FUTURE WORKS In this work a Process Framework for Software Development is presented that has the potential to help companies assess and improve their process improvement program. This framework is extracted from CMMI and is based on CSFs and critical barriers identified through literature and an empirical study. For the development of this framework we have analyzed the experiences, opinions and views of practitioners in order to identify factors that have a positive or negative impact on the implementation of a improvement programs. However, this model needs further improvement and evaluation. An expert panel review will be conducted in order to get feedback from different practitioners. Based on the evaluation feedback enhanced version of the maturity model will be produced. Future Works: A case study will also be conducted in order to validate and test this process framework in practice. Our ultimate aim of conducting empirical study and developing maturity model was to develop a SPI implementation framework. We have used our empirical findings and designed a framework for guiding the design of effective implementation strategies for software process improvement.
REFERENCES [1] Alagarsamy. K., Justus S and K. Iyakutti, “Implementation Specification of a SPI supportive Knowledge Management Tool”, IET Software, Vol. 2, No.2, pp. 123-133, April 2008. [2] Alagarsamy. K., Justus S and K. Iyakutti, “Knowledge based Software Process Improvement”, Proc. of IEEE International Conference on Software Engineering Advances, France, 2007. [3] Anne Walker, and Kathleen Millington “Business intelligence and knowledge management: tools to personalize knowledge delivery” Information Outlook, August, 2003. http://www.findarticles.com/p/articles/mi_m0FWE/is_8_7/ai_ 106863496 [4] Baccarini. D, “The logical framework method for defining project success”, Project Management Journal, pp.25-32 1999. [5] Baddoo, N., Hall, T., “De-motivators of software process improvement: an analysis of practitioner’s views”, Journal of Systems and Software, Vol. 66, No. 1, pp. 23–33, 2003 [6] Baddoo, N., Hall, T., Wilson, D., “Implementing a people focused SPI programme”, Proc. in 11th European Software Control and Metrics Conference and the Third SCOPE Conference on Software Product Quality, 2000 [7] Beecham, S., Hall, T., Rainer, A., Building a requirements process improvement model, Department of Computer Science, University of Hertfordshire, Technical Report No. 378, 2003. [8] Belout. A, “Effects of human resource management on project effectiveness and success: toward a new conceptual framework”, International Journal of Project Management, Vol.16, pp. 21–26, 1998. [9] Coolican, H., Research Methods and Statistics in Psychology. Hodder and Stoughton, London, 1999. [10] CMMI of Development, ver 1.2, SEI, Pitsburgh, 2006, PA 15213-390, CMU/SEI-2006-TR-008, 2006. [11] David N Card, “Research Directions in Software Process Improvement”, Proc. in 28th Annual International Computer Software and Application Conference, 2004. [12] Emam. K.E., J.N. Drouin, and W. Melo, “SPICE:The Theory and Practice of Software Process Improvement and Capability Determination”, IEEE Computer Society, Los Alamitos, 1998. [13] Haris Papoutsakis, and Ramon Salvador Vallès, “Linking Knowledge Management and Information Technology to Business Performance: A Literature Review and a Proposed
Velammal College of Engineering and Technology, Madurai
Page 349
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Model”, Journal of Knowledge Management Practice, Vol. 7, No. 1, March, 2006. [14] Jesper Arent, and Jacob Norbjerg., “Software Process Improvement as Organizational Knowledge Creation: A Multiple Case Analysis”, Proc. Of the 33rd Hawaii International Conference on System Sciences, 2000. [15] Kathleen Coleman Dangle, Patricia Larsen, and Michele Shaw, “Software Process Improvement in Small Organizations: A Case Study”, IEEE Software, pp 68-75, 2005. [16] Kautz. K, W.H. Henrik and Kim Thaysen, “Applying and Adjusting a Software Process Improvement Model in Practice: The Use of the IDEAL Model in a Small Software Enterprise”, Proc. International Conference on Software Engineering, Limerick, Ireland, SACM 23000 1-58113-2069/00/06, 2000. [17] Kellyann Jeletic, Rose Pajerski and Cindy Brown, Software Process Improvement Guidebook, NASA, Greenbelt, Maryland, March 1996. [18] Knowledge Management, APQC Online Discussion, 2005. http://www.apqc.org/portal/apqc/site/generic?path=/site/km/o verview.jhtml [19] Knowledge Management and e-Learning Integration, 2004. “http://knowledgemanagement.ittoolbox.com/documents/popu lar-q-anda/knowledge-management-and-elearning-integration2846”. [20] Koren. E.F. and S. Elisabeth Hove, “Using a Knowledge Survey to plan Software Process Improvement Activities – a Case Study”, 2002, http://heim.ifi.uio.no/~isu/INCO/Papers/Using_efk_eurospi20 02_paper.pdf [21] Krippendorf, K., Content Analysis: An Introduction to its Methodologies, Sage, London, 1980. [22] Marwick. A.D., “Knowledge Management Technology”, IBM System Journal, Vol. 40, No 4, 2001. [23] McFeeley. B., IDEAL: A User's Guide for Software Process Improvement, SEI, Pittsburgh, Handbook CMU/SEI96-HB-001, 1996. [24] Medha Umarji and Carolyn Seaman, “Predicting Acceptance of Software Process Improvement”, Human and Social Factors of Software Engineering, St. Louis, Missouri, USA, 2005. [25] Michael, S., Lewis, B., Research Practice: International Handbooks of Quantitative Application in Social Sciences, Sage, 1994. [26] Paulk, Mark C. & Chrissis, Mary Beth, The 2001 High Maturity Workshop (CMU/SEI-2001-SR-014). Pittsburgh, PA: Software Engineering Institute, Carnegie Mellon University, January 2002. http://www.sei.cmu.edu/publications/documents/01.reports/01 sr014.html. [27] PMI Standards Committee, William R Duncan, A Guide to the Project Management Body of Knowledge, Project management Institute, 1998.
[28] Pourkomeylian. P., Software Process improvement, Doctoral Dissertation, Goteborg University, ISSN 1400-741X, 2002. [29] SEI, 2002. Capability Maturity Model Integration, (CMMISM), Version 1.1. SEI, CMU/SEI-2002-TR-029. [30] Stelzer, D., Werner, M., “Success factors of organizational change in software process improvement”, Software Process Improvement and Practice, Vol. 4, No. 4, 1999. Wesley Vestal., Integrating Knowledge Management and Organizational Learning, American Productivity & Quality Center, 2005. http://www.apqc.org/portal/apqc/site/content?docid=11978
Velammal College of Engineering and Technology, Madurai
Page 350
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Exploiting Parallelism in Bidirectional Dijkstra for Shortest-Path Computation R. Kalpana#1, Dr. P.Thambidurai#2, R. Arvind kumar #3, S. Parthasarathi #4, Praful Ravi #5 #
Department of Computer Science & Engineering , Pondicherry Engineering College Puducherry, India 1
[email protected]
Abstract— The problem of finding a shortest path between two given nodes in a directed graph is a common task which is traditionally solved using Dijkstra’s algorithm [1]. There are many techniques available to speedup the algorithm while guaranteeing the optimality of the solution. Almost all of these speed-up techniques have a substantial amount of parallelism that can be exploited to decrease the running time of the algorithm. The rapidly evolving field of multiprocessing systems and multi-core processors provide opportunities for such improvements. The main focus of this paper is to identify features of the Bidirectional search speed-up technique suitable for parallelization and thereby reduce the running time of the algorithm. Keywords— Dijkstra’s Algorithm, algorithm, running time, speed-up
Graph
Theory,
parallel
VI. INTRODUCTION Computing shortest paths is a basic operation which finds application in various fields. The most common applications[2] for such a computation includes route planning systems for bikes, cars and hikers, and timetable based information systems for scheduled vehicles such as buses and trains. In addition to these, shortest path computation also finds its application in spatial databases and even in web searching. The core algorithm that serves as the base for the above applications is a special case of single-source shortest-path problem on a given directed graph with non-negative edge weights called Dijkstra’s algorithm[1]. The particular graphs[3], [4], [5] considered for the above applications are huge and the number of shortest path queries to be computed within a short time is huge as well. This is the main reason for the use of speed-up techniques for shortest-path computations. The focus of these techniques is to optimize the response time for the queries, that is, a speed-up technique is considered as a technique that reduces the search space of Dijkstra’s algorithm by using pre-computed values or by using problem specific information. Often the underlying data contains geographic information, that is, a layout of the graph is easily obtained. In
Velammal College of Engineering and Technology, Madurai
many applications the given graph can be considered static which enables preprocessing. Another possibility for improving the runtime of speed-up techniques is to identify the regions, if any, that are suitable for parallel processing. It is observed that the various speed-up techniques posses, to varying extents, operations that are possible candidates for parallel execution. Coupled with the improvement of processing capabilities of today’s modern systems the time required for computing shortest path can be effectively reduced. Also the speed-up techniques are such that, combination of two or more techniques is possible and the effective speed-up achieved by the combination is increased. In this paper, a study of the various speed-up techniques is made and the possible parallel regions present in speed-up techniques are analyzed. Experiments [6] were carried out to analyze the performance of a particular speed-up technique with the parallelism present in it utilized in different ways. The Parallelization adapted in Highway node routing [6] for solving shortest path problem gives a very good speedup compared to that of sequential algorithm. The bidirectional search is considered as one of the basic speedup techniques, which improves the speedup by searching the graph both forward and reverse. Using the OpenMp constructs with LEDA library the speedup of the algorithm is improved. VII. PRELIMINARIES The formal definitions and a brief description of Dijkstra’s Algorithm are given in this section. A. Definitions A graph G is a pair (V, E), where V is the set of nodes / vertices and E ⊆ V × V is a set of edges, where an edge is an ordered pair of nodes of the form (u, v) such that u , v ∈ V . Usually the number of nodes V is denoted by n and the number of edges E is denoted by m. For a node u ∈ V , the number of outgoing edges {( u , v ) ∈ E } is called the degree of the node.
Page 351
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A path in graph G is a sequence of nodes (u 1 ,..., u k ) so that (u i , u i +1 ) ∈ E for all 1 ≤ i < k . A path in which u 1 = u k is called a cycle or cyclic path. Given the edge weights l : E → R , the length of the path P = ( u 1 ,..., u k ) is the sum of the lengths of its edges l ( P ) :=
∑
1≤ i < k
l ( u i , u i +1 ) .
For two nodes s , t ∈ V , a shortest s-t path is a path of minimal length with u 1 = s and u k = t . The distance d ( s , t ) between s and t is the length of the shortest path s-t. A layout of a graph G = (V , E ) is a function L : V → R 2 , that assigns each node a position in R 2 . A graph is called sparse if m = O (n ) . B. Shortest Path Problem Consider a directed graph G = (V , E ) whose edge weights are given by a function l : E → R . The single-source singletarget shortest path problem is to find a shortest s-t path from a source s ∈ V to a specified target t ∈ V . The problem is said to be well defined for all pairs, if G doesn’t contain cycles with negative lengths. If negative weights are present but not negative cycles, using Johnson’s algorithm, it is possible to convert the original edge weights l : E → R to non-negative weights l ' : E → R 0+ that results in the same shortest paths. So it is safe to consider that edge weights are non-negative. The Dijkstra’s algorithm[1] maintains, for each node u ∈ V , a label dist(u) with the current distance values. A priority queue Q contains the nodes that form the search around source s. Nodes can be either unvisited, in the priority queue or already visited. In the computation of shortest path it has to be remembered that u is the predecessor of v if a shorter path v has been found. Dijkstra’s algorithm is used for computing shortest paths from a single source to all other nodes in the graph. If only one shortest path is required to a given target t ∈ V , the algorithm can halt when the target is eliminated from the priority queue. The time complexity of the algorithm is decided by the priority queue chosen for implementation. For usual graphs, Fibonacci heaps, provide the best theoretical worst-case time of O ( m + n log n ) . Binary heaps also result in the same time complexity for sparse graphs. It is to be noted that binary heaps are easier to implement and also performs better in many practical situations. It is arguable that a better speed-up technique reduces the search front and the priority queue will be of lesser importance in such a case.
A. Bidirectional Search In bidirectional search[2] a normal or forward variant of the algorithm starting at the source and a reverse or backward variant of the algorithm starting at the destination are run simultaneously. The backward variant of the algorithm is applied to the reverse graph, that is, a graph having the same node set V as the given graph and the reverse edge set E = {(u , v ) | ( v , u ) ∈ E }. The algorithm can be terminated when a node has been labeled permanent by both the forward and backward searches. The shortest path can be obtained by combining the path obtained in the forward search with the one obtained in backward search. B. Goal-Directed Search This technique [2], [8] alters the priority of the active nodes to change the order in which the nodes are processed. More precisely a potential p t (v ) depending on a target ‘t’ is added to the priority dist (v ) . So the modified priority of a node v ∈ V is given by dist ( v ) + p t ( v ) . With an appropriate potential the search can be directed to the target thereby reducing the running time and at the same time finding a shortest path. An alternate approach to modify the priority, is to change the edge lengths so that the search is directed towards the target ‘t’. In this case, the length of an edge ( u , v ) ∈ E is replaced by l ' (u , v ) = l (u , v ) − p t ( v ) + p t (u ) . It can be easily verified that a path from s to t is a shortest s-t path according to l ' , if and only if it is a shortest s-t path according to l. C. Multilevel Approach This speed-up technique has a preprocessing step in which the input Graph G is decomposed into l + 1 levels ( l ≥ 1) and enriched with extra edges which represent shortest path between certain nodes[2],[9],[10]. The decomposition of the graph depends on the selected nodes Si, at level i such that S 0 := V ⊇ S 1 ⊇ K ⊇ S l . The decomposition of the node sets can be based on various criteria. Selecting desired number of nodes with the highest degree in the graph, works out to be an appropriate criterion. Three different types of edges are added to the graph: upward edges, going from a node that is not selected at a particular level to a node selected at that level, downward edges, connecting selected to non-selected nodes and level edges, connecting selected nodes at one level. For finding a shortest path between two nodes, it is sufficient to consider a smaller subgraph of the “multilevel graph”.
VIII. RELATED WORK There is a high volume of ongoing research on shortest paths. The basic speedup techniques and the combined speedup techniques [2], [4], [7] for Dijkstra’s algorithm that usually reduce the number of nodes visited and running time of the algorithm in practice which are discussed here.
D. Shortest-Path Containers This speed-up technique requires a preprocessing computation of all shortest-path trees. For each edge, e ∈ E , a node set, S (e ) , to which the shortest path begins with e is computed[2],[11]. By using a layout, for each edge e ∈ E , the bounding box of S ( e ) is stored in an associative array C with
Velammal College of Engineering and Technology, Madurai
Page 352
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
index set E. The preprocessed information is used for determining if a particular edge will be present on a shortest path to a given target. E. Combination of speed-up Techniques The different speed-up techniques mentioned above can be combined in with each other, which further improve the efficiency during shortest path combination. Some of the combinations were implemented and analyzed[2]. Reach based routing[12] is also combined with many of the basic techniques. IX. OPENMP This section contains parallel programming constructs used to achieve parallelism. OpenMP is an Application Programming Interface (API)[13] for portable shared memory parallel programming. The API allows an incremental approach to parallelize an existing code, so that portions of a program can be parallelized in successive steps. There is a marked difference between OpenMP and other parallel programming paradigms where the entire program must be converted in one step. A. Fork and Join Model of OpenMP Multithreaded programs can be written in different ways that allows complex interaction between threads. OpenMP provides a simple structured approach to multithreaded programming. OpenMP supports a fork-join programming model as shown in Fig. 1.
to assign work to them according to the strategy specified by the programmer. B. Features provided by OpenMP OpenMP provides the following features that can be utilized by a user (i) create teams of threads for parallel execution (ii) specify how work must be shared among the threads (iii) scope for variables (iv) synchronization constructs. Creating a team of threads is achieved using parallel construct. #pragma omp parallel [clause [[,] clause]…] Structured block Although this construct ensures that computations in the associated parallel region are performed in parallel, it does not distribute the work among the threads in a team. There is an implied barrier at the end of the parallel region that forces all threads to wait until the work inside the region has been completed. There are three work sharing constructs available in OpenMP. 1) #pragma omp for Used for sharing iterations in a loop 2) #pragma omp sections Specify different work for each thread individually 3) #pragma omp single Allow only one thread to execute the enclosed region OpenMP provides many constructs for synchronization. Locks are also supported. The shortest path problem is solved in parallel with OpenMP in [6]. Although there is improvement in speedup, there are many speedup techniques which can be parallelized and the performance can be improved. X. SPEEDUP TECHNIQUES The modules in bidirectional search are segregated in such a way that the phases where the parallelism can be applied are identified. Those portions are executed in parallel using OpenMP based on the Multicore architecture.
Fig. 1 Fork-Join programming model of OpenMP
In this approach, the program starts as a single thread of execution. Whenever an OpenMP parallel construct is encountered by a thread while it is executing the program, it creates a team of threads (this is the fork), becomes the master of the team, and collaborates with the other members of the team to execute the code enclosed by the construct. At the end of the construct, only the original thread continues; all others terminate (this is the join). Each portion of code enclosed by a parallel construct is called a parallel region. OpenMP expects the application developer to give a high-level specification of the parallelism in the program and the method for exploiting that parallelism. The OpenMP implementation will take care of the low-level details of actually creating independent threads to execute the code and
Velammal College of Engineering and Technology, Madurai
A. Bidirectional Search The bidirectional search[4] algorithm has a considerable portion of its operations suitable for parallel execution. Initialization (Forward Search) [IFS] Initialization (Backward Search) [IBS] Node Selection (Forward Search) [SFS] Distance updation (Forward Search) [UFS] Node Selection (Backward Search) [SBS] Distance updation (Backward Search) [UBS]
Fig. 2, Bidirectional Search. The inner box with dash-dot line indicates a loop.
Page 353
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
The various modules for bidirectional search is outlined in Fig. 2, and detailed in Algorithm 1. In bidirectional search two searches are carried out simultaneously: a forward variant of the algorithm, starting from the source, and a reverse variant of the algorithm, starting from the target. The reverse variant of the algorithm is applied to the reverse graph. The search terminates when a node is marked permanent by both searches. In this scenario, the forward and reverse searches are mutually independent to certain extent, except for the set “perm” which contains permanent nodes marked by both.
search process. Figure 3, describes bidirectional search in which only the updation of neighbor distance values is parallelized. IFS
IBS SFS SBS
//Initialization phase for all nodes u ∈ V set dist ( u ) := ∞
UFS
UBS
initialize priority queue Q with source s and
dist ( s ) := 0 for all nodes u ∈ V set rdist (u ) := ∞ initialize priority queue rQ with target t and
Initialization parallelized using OpenMP Sections
rdist ( t ) := 0 initialize a perm to empty set // Node selection phase begins while priority queue Q and rQ are not empty { get node u with smallest tentative distance dist(u) in Q if( u = t or u ∈ perm ) halt. add u to perm get node ru with smallest distance rdist(ru) in rQ if( ru = s or ru ∈ perm ) halt. add ru to perm // update phase forward search for all neighbor nodes v of u set new-dist := dist(u) + w(u, v) update dist(v) of forward priority queue Q // update phase reverse search for all neighbor nodes rv of ru set new-dist := rdist(ru) + w(ru, rv) update dist(rv) of reverse priority queue rQ } Algorithm 1: Bidirectional Dijkstra’s Algorithm
B. Bidirectional Search with parallelized updates There are two regions where parallelism can be exploited. The initialization of required data structures for both forward and reverse searches can be performed in parallel, that is, data level parallelism can be adopted. The parallel bidirectional algorithm is given in Algorithm 2. The initialization of both forward and reverse search can be parallelized. Another operation that lends itself to parallelization is the updation of distance values, for the neighbors, of a node which is marked permanent. In this region also data parallelism is possible. It is to be noted that another possibility for parallelism is to run both the forward and reverse variants of the algorithm simultaneously as independent threads of the
Velammal College of Engineering and Technology, Madurai
Updates parallelized using OpenMP Sections
Fig. 3, Bidirectional Search with parallelized updates
//Initialization phase # pragma omp section { initialize forward priority queue Q with dist(u)=∞ and dist(s)=0 } # pragma omp section { initialize reverse priority queue rQ with rdist(u)=∞ and rdist(t)=0 } initialize perm array with empty set //Node selection phase begins While (priority queue Q and rQ are not empty) { get node with min dist(u) in Q and assign it to u if (u==t or u € perm) then exit else add u to perm get node with min rdist(u) in rQ and assign it to ru if (ru==s or ru € perm) then exit else add ru to perm //Update phase forward search # pragma omp section { for all neighbor nodes v of u set new-dist := dist(u) + w(u, v) update dist(v) of forward priority queue Q } //Update phase reverse search # pragma omp section { for all neighbor nodes rv of ru set new-dist := rdist(ru) + w(ru,rv) update dist(rv) of reverse priority queue rQ } } Algorithm 2: Bidirectional Search with parallelized updates
XI. EXPERIMENTAL RESULTS
Page 354
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Dijkstra’s algorithm, the bidirectional search and the bidirectional search technique with parallelized updates were implemented and the results were analyzed. The different algorithms mentioned were implemented in C++ with the help of LEDA library. The graph and priority queue data structures as well as other utilities such precise time measurements provided by LEDA were used in the implementation. For parallelizing the regions of code in bidirectional search the OpenMP API was utilized. The use of OpenMP helped to maintain a single source for both the sequential and parallel versions of the algorithm. The code was compiled using GNU C++ Compiler (g++ version 4.2) and the experiments were performed on an Intel Core2Duo machine (2.20 GHz) with 1 GB RAM running on Ubuntu GNU/Linux 8.04 (kernel version 2.6) The main result of this work is shown in Fig. 5. The running time of the different algorithms is plotted (along y-axis) against the number of nodes (along x-axis). The graphs generated were random graphs with the number of edges in O(n), that is, the graphs generated were sparse. The experiments were conducted for 30 runs, and the results of running time of the algorithm are tabulated in Table 1. From Fig. 5, it is observed that the bidirectional search reduces the running time of Dijkstra’s algorithm by a factor of 2.2 while the parallelized bidirectional search reduces the runtime by a factor of 3.1. Run No. 1
Nodes 1000
0.001
0.0015
Parallel Bi+Dij 0.003
5
5000
0.0115
0.006
0.0055
10
10000
0.0295
0.014
0.013
15
15000
0.0495
0.0225
0.0175
20
20000
0.0685
0.036
0.023
25
25000
0.103
0.043
0.0305
30
30000
0.126
0.05
0.0355
1.692
0.7695
0.546
0.0564
0.0257
0.0182
2.1988304 09
3.098901
Total Runtime (30 runs) Average Runtime Speed-Up
Dij
Bi+Dij
TABLE V RUNNING TIME IN DIJKSTRA AND BIDIRECTIONAL DIJKSTRA WITH PARALLELIZED UPDATES
Velammal College of Engineering and Technology, Madurai
Fig. 5. Runtime comparison of the different speed-up techniques Alg 1 – Traditional Dijkstra Alg 2 – Bidirectional Dijkstra Alg 3 – Bidirectional Dijkstra with updates parallelized
RRun No.
Nodes
Dij
Bi+Dij
1
1000
576
55
5
5000
2103
134
10
10000
4190
170
15
15000
7263
236
20
20000
11633
271
25
25000
11932
202
30
30000
14692
275
238203
6020
7940.1
200.6666667
Total Vertex Visit Count (30 runs) Average Vertex Visit Count Speed-Up
39.56860465
TABLE IVI NUMBER OF NODES VISITED IN DIJKSTRA AND BIDIRECTIONAL DIJKSTRA
The experiments were conducted for 30 runs, and the results of Number of nodes visited of the algorithm are tabulated in Table II. Fig 6. plots the number of nodes visited during shortest path computation, using Dijkstra’s algorithm, against the number of nodes available in the graph. Similarly the graph in Fig. 6 takes into account the number of nodes visited during shortest path calculation using bidirectional search in Dijkstra’s algorithm.
Page 355
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
No.of Nodes visited during shortest path computation
[2] Martin Holzer, Frank Schulz, Dorothea Wagner, and Thomas Willhalm, Combining Speed-up Techniques for Shortest-Path Computations, ACM Journal of Experimental Algorithmics, Vol. 10, Article No. 2.5, 2005. [3] Johnson, D. B, Efficient algorithms for shortest paths in sparse networks, Journal of the Association for Computing Machinery,1959, 1–13.
16000
No.of Nodes Visited
14000
[4] Frederikson, G. N, Fast algorithms for shortest paths in planar graphs with applications, SIAM Journal on Computing, 16(6), 1987, 1004–1022.
12000 10000 8000
Dij Bi+Dij
6000
[5] Agarwal. R. and Jagadish, H. V, Algorithms for searching massive graphs, IEEE Transaction son Knowledge and Data Engineering, 6(2),1994, 225–238. [6] , Dominik Schultes, Johannes Singler, Peter Sanders, Parallel Highway Node Routing, A Technical Report, Feb,2008. algo2.iti.kit.edu/schultes/hwy/parallelHNR.pdf
4000 2000 0 1000
5000
10000
15000
20000
25000
30000
No.of Nodes
[7] Willhalm, T, Engineering shortest paths and layout algorithms for large graphs., Ph.D.thesis, Faculty of Informatics, University of Karlsruhe, 2005 [8 Andrew V.Goldberg, Chris Harrelson, Computing the Shortest Path: A* Search Meets Graph Theory, Microsoft Research, 1065 La Avenida, Mountain View, CA 94062. 2005
Fig. 6. Number of nodes visited during shortest path computation Dij – Dijkstra’s Algorithm Bi+Dij – Bidirectional Dijkstra
The result will be the same for Bidirectional Dijkstra with updates parallelized. It is observed that the number of nodes visited during shortest path computation is comparatively less in bidirectional search. The Number of nodes visited in Bidirectional Dijkstra is reduced by a factor of 40. XII. CONCLUSION The speed-up technique for Dijkstra’s algorithm reduces the runtime of the algorithm while retaining optimality. Even though the number of nodes visited during bidirectional search and bidirectional search with updates parallelized are the same. The running time is still reduced in the case of bidirectional search with updates parallelized. Also the speedup techniques can be combined to increase the speed-up considerably. These techniques lend themselves for parallelization which can be effectively utilized in large applications like route planning systems, time table information systems, etc., where multi-core processors are utilized so that the system efficiency can be improved. Locks can be applied during parallelizing the bidirectional Dijkstra for special case of graphs.
[9] Peter Sanders and Dominik Schultes , Highway Hierarchies Hasten Exact Shortest Path Queries, , ESA 2005, LNCS 3669, Springer-Verlag Berlin Heidelberg 2005, 568–579.. [10] Holzer, Hierarchical speedup techniques for shortest path algorithms, M, Tech. report, Dept of Informatics, University of Konstanz, Germany, 2003. [11] Dorothea Wagner and Thomas Willhalm, Geometric Containers for Efficient Shortest-Path Computation, ACM Journal of Experimental Algorithmics, 10(1.3), 2005. [12] Gutman. R.J, Reach-based routing: A new approach to shortest path algorithms optimized for road networks, In Proceedings of the Sixth Workshop on Algorithm Engineering and Experiments and the First Workshop on Analytic Algorithmics and Combinatorics, 2004. [13] OpenMP Website (http://www.openmp.org)
REFERENCES [1] Dijkstra. E. W , A note on two problems in connection with graphs, In Numerische Mathematik. Vol. 1. Mathematisch Centrum, Amsterdam, The Netherlands, 1959, 269–271.
Velammal College of Engineering and Technology, Madurai
Page 356
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Hiding Sensitive Frequent Item Set by Database Extension B. Mullaikodi, M.E#1,Dr. S.Sujatha M.E.*2, #
Computer Science and Engineering Department, Anna University, Trichy 1
[email protected]
Abstract -Sharing data among organizations often leads to mutual benefit. Recent technology in data mining has enabled efficient extraction of knowledge from large databases. This, however, increases risks of disclosing the sensitive knowledge when the database is released to other parties. To address this privacy issue, one may sanitize the original database so that the sensitive knowledge is hidden. Sensitive knowledge hiding in large transactional databases is one of the major goals of privacy preserving data mining. A novel approach that performs sensitive frequent item set hiding by applying an extension to the original database was proposed. The extended portion of data set contains transactions that lower the importance of the sensitive patterns, while minimally affecting the importance of the nonsensitive ones. We present the border revision to identify the revised border for the sanitized database and then we compute the minimal size for the extension. The hiding process involves the construction of a Constraints Satisfaction Problem, by using item sets of revised borders and its solution through Binary Integer Programming. Keywords – Privacy preserving data mining, knowledge hiding, association rule mining, binary integer programming.
I. INTRODUCTION A subfield of privacy preserving data mining is “knowledge hiding”. The paper presents a novel approach that strategically performs sensitive frequent itemset hiding based on a new notion of hybrid database generation. This approach broadens the regular process of data sanitization by applying an extension to the original database instead of either modifying existing transactions, or rebuilding the dataset from scratch. The extended portion of the dataset contains a set of carefully crafted transactions that achieve to lower the importance of the sensitive patterns to a degree that they become uninteresting from the perspective of the data mining algorithm, while minimally affecting the importance of the nonsensitive ones. The hiding process is
Velammal College of Engineering and Technology, Madurai
guided by the need to maximize the data utility of the sanitized database by introducing the least possible amount of side-effects, such as (i) the hiding of non-sensitive patterns, or (ii) the production of frequent patterns that were not existent in the initial dataset. The released database, which consists of the initial part (original database) and the extended part (database extension), can guarantee the protection of the sensitive knowledge, when mined at the same or higher support as the one used in the original database. The approach introduced in this paper is exact in nature. On the contrary, when an exact solution is impossible, the algorithm identifies an approximate solution that is close to the optimal one. To accomplish the hiding task, the proposed approach administers the sanitization part by formulating a Constraint Satisfaction Problem (CSP) and by solving it through Binary Integer Programming (BIP). II. KNOWLEDGE HIDING FORMULATION This section provides the necessary background regarding sensitive itemset hiding and sets out the problem at hand, as well as the proposed hiding methodology.
A. Frequent Itemset: Let I = {i1, i2, …………., iM} be a finite set of literals, called items, where M denotes the cardinality of the set. Any subset I ⊆ I is an item set over I. A transaction T over I is a pair T = (tid, I), where I is the item set and tid is a unique identifier, used to distinguish among transactions that correspond to the same item set. A transaction database D = {T1, T2, ....TN} over I is an N x M table consisting of N transactions over I carrying different identifiers, where entry Tnm = 1 if and only if the mth item (m ∈ [1, M]) appears in the nth transaction (n ∈ [1, N]). Otherwise, Tnm = 0. A transaction T = (tid, J) supports an item set I over I, if I ⊆ J. Let S be a set of items; notation p(S) denotes the powerset of S, which is the set of all subsets of S.
Page 357
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Given an item set I over I in D, sup (I, D) denotes the number of transactions T ∈ D that support I and freq(I, D) denotes the fraction of transactions in D that support I. An item set I is called large or frequent in database D if and only if its frequency in D is at least equal to a minimum threshold mfreq. A hybrid of Apriori and FP-Tree algorithms are proposed to be used to find the frequent item set. A hybrid of Apriori and FP-Tree algorithms are proposed to be used to find an optimized set of frequent item set. 1) The Apriori Algorithm - Finding Frequent Itemsets Using Candidate Generation:
Apriori is an influential algorithm for mining frequent itemsets for Boolean association rules. Apriori employs an iterative approach known as a level-wise search, where k-itemsets are used to explore (K+1) itemsets. First , the set of frequent 1-itemsets is found. This set is denoted L1. L1 is used to find L2, the set of frequent 2itemsets , which is used to find L3 , and so on, until no more frequent k-itemsets can be found. The finding of each Lk requires one full scan of the database. To improve the efficiency of the level-wise generation of frequent itemsets, an important property called the Apriori property, i.e., all nonempty subsets of a frequent itemset must also be frequent, is used to reduce the search space. A two step process join and prune actions are used to find Lk from Lk-1.
• If the conditional FP-tree contains a single path, simply enumerate all the patterns
B.Hiding methodology To properly introduce the hiding methodology, one needs to consider the existence of three databases, all depicted in binary format. They are defined as follows: ¾ Database Do, is the original transaction database that, when mined at a certain support threshold msup, and leads to the disclosure of some sensitive knowledge in the form of sensitive frequent patterns. This sensitive knowledge needs to be protected. ¾ Database Dx ,is a minimal extension of Do that is created by the hiding algorithm during the sanitization process, in order to facilitate knowledge hiding. ¾ Database D ,is the union of database Do and the applied extension Dx and corresponds to the sanitized outcome that can be safely released. TABLE 1 Sample : Sanitized Database D as a Mixture of the Original Database DO and the Applied Extension DX
2) FP-Growth Algorithm-Mining Frequent patterns without candidate generation : Frequent pattern growth or simply FP-growth adopts a divide-and – conquer strategy as follows: compress the database representing frequent items into a frequent pattern tree, or FP tree, but retain the itemset association information and then divide such a compressed database into a set of conditional databases, each associate with one frequent item, and mine each such database separately. TABLE 2
Major steps to mine FP-tree
Sample : Frequent Item Sets for DO and DX at msup = 3 (for table 1)
1) Construct conditional pattern base for each node in the FP-tree 2) Construct conditional conditional pattern-base
FP-tree
from
each
3) Recursively mine conditional FP-trees and grow frequent patterns obtained so far
Velammal College of Engineering and Technology, Madurai
Page 358
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Given the sanitized database D, its original version DO, and the produced extension DX , the quality of database D is measured both in the size of DX and in the number of binary variables set to “1” in the transactions of DX (i.e., the distance metric). In both cases, lower values correspond to better solutions.
III. HYBRID SOLUTION METHODOLOGY
Definition 3 (ideal solution). A solution to the hiding of the sensitive item sets is considered as ideal if it has the minimum distance among all the existing exact solutions and is obtained through the minimum expansion of DX. In that sense, ideal is a solution that is both minimal (with respect to distance and size of extension) and exact.
A. Computation of size of database extension
C. Border Revision
Database DO is extended by DX to construct database D. An initial and very important step in the hiding process is the computation of the size of DX. A lower bound on this value can be established based on the sensitive item set in S, which has the highest support. The rationale here is given as follows: by identifying the sensitive item set with the highest support, one can safely decide upon the minimum number of transactions that must not support this item set in DX, so that it becomes infrequent in D.
The rationale behind this process is that hiding of a set of item sets corresponds to a movement of the original borderline in the lattice that separates the frequent item sets from their infrequent counterparts , such that the sensitive item sets lie below the revised borderline. There are four
Lower bound Q is
⎢ sup( I N , DO ) ⎥ Q=⎢ − N⎥ +1 ⎣ mfreq ⎦
………….1
Equation (1) provides the absolute minimum number of transactions that need to be added in DX, to allow for the proper hiding of the sensitive item sets of DO. However, this
lower bound can, under certain circumstances, be insufficient to allow for the identification of an exact solution, even if one exists. To circumvent this problem, one needs to expand the size Q of DX as determined by (1), by a certain number of transactions. A threshold, called safety margin (denoted
hereon as SM), is incorporated for this purpose. Safety margins can be either predefined or be computed dynamically, based on particular properties of database DO and / or other parameters regarding the hiding process.
B. Exact and Ideal Solutions Definition 1 (feasible/exact/approximate solution). A solution to the hiding of the sensitive knowledge in DO is considered as feasible if it achieves to hide the sensitive patterns. Any feasible solution, introducing no side effects in the hiding process, is called exact. Finally, any non exact feasible solution is called approximate.
possible scenarios involving the status of each item set I prior and after the application of border revision:
C1 : Item set I was frequent in DO and remains frequent in D. C2 : Item set I was infrequent in DO and is infrequent in D. C3 : Item set I was frequent in DO and became infrequent in D. C4 : Item set I was infrequent in DO and became frequent in D. Since the borders are revised to accommodate for an exact solution, the revised hyper plane is designed to be ideal in the sense that it excludes only the sensitive item sets and their supersets from the set of frequent patterns in D, leaving the rest of the item sets in their previous status as in database DO. The first step in the hiding methodology rests on the identification of the revised borders for D. The hiding algorithm relies on both the revised positive and the negative borders, denoted as Bd+ (F1D) and Bd– (F1D), respectively. After identifying the new (ideal) borders, the hiding process has to perform all the required minimal adjustments of the transactions in Dx to enforce the existence of the new borderline in the result database.
Definition 2 (database quality).
Velammal College of Engineering and Technology, Madurai
Page 359
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
more, by carefully selecting the item sets of set C, the hiding algorithm ensures that the exact same solution to the one of solving the entire system of inequalities is attained. This is accomplished by exploiting cover relations existing among the item sets in the lattice. Set C is chosen appropriately to consist of all the item sets of the revised border. The proposed hiding algorithm is capable of ensuring that if (3) and (4) are satisfied for all the item sets in C, then the produced solution is exact and is identical to the solution involving the whole system of the 2M – 1 inequalities. Cover relations governing the various item sets in the lattice of DO ensure that the formulated set of item sets C has an identical solution to the one of solving the system of all 2M – 1 inequalities for D. Fig. 1 An sample item set lattic demonstration (a) the original border and the sensitive item sets, and (b) the revised border for Table1
D. Problem Size Reduction To enforce the computed revised border and identify the exact hiding solution, a mechanism is needed to regulate the status (frequent versus infrequent) of all the item sets in D. Let C be the minimal set of border item sets used to regulate the values of the various uqm variables in DX. Moreover, suppose that I ∈ C is an item set, whose behavior we want to regulate in D. Then, item set I will be frequent in D if and only if sup (I, Do) + Sup(I, Dx) ≥ m freq x (N+Q), or equivalently if Q
Sup (I,DO) +
∑∏
uqm
≥
mfreq x (N + Q) …(3)
q =1 i M ∈I
equivalently if Q
sup (I,DO) +
∑∏
uqm < mfreq x (N + Q) ...(4)
q =1 i M ∈I
Inequality (3) corresponds to the minimum number of times that an item set I has to appear in the extension DX to remain frequent in D. On the other hand, (4) provides the maximum number of times that an item set I has to appear in DX to be infrequent in database D. To identify an exact solution to the hiding problem, every possible item set in P, according to its position in the lattice—with respect to the revised border—must satisfy either (3) or (4). However, the complexity of solving the entire system of the 2M _ 1 inequalities is well known to be NP-hard . Therefore, one should restrict the problem to capture only a small subset of these inequalities, thus leading to a problem size that is computationally manageable. The proposed problem formulation achieves this by reducing the number of the participating inequalities that need to be satisfied. Even
Velammal College of Engineering and Technology, Madurai
The cover relations that exist between the item sets of Bd+ (F`D) and those of F`D. In the same manner, the item sets of Bd– (FD) are generalized covers for all the item sets of P \ (Bd+ (F`D) ∪ {∅}). Therefore, the item sets of the positive and the negative borders cover all item sets in P. Optimal solution set C: The exact hiding solution, which is identical to the solution of the entire system of the 2M – 1 inequalities, can be attained based on the item sets of set C = Bd+ (F`D) ∪ Bd– (F`D) Based on (7), the item sets of the revised borders Bd+ (F`D) and Bd– (F`D) can be used to produce the corresponding inequalities, which will allow for an exact hiding solution for DO.
D. Handling of Suboptimality Since an exact solution may not always be feasible, the hiding algorithm should be capable of identifying good approximate solutions. There are two possible scenarios that may lead to nonexistence of an exact solution. Under the first scenario, DO itself does not allow for an optimal solution due to the various supports of the participating item sets. Under the second scenario, database DO is capable of providing an exact solution, but the size of the database extension is insufficient to satisfy all the required inequalities of this solution. To tackle the first case, the hiding algorithm assigns different degrees of importance to different inequalities. To be more precise, while it is crucial to ensure that (4) holds for all sensitive item sets in D, thus they are properly protected from disclosure, satisfaction of (3) for an item set rests in the discretion of ensuring the minimal possible impact of the sanitization process to DO. This inherent difference in the significance of the two inequalities, along with the fact that solving the system of all inequalities of the form (4) always leads to a feasible solution (i.e., for any database DO), allows the relaxation of
Page 360
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the problem, when needed, and the identification of a good approximate solution. To overcome the second issue, the hiding algorithm incorporates the use of a safety margin threshold, which produces a further expansion of DX by a certain number of transactions. These transactions must be added to the ones computed by using (1). The introduction of a safety margin can be justified as follows: Since (1) provides the lower bound on the size of database DX , it is possible that the artificially created transactions are too few to accommodate for the proper hiding of knowledge. This situation may occur due to conflicting constraints imposed by the various item sets regarding their status in D. These constraints require more transactions (or to be more precise, more item modifications) in order to be met. Thus, a proper safety margin will allow the algorithm to identify an exact solution if such a solution exists. The additional extension of DX, due to the incorporation of the safety margin, can be restricted to the necessary degree. A portion of transactions in DX is selected and removed at a later point, thus reducing its size and allowing an exact solution. Therefore, the only side effect of the use of the safety margin in the hiding process is an inflation in the number of constraints and associated binary variables in the problem formulation, leading to a minuscule overhead in the runtime of the hiding algorithm.
E. Formulation and Solution of the CSP A CSP is defined by a set of variables and a set of constraints, where each variable has a nonempty domain of potential values. The constraints involve a subset of the variables and specify the allowable combinations of values that these variables can attain. Since in this work all variables involved are binary in nature, the produced CSP is solved by using a technique called BIP that transforms it to an optimization problem. To avoid the high degree of constraints, the application of a Constraints Degree Reduction(CDR) approach is essential. On the other hand, the resulting inequalities are simple in nature and allow for fast solutions, thus adhere for an efficient solution of the entire CSP. The proposed CSP formulation is
F. Validity of Transactions The incorporation of the safety margin threshold in the hiding process may lead to an unnecessary extension of DX.
Velammal College of Engineering and Technology, Madurai
it is possible to identify and remove the extra portion of DX that is not needed, thus minimize the size of database D to the necessary limit. To achieve that, one needs to rely on the notion of null transactions, appearing in database DX. A transaction Tq is defined as null or empty if it does not support any valid item set in the lattice. Null transactions do not support any pattern from P \ {φ}. Apart from the lack of providing any useful information, null transactions are easily identifiable, thus produce a privacy breach in the hiding methodology. They may exist due to two reasons: 1) an unnecessarily large safety margin or 2) a large value of Q essential for proper hiding. In the first case, these transactions need to be removed from DX, while in the second case the null transactions need to be validated, since Q denotes the lower bound in the number of transactions to ensure proper hiding. After solving the CSP, all the null transactions appearing in DX are identified. Suppose that Qinv such transactions exist. The size of database DX will then equal the value of Q plus the safety margin SM. This means that the valid transactions in DX will be equal to υ = Q + SM – Qinv. To ensure minimum size of DX , the hiding algorithm keeps only k null transactions, such that k = max (Q - υ, 0) ⇒ k = max (Qinv – SM, 0). As a second step, the hiding algorithm needs to ensure that the k empty transactions that remain in DX become valid prior to releasing database D to public. A heuristic is applied for this purpose, which effectively replaces null transactions of DX with transactions supporting item sets of the revised positive border. After solving the CSP in Fig. 2, the outcome is examined to identify null transactions. Then, the null transactions are replaced with valid ones, supporting item sets of Bd+ (F`D).
IV. EXPERIMENTAL EVALUATION
The algorithm was tested on different datasets using different parameters such as minimum support threshold and number/size of sensitive itemsets to hide. The thresholds of minimum support were properly set to ensure an adequate amount of frequent itemsets, among which a set of sensitive itemsets were randomly selected. We compare the solutions of the hybrid algorithm against three state-of-the-art BBAs: the BBA, the Max-Min 2 algorithm, and inline algorithm at terms of side effects introduced by the hiding process. The hybrid algorithm consistently outperforms the three other schemes, with the inline approach being the second best. An interesting insight from the conducted experiments is the fact that the hybrid approach, when compared to the inline algorithm and the heuristic approaches can better preserve the quality of the
Page 361
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
border and produce superior solutions. Indeed, the hybrid approach introduces the least amount of side effects among the four tested algorithms.
V. CONCLUSION
A novel, exact border-based hybrid approach to sensitive frequent item set hiding, through the introduction of a minimal extension to the original database was presented. A hybrid approach of combining Apriori and FP_Tree was done to find optimized frequent item set. This methodology is capable of identifying an ideal solution whenever one exists, or approximate the exact solution and the solution of this approach is of higher quality than previous approaches.
REFERENCES [1] Aris Gkoulalas – Divanis, Vassilios S. Verykois, (May 2009) “Exact Knowledge Hiding through Database Extension” IEEE Transaction on Knowledge and Data Engineering, Vol. 21, No. 5, [2] Gkoulalas-Divanis. A and Verykios. V.S, (Nov. 2006) “An Integer Programming Approach for Frequent Itemset Hiding,” Proc. ACM Conf. Information and Knowledge Management (CIKM ’06), pp. 748-757, [3] Oliveira. S.R.M. and Zaı¨ane. O.R.,, (2003) “Protecting Sensitive Knowledge by Data Sanitization,” Proc. Third IEEE Int’l Conf. Data Mining (ICDM ’03), pp. 211-218. [4] Sun. X and Yu. P.S, (2005) “A Border-Based Approach for Hiding Sensitive Frequent Itemsets,” Proc. Fifth IEEE Int’l Conf. Data Mining (ICDM ’05), pp. 426-433. [5] Verykios. V.S., Emagarmid. A.K., Bertino. E., Saygin. Y, and Dasseni. E, (Apr. 2004) “Association Rule Hiding,” IEEE Trans. Knowledge and Data Eng., vol. 16, no. 4, pp. 434-447,
Velammal College of Engineering and Technology, Madurai
Page 362
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Denial of service: New metrics and Their measurement
Dr.KannanBalasubramanian1, Kavithapandian.P2 Professor, Department of Computer Science & Engineering Mepco Schlenk Engineering College, Sivakasi, Tamilnadu, India.1 II M.E [CSE], Anna University, Thiruchirapall, Tamilnadu, Indiai.2 [email protected], [email protected]
Abstract - In this paper, we propose a novel, user-centric approach to DoS impact measurement. Our key insight is that DoS always causes degradation of service quality, and a metric that holistically captures a human user’s QoS perception will be applicable to all test scenarios. For each popular application, we specify its QoS requirements, consisting of relevant traffic measurements and corresponding thresholds that define good service ranges. We observe traffic as a collection of high-level tasks, called “transactions”. Each legitimate transaction is evaluated against its application’s QoS requirements; transactions that do not meet all the requirements are considered “failed.” We aggregate information about transaction failure into several intuitive qualitative and quantitative composite metrics to expose the precise interaction of the DoS attack with the legitimate traffic. KeywordsNetwork-level security and protection, communication/networking and information technology, computer systems organization, measurement techniques, performance of systems.
I. INTRODUCTION Denial of service (DoS) is a major threat. DoS severely disrupts legitimate communication by exhausting some critical limited resource via packet floods or by sending malformed packets that cause network elements to crash. The large number of devices, applications, and resources involved in communication offers a wide variety of mechanisms to deny service. Effects of DoS attacks are experienced by users as a severe slowdown, service quality degradation, or service disruption. DoS attacks have been studied through network simulation or testbed experiments. Accurately measuring the impairment of service quality perceived by human clients during an attack is essential for evaluation and comparison of potential DoS defenses, and for study of novel attacks. Researchers and developers need accurate, quantitative, and versatile DoS impact metrics whose use does not require significant changes in current simulators and experimental tools. Accurate metrics produce measures of service denial existing metrics agree with human perception of service denial. .
We demonstrate that our metrics meet the goals of being accurate, quantitative, and versatile 1) through testbed experiments with multiple DoS scenarios and rich legitimate traffic mixes, 2) through OPNET simulations and 3) through experiments involving human users. This project’s contributions are threefold: 1) We propose a novel approach to DoS impact measurement relying on application-specific QoS requirements. Although our proposed metrics combine several existing approaches, their novelty lies in a) the careful specification of traffic measurements that reflect service denial for the most popular applications and b) the definition of QoS thresholds for each measurement and each application class, based on extensive study of the QoS literature. 2) We aggregate multiple measurements into intuitive and informative DoS metrics that can be directly applied to existing testbed experiments and simulations, and to a variety of DoS scenarios. 3) We demonstrate that our metrics accurately capture human perception of service denial by conducting experiments with human users. II. WORK FLOW To build a network model the workflow centers on the Project Editor. This is used to create network models, collect statistics directly from each network object or from the network as a hole, execute a simulation and view results.
A. Data Flow Diagram Interactive Application
QOS Measurement
Introduce DOS
Calculate DOS Metrics
Analysis Traffic
UDP Bandwidth Flood
Analysis
Velammal College of Engineering and Technology, Madurai
Graph Representation
Page 363
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
III. Existing metrics Prior DoS research has focused on measuring DoS through selected legitimate traffic parameters: 1. Packet loss, 2. Traffic throughput or goodput, 3. Request/Response delay, 4. Transaction duration, and 5. Allocation of resources. A. Packet Loss Loss is defined as the number of packets or bytes lost due to the interaction of the legitimate traffic with the attack or due to collateral damage from a defense’s operation. The loss metric primarily measures the presence and extent of congestion in the network due to flooding attacks. It cannot be used for attacks that do not continually create congestion, or do not congest network resources at all. Examples of such attacks are pulsing attacks, TCP SYN floods attacks that target application resources, and vulnerability attacks that crash applications and hosts. Further, the loss metric typically does not distinguish between the types of packets lost, while some packet losses have a more profound impact than others (for example, a lost SYN versus data packet) on service quality. B. Throughput Throughput is defined as the number of bytes transferred per unit time from the source to the destination. Goodput is similar, but does not count retransmitted bytes. Both are meaningful for TCP-based traffic, which responds to congestion by lowering its sending rate.
C. Request/Response delay Request/response delay is defined as the interval between the time when a request is issued and the time when a complete response is received from the destination. It measures service denial of interactive applications (e.g., telnet) well, but fails to measure it for non-interactive applications (e.g., email), which have much larger thresholds for acceptable request/response delay. Further, it is completely inapplicable to one-way traffic (e.g., media traffic), which does not generate responses but is sensitive to one-way delay, loss and jitter. D. Transaction duration It is the time needed for an exchange of a meaningful set of messages between a source and a destination. This metric
Velammal College of Engineering and Technology, Madurai
depends heavily on the volume of data being transferred and whether the application is interactive and congestion sensitive. It accurately measures service denial for interactive applications such as Web browsing. For oneway traffic, such as media streaming that may not respond to congestion and runs over UDP,transaction duration will not be affected by the attack. Duration of many noninteractive transactions can be extended without causing service denial because humans expect that such transactions may occur with some delay. E. Allocation of resources Allocation of resources is defined as the fraction of a critical resource (usually bandwidth) allocated to legitimate traffic vs. attack traffic. This metric does not provide any insight into the user-perceived service quality. Instead it assumes that the only damage to legitimate traffic is inflicted by the lack of resources, and is only applicable to flooding attacks.
IV. PROPOSED METRICS We now introduce several definitions needed for DoS impact metrics. The client is the host that initiates communication with another party, which we call the server.
Definition 1. A conversation between a client and a server is the set of all packets exchanged between these two hosts to provide a specific service to the client, at a given time. A conversation is explicitly initiated by a client application (e.g., by opening a TCP connection or sending a UDP packet to a well-known service) and ends either explicitly (a TCP connection is closed, UDP service is rendered to the client) or after a long period of inactivity. If several channels are needed to render service to a client, such as FTP control and data channels, all related channels are part of a single conversation. Definition 2. A transaction is the part of a conversation that represents a higher-level task whose completion is perceptible and meaningful to a user. Transaction usually involves a single request-reply exchange between a client and a server, or several such exchanges that occur close in time. A conversation may contain one or several transactions. Definition 3. A transaction is successful if it meets all the QoS requirements of its corresponding application. If at least one QoS requirement is not met, a transaction has failed. Transaction success/failure is the core of our proposed metrics.
Page 364
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Interactive applications such as Web, file transfer, telnet,email (between a user and a server), DNS, and ping involve a human user requesting a service from a remote server, and waiting for a response. Their primary QoS requirement is that a response is served within a user-acceptable delay. Research on human perception of Web traffic delay shows that people can tolerate higher latencies for entire task completion if some data is served incrementally. Media applications such as conversational and streaming audio and video have strict requirements for low loss,low jitter, and low one-way delay. These applications further involve a media channel (where the audio and video traffic are sent, usually via UDP) and a control channel (for media control). Both of these channels must provide satisfactory service to the user. We treat control traffic as interactive traffic requiring a 4-second partial delay. Chat applications can be used for text and media transfer between two human users. While the request/response delays depend on human conversation dynamics,the receipt of user messages by the server must be acknowledged within a certain time. This delay requirement as a 4-second threshold on the round-trip time between the client and the server. Additionally, we apply the QoS requirements for media applications to the media channel of the chat application. DoS Metrics We aggregate the transaction success/failure measures into several intuitive composite metrics. Percentage of failed transactions (pft) This metric directly captures the impact of a DoS attack on network services by quantifying the QoS experienced by users. For each transaction that overlaps with the attack, we evaluate transaction success or failure applying Definition 3. A straightforward approach to the pft calculation is dividing the number of failed transactions by the number of all transactions during the attack. This produces biased results for clients that generate transactions serially.
because in some experiments it may be useful to produce a single number that describes the DoS impact. But we caution that DoS-level is highly dependent on the chosen application weights and thus can be biased.
QoS-ratio It is the ratio of the difference between a transaction’s traffic measurement and its corresponding threshold, divided by this threshold. The QoS metric for each successful transaction shows the user-perceived service quality, in the range (0, 1], where higher numbers indicate better quality. It is useful to evaluate service quality degradation during attacks. We compute it by averaging QoS-ratios for all traffic measurements of a given transaction that have defined thresholds.
The failure ratio It shows the percentage of live transactions in the current (1-second) interval that will fail in the future. The failure ratio is useful for evaluation of DoS defenses, to capture the speed of a defense’s response, and for time-varying attacks. Transactions that are born during the attack are considered live until they complete successfully or fail. Let A be a client that initiates some conversation with server B. A request is defined as all data packets sent from A to B, before any data packet is received from B. A reply is defined as all data packets sent from B to A, before any new request from A.
The DoS-hist metric It shows the histogram of pft measures across applications, and is helpful to understand each application’s resilience to the attack.
The DoS-level metric It is the weighted average of pft measures for all applications of interest: DoS-level =∑kpft(k).wk, where k spans all application categories, and wk is a weight associated with a category k. We introduced this metric
Velammal College of Engineering and Technology, Madurai
Figure illustrates request and reply identification, and measurement of partial delay, echo delay, and whole delay. We specify two types of delay requirements for e-mail, Web, telnet, and file transfer transactions where a user can utilize a partial response: 1) Partial delay measured between receipts of any two data packets from the server. For the first data packet, partial delay is measured from the end of a user request and 2) Whole delay measured from the end of a user request until the entire response has been received. Additionally, a
Page 365
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
telnet application serves two types of responses to a user: it echoes characters that a user types, and then generates a response to the user request. The echo generation must be faster than the rest of the response, so we define the echo delay requirement for telnet transactions. We identify the echo delay as the delay between a user’s request and the first response packet. V. EXPERIMENTAL SETUP
Topology Four legitimate networks and two attack networks are connected via four core routers. Each legitimate network has four server nodes and two client nodes, and is connected to the core via an access router. Links between the access router and the core have 100Mbps bandwidth and 10-40-ms delay, while other links have 1-Gbps bandwidth and no added delay. The location of bottlenecks is chosen to mimic high-bandwidth local networks that connect over a limited access link to an over provisioned core. Attack networks host two attackers each, and connect directly to core routers.
UDP flood Attack This attack can deny service in two ways: 1) By generating a large traffic volume that exhausts bandwidth on bottleneck links (more frequent variant) and 2) By generating a high packet rate that exhausts the CPU at a router leading to the target. We generate the first attack type: a UDP bandwidth flood. Packet sizes had range [750 bytes, 1.25 Kbytes] and total packet rate was 200 Kbps.
TCP SYN flood Attack Another popular attack with both attackers and researchers is the TCP SYN flood. It denies service by sending many TCP SYN packets that consume OS memory at the target. This attack can be largely countered if the target deploys the TCP SYN-cookie defense, which allocates memory only after the three-way handshake is completed.
VI. SYSTEM IMPLEMENTATION
1) Programming in OPNET
Figure Experimental topology shows the traffic patterns. Traffic patterns for IRC and VoIP differ because those application clients could not support multiple simultaneous connections. All attacks target the Web server in network 4 and cross its bottleneck link, so only this network’s traffic should be impacted by the attacks
Background Traffic Each client generates a mixture of Web, DNS, FTP, IRC, VoIP, ping, and telnet traffic. We used open-source servers and clients when possible to generate realistic traffic at the application, transport, and network level. For example, we used an Apache server and wget client for Web traffic, bind server and dig client for DNS traffic, etc. Telnet, IRC, and VoIP clients and the VoIP server were custom-built in Perl. Clients talk with servers in their own and adjacent networks.
Velammal College of Engineering and Technology, Madurai
OPNET is a simulator built on top of a discrete event system. It simulates the system behavior by modeling each event happening in the system and processes it by user-defined processes. It uses a hierarchical strategy to organize all the models to build a whole network. The hierarchy models entities from physical link transceivers, antennas, to CPU running processes to manage queues or running protocols, to devices modeled by nodes with process modules and transceivers, to network model that connects all different kinds of nodes together. OPNET also provides programming tools for us to define any type of packet format we want to use in our own protocols. Programming in OPNET includes the following major tasks: define protocol packet format, define the state transition machine for processes running the protocol, define process modules and transceiver modules we need in each device node, finally define the network model by connecting the device nodes together using user-defined link models.
Page 366
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Sample Node and Link configuration without attacks Failure ratio for transaction
Sample Node and Link configuration with attacks
Life diagram of failed transaction
Transaction cdf with respect to loss
QoS degrade measure for failed transaction
VII.CONCLUSION Our metrics are usable and they offer the real opportunity to compare and contrast different DoS attacks and defenses on an objective head-to-head basis. This work will advance DoS research by providing a clear measure of success for any proposed defense, and helping researchers gain insight into strengths and weaknesses of their solutions.
REFERENCES [1] B.N. Chun and D.E. Culler, “User-Centric Performance Analysis of Market-Based Cluster Batch Schedulers,” Proc. Second IEEE Int’l Symp. Cluster Computing and the GridProc. Second IEEE/ACM Int’l Conf. Cluster Computing and the Grid (CCGRID ’02), May 2002.
Velammal College of Engineering and Technology, Madurai
Page 367
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[2] CERT Advisory CA-1996-21 TCP SYN Flooding and IP Spoofing Attacks, CERT CC, http://www.cert.org/advisories/CA-1996- 21.html, 1996. [3] M. Guirguis, A. Bestavros, and I. Matta, “Exploiting the Transients of Adaptation for RoQ Attacks on Internet Resources,” Proc. 12th IEEE Int’l Conf. Network Protocols (ICNP ’04), Oct. 2004. [4] A. Kuzmanovic and E.W. Knightly, “Low-Rate TCP-Targeted Denial of Service Attacks (The Shrew versus the Mice and Elephants),” Proc. ACM SIGCOMM ’03, Aug. 2003. [5] S. Kandula, D. Katabi, M. Jacob, and A. Berger, “Botz-4-Sale: Surviving Organized DDoS Attacks that Mimic Flash Crowds,” Proc. Second Symp. Networked Systems Design and Implementation (NSDI), 2005.
Velammal College of Engineering and Technology, Madurai
Page 368
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
High Performance Evaluation of 600-1200V, 140A Silicon Carbide Schottky Barrier Diodes and Their Applications Using Mat Lab Manickavasagan K, P.G. Student, 401, Adhipraskthi Engineering College. Melmaruvathur, Tamil Nadu-603319. [email protected] Abstract - High performance evaluation of a 600-1200 V / 140 A range Sic schottky Barrier Diode and their applications is experimentally evaluated. The SiC Schottky Barrier Diode (SBD) is commercially available in the 600-1200 V / 1-40 A range. The main advantage of a high voltage SiC SBD lies in its superior dynamic performance in this respect: (a) The reverse recovery charge in the SiC SBD is extremely low (< 20 nC) and is the result of junction capacitance, not stored charge. Furthermore, unlike the Si PiN diode, it is independent of di/dt, forward current and temperature. (b) Higher junction temperature operation up to 175°C, (c) Reduction in the number of MOSFETs by 50%, (d) Faster switching up to 500 kHz to reduce the EMI Filter size and other passives, and (e) Reduction or elimination of the active or passive snubber. The performance of a 600 V, 4 A silicon carbide (SiC) Schottky diode is experimentally evaluated. A 300 W boost power factor corrector with average current mode control (PFC) is considered as a key application. Measurements of overall efficiency, switch and diode losses and conducted electromagnetic interference (EMI) are performed both with the SiC diode and with two ultra-fast, soft-recoveries, silicon power diodes, and the recently presented. The paper compares the results to quantify the impact of the recovery current reduction provided by SiC diode on these key aspects of the converter behavior. Keywords: (SBD) SiC Schottky Barrier Diode, (SiC) silicon carbide, Galium Arsenide (GaAs)
1.
Introduction
I should bring out my taughts and ideas through this paper in order to convey the topic. High performance evaluation of a 600-1200 V / 1-40 A range Sic schottky Barrier Diode and their applications is experimentally evaluated. Such that increase of power density is one of the main tasks for power electronics today: system sizes shall be reduced although in general the power output of the user applications increases. There are two ways to meet this challenge: 1. Reduction of losses by more efficient power electronic devices; 2. reduction of active and passive components’ number, weight and size, generally by increasing switching frequencies.
Velammal College of Engineering and Technology, Madurai
Unfortunately, Silicon (Si) bipolar diodes always show significant switching losses due to their reverse recovery behavior, especially at elevated temperatures as occurs in operation. Unipolar (“pure Schottky”) diodes on Si can only be made for voltages up to about 100V. To overcome this silicon limit, high band gap semiconductors have come into focus during the last few years. Galium Arsenide (GaAs) and Silicon Carbide (SiC) Schottky diodes have been made with breakdown voltages up to 600V and even several thousands of Volts respectively without (or more precisely with extremely small) reverse recovery. These devices are available for several years now, and their advantages have been shown in numerous applications. The power electronic systems operating in the 600-1200 voltage range currently utilize silicon PiN diodes, which tend to store large amounts of minority carrier charge in the forward-biased state. The stored charge has to be removed by majority carrier recombination before the diode can be turned off. This causes long storage and turn-off times. The prime benefits of the SiC SBD lie in its ability to switch fast (< 50 ns), with almost zero reverse recovery electron mobility than 6H-SiC. charge and the high junction temperature operation. The 600 V GaAs SBDs can be made but suffer from limitations with regards to the high junction temperature operation and 5x bigger foot-print for the same current rating. The comparable Silicon PiN diodes (Si SBDs are not viable in the 600 V range because of their large on-state voltage drops) have a reverse recovery charge of 100-500 nC and take at least 100 ns to turn-off. This places a tremendous burden on other switching elements in the system in terms of the required forward safe operating area and the switching losses incurred.
2. Electronic Properties Although there are about 170 known crystal structures, or polytypics of SiC, only two (4H-SiC and 6H-SiC) are available commercially. 4H-SiC is preferred over 6HSiC for most electronics applications because it has higher and more isotropic electron mobility than 6H-SiC. Table 1 compares the key electronic
Page 369
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
properties of 4H-SiC to Si and GaAs. The higher breakdown electric field strength of SiC enables the potential use of SiC SBDs in 600-2000 V range. Specific benefits of SiC electronic properties are: The 10x higher breakdown electric field strength of SiC reduces the specific on resistance compared to the Si and GaAs SBDs. This is illustrated in Fig. 1. At 600 V, a SiC SBD offers a Ron of 1.4 mΩ-cm2, which is considerably, less than 6.5 mΩ-cm2 for a GaAs SBD and 73 mΩ-cm2 for a Si SBD. This means that the SiC SBD will have a much smaller foot-print. The higher band gap results in much higher schottky metal-semiconductor barrier height as compared to GaAs and Si, resulting in extremely low leakage currents at elevated junction temperatures due to reduced thermionic electron emission over the barrier. The very high thermal conductivity of SiC reduces the thermal resistance of the die.
3.
SiC Schottky Diodes
Fig. 2 shows a typical temperature dependent forward characteristic of a 10 A / 600 V 4H-SiC SBD. The onresistance increases with temperature due to the reduction in the electron mobility at elevated temperatures. The diode shows 10 A at a VF of 1.5 V at 25˚C. The current reduces to approximately 5.7 A at the same VF at 200˚C. This negative temperature coefficient of forward current allows us to parallel more than one die in a package without any unequal current sharing issues. This behavior is unlike high voltage Si PiN diodes. Fig. 3 shows the reverse characteristics of the 10 A / 600 V SBD. The typical leakage current is less than 50 µA at 600 V at 25 C which increases to 70 µA at 200˚C – a very nominal increase for such a wide temperature range. The devices were packaged in plastic TO-220 packages with a thermal impedance of 1.1˚C/W. The current de-rating curve for a packaged part is shown in Fig. 4. These parts are rated for a maximum junction temperature of 175˚C. For a case temperature of up to 150˚C, the junction temperature remains below 175˚C. When the case temperature is above 150˚C, the current has to be appropriately de-rated to keep the junction temperature below 175˚C. The reverse capacitance vs. voltage curve is shown in Fig. 5. At 10 V reverse bias, the input capacitance is about 240 pF, which drops to 90 pF at 100 V and saturates to 50 pF above 300 V. This capacitance is comparable to low voltage Si Schottky Diodes. The turn-off characteristics of the 10 A/600 V 4H-SiC SBD are compared with a Si FRED at different temperatures (Fig. 6). The SiC diode, being a majority
Velammal College of Engineering and Technology, Madurai
carrier device, does not have any stored minority carriers. Therefore, there is no reverse recovery current associated with the turn-off transient of the SBD. However, there is a small amount of displacement current required to charge the Schottky junction capacitance (< 2 A) which is independent of the temperature, current level and di/dt. In contrast to the SiC SBD, the Si FRED exhibits a large amount of the reverse recovery charge, which increases dramatically with temperature, on-current and reverse di/dt. For example, the Qrr of the Si FRED is approximately 160 nC at room temperature and increases to about 450 nC at 150˚C. This excessive amount of Qrr increases the switching losses and places a tremendous burden on the switch and diode in typical PFC or motor control applications. In a switching application, the diode will be subjected to peak currents that are greater than the average rated current of the device. Fig. 7 shows a repetitive peak forward surge current of 50 A at 25˚C for the 10A / 600 V SiC SBD. This 60 Hz half sine wave measurement indicates a repetitive peak current of 5X the average. The quality of the SiC wafers has been continuously improving over the last 5 years. It is now possible to make larger area chips. Fig. 8 shows an example of a single SiC SBD chip rated at 600 V / 30 A. This device showed a leakage current of 70 µA at a reverse bias of 600 V. As was mentioned before, the negative temperature coefficient of the current makes it possible to easily parallel several chips in a single package without encountering current sharing problems. An example is shown in Fig. 9, where three chips of the type shown in Fig. 8 were packaged together to provide an 80 A / 600 V part. This package had a leakage current of 125 µA at 600 V and 25˚C. This demonstrates that the SiC SBD technology is scalable to higher currents. While the material is continuously improving, relatively high current parts can be obtained now by paralleling several chips in a package.
4. Reliability of the SiC SBD SiC is inherently a very robust and reliable material. Our 600 V SBDs have undergone extensive reliability testing and can be used as an example of SiC device reliability. To date the diodes have completed a total of 145,000 device hours of High Temperature Reverse Bias Testing (HTRB), 11,000 device hours of continuous current “burn in” testing, and 35,000 device hours of power cycle testing with no failures. The HTRB testing involved seven separate lots with test conditions of -600 volts DC at a temperature of 200˚C. The “burn in” was done at the rated device forward current, with the device junction temperature held at 200˚C. The power cycle consisted of 7
Page 370
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
minute on/off cycles (3.5 min. on / 3.5 min off) with the on-current set to the device rated current, a maximum junction temperature of 175˚C and a junction temperature delta of greater than 100˚C during the cycle.
5. Power Factor Correction (PFC) One of the largest applications for SiC Schottky rectifiers in the near future is in the continuous conduction mode (CCM) power factor correction (PFC) circuit. In traditional off-line AC-DC power supplies used in computer and telecom applications, the AC input sees a large inductive (transformer) load which causes the power factor to be substantially lower than 1. A PFC circuit allows the AC input line to see near-unity power factor, as required by new legal requirements. As shown in Fig. 10, chopping the full wave rectified input with a fast switch (MOSFET), and then stabilizing the resulting DC waveform using a capacitor accomplishes this function. When the MOSFET is ON, it is necessary to prevent the current to flow from the output capacitor or the load through the MOSFET. Hence, when the FET is ON, the Diode is OFF, and vice versa. During the switching transient when the Diode is turning OFF and the MOSFET is turning ON, the reverse recovery current from the Diode flows into the MOSFET, in addition to the rectified input current. This results in a large inrush current into the MOSFET, requiring a substantially large sized MOSFET, than that required if the Diode had no reverse recovery current. This large MOSFET represents a substantial cost in this circuit. These switching losses also limit the frequency of operation in the circuit, and hence its cost, size, weight and volume. A higher frequency would allow the size of the passive components to be correspondingly smaller. Many fast silicon rectifiers also show “snappy” reverse recovery, which results in a large EMI signature, which are also unacceptable to the new European requirements. A fast rectifier with smooth switching characteristics will allow for high efficiency PFC circuits, which also comply with new legal requirements. A 4H-SiC diode is such a rectifier. This near-zero reverse recovery SiC Schottky rectifier offers low switching losses while still showing comparable on-state performance of conventional silicon rectifiers. Due to the majority carrier transport properties of these rectifiers, they show only a capacitive current during their turn-off transient, which flows through the power MOSFET. In order to measure the benefit of these high performance rectifiers, a 250 Watt PFC test circuit was compared with an ultrafast Silicon Diode as well as SiC SBD. This test circuit used a 14 A, 500 V International Rectifier MOSFET (IRFP450), and a 6 A, 600 V ultrafast IR Si PiN diode (HFA08TB60). The input voltage was kept at a constant
Velammal College of Engineering and Technology, Madurai
120 V RMS, and the output voltage was 370 V DC. The operating frequency was 90 kHz, and the gate resistance at the MOSFET was 50 Ω. The current rating of the MOSFET was higher than the average rating to accommodate the reverse recovery current of the diode, and to maintain a high efficiency of the circuit. Under full load condition, a 600 Ω resistor was utilized, while at half load condition, 1200 Ω was used. Voltage and current measurements were taken on both the MOSFET as well as the diode, in order to estimate the power losses in these components. The input and output power was also measured to calculate the efficiency of the circuit. Under full load conditions, the temperature on the MOSFET case was measured with and without an external fan on the device. After all these measurements were taken using the ultrafast Si diode, they were repeated using Cree’s 4 A, 600 V SiC SBD. Fig. 11 shows the comparison of the switching energy losses per switching cycle in the MOSFET and Diode under half load and full load conditions. Further, the turnON and turn-OFF losses within each device are separated. Under half load conditions, the total switching losses decrease by about 25% from 266 µJ to 200 µJ when the Si diode is replaced by SiC SBD. The 50% decrease during Diode turn OFF losses, and 27% decrease during MOSFET turn ON are primarily responsible for this overall reduction in losses when a SiC SBD is used in the circuit as compared to when a Si diode is used. The MOSFET turn OFF losses and Diode turn ON losses are similar when Si and SiC Diodes are used in this circuit. Under full load conditions, Diode turn OFF losses decrease by 44%, MOSFET turn ON losses decrease by 39%, and Diode turn ON losses decrease by 29% when a SiC diode is used in this circuit as compared to a Si diode. The MOSFET turn OFF losses remain similar in both cases. An overall decrease of 27% in switching losses is measured when the circuit uses a SiC diode as compared to a Si diode. It is worth noting that diode turn ON losses are significantly lower as compared to Si PiN diodes under full load conditions because of a slower turn ON process in a PiN diode as compared to a SBD under higher current operation. These results also show that the dominant reduction in switching losses occurs due to the small reverse recovery losses in the SiC diode as compared to the case with Si diodes. Fig. 13 shows the comparison of the measured efficiency of the entire PFC circuit between Si and SiC diodes. At half load condition, the circuit efficiency increases from 88.4% with the Si diode to 95% with the SiC diode. At full load condition, the circuit efficiency increases from 90% with the Si diode as compared to 93% with the SiC diode. Ostensibly, the slightly higher on-state losses in the SiC SBD result in the relatively smaller gain in the overall circuit efficiency under full load operating condition.
Page 371
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 14 shows the measured MOSFET case temperature as a function of time after initial power up. Initially, the devices were in thermal equilibrium at room temperature. This measurement was done when a full load operating condition was impressed on this circuit. Two conditions were used for these measurements: the first was without a fixed position and speed fan, for case cooling (“in air” condition); and the second was with such a fan. With no fan, the temperature does not stabilize even after 15 minutes of the circuit power up. However, the temperature on the MOSFET was 41oC lower (86oC vs. 127oC) when a SiC SBD was used as compared to when a Si diode was used. When the fan was used for appropriate thermal dissipation, the MOSFET case temperature was only 40oC when a SiC SBD was used as compared to 50oC when a Si PiN diode was used. This increases the thermal “headroom” a circuit designer needs for more rugged operation of the circuit. Based on the measurements presented above, the most significant system advantages offered by SiC SBDs vis-à-vis Si PiN diodes in a PFC circuit are higher circuit efficiency and lower FET case temperature. These advantages can be very effectively harnessed for lowering the cost of the circuit. For a given efficiency, a higher frequency of operation of the circuit can result in smaller (and hence cheaper) inductors and MOSFETs, which are typically the most expensive components in the PFC circuit. For an identical case temperature, a smaller and cheaper MOSFET and heat sinks can be used in the circuit. Another simple circuit modification to lower the total circuit losses involves reducing the gate resistor of the MOSFET. A higher gate resistor is used in typical PFC circuit in order to limit the di/dt in the Si PiN diode, which might result in excessive reverse recovery current, and EMI emissions. Since SiC SBDs can operate under very high di/dt, a smaller MOSFET gate resistance can be utilized. Such a modification will result in lowering the MOSFET turn-OFF losses, which showed little change with direct replacement of SiC SBD with Si PiN diode in the PFC circuit described above.
FIGURES
Fig:1
Fig:2
6. TABLES AND FIGURES TABLE 1 KEY ELECTRONIC PROPERTIES OF Si,
Property
Silicon
Band gap,Eg(eV) 1.12 Electron mobility 1400 Hole mobility,µn 450 Intrinisic carrier 1.5*1010 concentration, Electron mobility, µp(cm2/vs) 1.0 Critical breakdown electicfield, Ecrit(MV/cm) 0.25 Thermal conductivity,(W/cm*K) 1.5
GaAs, and 4H-SiC GaAs 4H-SiC 1.5 92000 400 2.1*106 1.5
3.26 800 140 5*10-9 3.26
0.3
2.2
0.5
3.0-3.8
\
Velammal College of Engineering and Technology, Madurai
Page 372
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Velammal College of Engineering and Technology, Madurai
Page 373
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
7. Software In Mat Lab Programming MATLAB provides the foundation for making progress in vital areas of engineering and science. Over one million people around the world use MATLAB to help them develop cancer therapies, search for new sources of energy, make our cars safer and more fuel-efficient, and explore outer space. By combining a powerful numeric engine and technical programming environment with interactive exploration and visualization tools, MATLAB has become the language of technical computing. MATLAB is a high-level technical computing language and interactive environment for algorithm development, data visualization, data analysis, and numeric computation. Using the MATLAB product, I can solve technical computing problems faster than with traditional programming languages, such as C, C++, and Fortran. I can use MATLAB in a wide range of applications, including signal and image processing, communications, control design, test and measurement, financial modeling and analysis, and computational biology. Add-on toolboxes (collections of special-purpose MATLAB functions, available separately) extend the MATLAB environment to solve particular classes of problems in these application areas. MATLAB provides a number of features for documenting and sharing our work. I can integrate our MATLAB code with other languages and applications, and distribute our MATLAB algorithms and applications. For solving such researching tasks modern sensitive equipment and powerful software environment needed. I chose Mat lab for its advantages: Easy to use and to create programs· Big set of built-in functions for experimental data processing· External hardware interfaces · Good visualization facilities Note: All Program will Explain in conference presentation with simulation output. Example Programming as given below.
Velammal College of Engineering and Technology, Madurai
Page 374
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
% C-V Performance of Schottky diode V=[0:4]; %PhiB=[0.452 0.428 0.446]; %A=[1.6*10^(-6) 1.6*10^(-6) 1.6*10^(-6)]; q=1.6568e-025; Nd=10^(-3); C1=1.6*10^(-6)*sqrt((q*Nd)./((2*0.452)-V)); C2=1.6*10^(-6)*sqrt((q*Nd)./((2*0.428)-V)); C3=0.64*10^(-6)*sqrt((q*Nd./(2*0.446)-V)); %figure figure,plot(V,C1,':*r',V,C2,'--og',V,C3,'-.k'); xlabel('Diode Voltage (V)'); ylabel('Capacitor (C)'); legend('SBD1','SBD2','SBD3') C-V Performance of Schottky diode output
Switching Pulse For (Ms And Ms1)
Output Voltage
8. Hardware simulation Mat lab result
Output Current
Output Current
Ac Input Voltage
Velammal College of Engineering and Technology, Madurai
Page 375
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[4] F. Philippen, B. Burger, «A New High Voltage Schottky Diode based
OUTPUT VOLTAGE
th
th
on Silicon Carbide (SiC)» EPE 2001 Conf. Proc., Graz, August 27 - 29 , 2001.
8. CONCLUSION It may be concluded that SiC SBDs offer significant advantages over silicon PiN diodes in power electronic applications such as PFC. SiC SBDs are commercially available in 600-1200 V, 1-10 A range and can be utilized today to enhance the performance of the PFC circuit by improving the efficiency, reducing the switching losses in the diode and the MOSFET, reducing the MOSFET case temperature and reducing the number of MOSFETs. Additionally, they can be used to simplify or even eliminate the snubber circuits, reducing the heat sink size, or increase the frequency and reduce the size of the magnetic components. In a typical 250 W PFC circuit, an overall decrease of 27% in switching losses is measured when the circuit uses a SiC diode as compared to a Si diode. At full load condition, the circuit efficiency increases from 90% with the Si diode as compared to 93% with the SiC diode. Silicon Carbide is a wide band gap, high breakdown field material allowing high voltage Schottky diodes to be made. SiC Schottky diodes with 300, 600 and 1200-volt are commercially available at CREE. The 600-volt diodes are available with 1, 2, 4, 6, 10 and 20-amp current ratings. The 1200-volt diodes are available with 5, 10 and 20-amp current ratings. The main advantage of a high voltage SiC Schottky diode lies in its superior dynamic performance. Schottky diodes are majority carrier devices and thus do not store charge in their junctions. The reverse recovery charge in the SiC Schottky diode is extremely low and is only the result of junction capacitance, not stored charge. Furthermore, unlike the silicon PiN diode, the reverse recovery characteristics of SiC Schottkys are independent of di/dt, forward current and junction temperature. The maximum junction temperature of 175°C in the SiC Schottkys represents the actual operational temperature. The ultra-low junction charge in SiC Schottkys results in reduced switching losses in a typical hard switched CCM PFC boost converter application.
[5] M. Coyaud, J. P. Ferrieux, C. Schaeffer et al.: "Performances of SiC Schottky Rectifier in Power Factor Correction", Proc. of the 2001 IAS Annual Meeting., October, 2001, pp. 370-375. [6] M. Trivedi, K. Shenai, "Hard- and Soft-Switching Buck Converter Performance of High-Voltage 4H-SiC and Si P-i-N Diodes", Proc. of the 2001 IAS Annual Meeting., October, 2001, pp. 391-395. [7], S. Buso, G. Spiazzi: “Conducted EMI Issues in a 600-W Single Phase Boost PFC Design”, IEEE Transactions on Industry Applications, Vol. 36, No. 2, March/April 2000, pp. 578-585.
10. AUTHORS’ INFORMATION
K. MANICKAVASAGAN Place : 169, Irumbedu Village &Post, Vandavasi Taluk, Kildodungalur, Tiruvannamalai District; Tamil Nadu, India-604403. Date of Birth
: 30/07/1969
UG Studied in Name of college: Crescent Engg. College Vandalur, R.No: 92387,Chennai.Pin:600 048. Tmail Nadu, India.
9. References [1] I. Zverev, M. Treu, H. Kapels, O. Hellmund, R. Rupp, «SiC Schottky rectifiers: performance, reliability and key application», th
th
EPE 2001 Conf. Proc., Graz, August 27 - 29 , 2001. [2] M.E. Levinshtein et al., «High voltage SiC diodes with small recovery th
time», Electronics Letters, 6 July, 2000, vol. 36, no 14, pp. 1241-1242. [3] W. Wright et al. «Comparison of Si and SiC diodes during operation in th
three-phase inverter driving ac induction motor», Electronics Letters, 7 June, 2001, vol. 37, no 12, pp. 787-788.
Velammal College of Engineering and Technology, Madurai
Page 376
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Cascade Data Mining Approach for Network Anomaly Detection System Seelammal.C Department of Information Technology Kalasalingam University, India. [email protected]
Abstract— As the number of networked computers grows, intrusion detection is an essential component in keeping networks secure. Recently data mining methods have gained importance in addressing network security issues, including network intrusion detection a challenging task in network security. This paper, proposes “Heuristic+C4.5,” a method to cascade heuristic clustering and the C4.5 decision tree learning methods for classifying anomalous and normal activities in a computer network. The heuristic clustering method first partitions the training instances into clusters using similarity. The decision tree on each cluster refines the decisions by learning the subgroups within the cluster. To obtain a final decision on classification, the decisions of the heuristic and C4.5 methods are combined using nearest consensus rule. For the evaluation experiments on was conducted on 99KDDcup data set of computer networks. Results show that the detection accuracy of the Heuristic+C4.5 method is high at low false-positive-rate on KDD. Keywords—Anomaly Detection, Classification, Decision Trees, Heuristic Clustering
I. INTRODUCTION With rapid development in the computer based technology, new application areas for computer networks have emerged in Local Area Network and Wide Area Network .This became an attractive target for the abuse and a big vulnerability for the community. Securing this infrastructure has become the one research area. Network intrusion detection systems have become a standard component in security infrastructures. Intrusion detection is “the process of monitoring the events occurring in a computer system or network and analyzing them for signs of intrusions, defined as attempts to compromise the confidentiality, integrity, availability, or to bypass the security mechanisms of a computer or network”. There are generally two types of attacks in network intrusion detection. In misuse detection, each instance in a data set is labeled as ‘normal’ or ‘intrusion’ and a learning algorithm are trained over the labeled data. . An anomaly detection technique builds models of normal behavior, and automatically detects any deviation from it, flagging the latter as suspect. Attacks fall into four main categories[8]
Velammal College of Engineering and Technology, Madurai
they are Denial of Service (DoS) is a class of attacks where an attacker makes some computing or memory resource too busy or too full to handle legitimate requests, A remote to user (R2L) attack is class of attacks where an attacker sends packets to a machine over a network, then exploits machine’s vulnerability to illegally gain local access a user, User to root exploits is a class of attacks where an attacker starts out with access to a normal user account on the system and is able to exploit vulnerability to gain root access to the system. Probing is a class of attacks where an attacker scans a network to gather information or find known vulnerabilities. The main reason for using Data Mining Techniques for intrusion detection system is due to the enormous volume of existing and newly appearing network data that require for processing. The data accumulated each day by a network is huge. Several data mining techniques such as clustering, classification, and association rules are proving to be useful for gathering different knowledge for intrusion detection. Unsupervised Anomaly Detection (UAD) algorithms have the major advantage of being able to process unlabeled data and detect intrusions that otherwise could not be detected. The goal of data clustering, or unsupervised learning, is to discovery a “natural” grouping in a set of patterns, points, or objects, without knowledge of any class labels. II. RELATED WORKS Network-based computer systems play increasingly vital roles in modern society, security of network systems has become important than ever before. As an active defense technology, IDS (Intrusion Detection Systems) attempts to identify intrusions in secure architecture. In 1970s, the U.S Department of Defense outlined security goals for audit mechanisms, among which were allowing the discovery of attempts to bypass protection mechanism. Over the past years there are many different types of intrusions, and different detectors are needed to detect normal and anomalous activities in network. Some Clustering algorithms have recently gained attention in related literature, since they can help current intrusion detection systems in several respects.
Page 377
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Several existing supervised and unsupervised anomaly detection schemes and their variations are evaluated on the DARPA 1998 data set of network connections [3] as well as on real network data using existing standard evaluation techniques. Recently, anomaly detection has been used for identifying attacks in computer networks [1], malicious activities in computer systems and misuse in web systems [6], [7]. A more recent class of Anomaly Detection Systems developed using machine learning techniques like artificial neural-networks [8], Kohonen’s self organizing maps [9] fuzzy classifiers [10] and others [7] have become popular because of their high detection accuracies at low false positives. Data mining for security knowledge [2] employs a number of search algorithms, such as statistical analysis, deviation analysis, rule induction, neural abduction, making associations, correlations, and clustering. Another data mining approach [6] for intrusion detection is the application of clustering techniques for effective intrusion .How ever , the existing works that is based on either K-Means or K-Medoids have two shortcomings in clustering large network datasets namely number of clusters dependency and lacking of the ability of dealing with character attributes in the network transactions. Among these the number of clusters dependency suggests that the value of K is very critical to the clustering result and should be fixed before clustering. Trained a Hidden Markov Model [7] implemented on the trained datasets that he used to train instance-based learner. Fixed-width and k-nearest neighbor clustering techniques [4] to connection logs looking for outliers, which represent anomalies in the network traffic. A more recent class of IDS developed using machine learning techniques like artificial neural networks, Kohenen’s self organizing maps, fuzzy classifiers, symbolic dynamics [10], and multivariate analysis. However, the ADS related studies cited above have drawbacks: the majority of these works evaluate the performance of anomaly detection methods with single machine learning techniques like artificial neural-networks, pattern matching, etc., have high false positive rates. While recent advances in machine learning show that fusion [15], selection [15], and cascading [20] of multiple machine learning methods have a better performance yield over individual methods. This paper, presents a novel anomaly detection method, called “Heuristic+C4.5,” developed by cascading two machine learning algorithms: 1) the heuristic clustering and 2) the C4.5 decision tree learning. In the first stage, heuristic clustering is performed on training instances to obtain disjoint clusters. Each heuristic cluster represents a region of similar instances, “similar” in terms of Euclidean distances between the instances and their cluster centroids. In the second stage of Heuristic+C4.5, the heuristic method is cascaded with the C4.5 decision tree learning by building
Velammal College of Engineering and Technology, Madurai
an C4.5 decision tree using the instances in each heuristic cluster. For the evaluation experiments are performed on 99KDDcup dataset. Performance evaluation of the Heuristic+C4.5 cascading approach is conducted using measures: 1. Detection accuracy or true positive rate (TPR), 2. False positive rate (FPR), 3. Precision, and 4. Total accuracy (or accuracy), The performance of Heuristic+C4.5 is empirically compared with the performance of individual heuristic clustering and the C4.5 decision tree classification algorithms.
A. Contributions of the Paper The contributions of the paper are enumerated as follows. The paper presents a novel method to cascade the heuristic clustering and C4.5 decision tree learning methods for classifying data from normal and anomalous behaviors in a computer network. This paper evaluates the performance of Heuristic+C4.5 classifier, and compares it with the individual heuristic clustering and C4.5 decision tree methods using three performance measures. The paper presents a novel method for cascading two successful data partition methods for improving classification performance. From an anomaly detection perspective, the paper presents a high performance anomaly detection system. The rest of the paper is organized as follows: In Section 3, briefly discuss the heuristic clustering and C4.5 decision tree learning-based anomaly detection methods. In Section 4, present the Heuristic+C4.5 method for anomaly detection. In Section 5, discuss the experimental results and in the final section conclude the work and future directions. III.
ANOMALY DETECTION WITH HEURISTIC CLUSTERING AND C4.5 DECISION TREE LEARNING METHODS
In this section, briefly discuss the heuristic [2] clustering and the C4.5 decision tree classification [10] methods for anomaly detection.
A.
Anomaly Detection with Heuristic Clustering
This section, describes the heuristic algorithm, and illustrates how to apply this algorithm to generate detection models from audit data. Here audit data refers to preprocessed time-stamped audit records, each with a number of attributes.
1) Data Preprocessing: Depending on the monitoring data and the data mining algorithm, it may be necessary or
Page 378
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
beneficial to perform cleaning and filtering of the data in order to avoid the generation of misleading or inappropriate results. Disturbances can be noise, known changes such as dependencies on the time of day, missing data fields in some of the data records etc. 2) Notations of Some Terms: Some notations needed in heuristic clustering algorithm are Notation1: Let H = {H1 , H 2 ,...., H m } be a set of attribute values, the m is number of attribute values. Some attribute values in KDD are duration, srcbytes, dest-bytes, flag, etc.., Notation 2: Let H = H N UH s and
HN I HS = ϕ where HN is the subset of numerical attribute (e.g., no of bytes), and HS is the subset of character attribute. (e.g., service, Protocol). Notation 3: Let, ei = ( hi1 , hi 2 ,.....him ) , ei is a record, the m is number of attribute values and hij is the value of Hm Notation 4: E = {e1 , e2 ,....., en } , E is the set of records; n is the number of packets.
3) Heuristic Clustering Algorithm (HCA): In the HCA algorithms no need to fix the value of K (K-means) in the beginning of clustering. A novel heuristic clustering technology is used to determine the number of cluster automatically. For clustering, similarity between ei and every center of cluster is calculated. Sim(ei,Cj) and the similarity between every center of cluster Sim(C), if the minimal of Sim(ei,Cj) is more than the minimal of Sim(C) , create a new cluster, and the center of cluster is ei , otherwise insert the ei into Cj. Step 1. Confirm two initial cluster centers by algorithm search ( ). Step 2. Import a new record ei. Repeat 3 to 5 until no more records. Step 3. Compute the similarity by algorithm Similar (), and find
Min( Sim(ei , C j )), Min( Sim(C )). Step 4.If Min( Sim(ei , C j )) > Min( Sim(C )) then
Ck +1 = {e j }, C = {C1 , C2 ,...., C k +1} create a new cluster, and the ei is the center of the new cluster.
Velammal College of Engineering and Technology, Madurai
Else C j = C j U {ei } , insert ei into Cj. 4) The Method of Computing Similarity:The audit data consists of numerical attribute (e.g. the receive Bytes) and character attribute. (e.g. Protocol ).But whether k-Means or K-medoids all lack the ability of dealing with character attribute. The HCA algorithm resolves this problem by processing the character attribute using the method of attribute matching. The similarity of character attributes using equation (1) let ei and ej be two records in the E. all containing m attributes (including P character attributes), the nhik and nhjk is the number of hik and hjk respectively.
(nhik + nhjk ) *A hik * nkjk ) k =1 p
Sim P (ei,ej ) =
∑ (n
(1)
if (hik=hjk) then A= 0 else A=1. The similarity of numerical attribute (to the numerical attribute, still use the classical Euclidean distance to computer similarity) q
Sim N (ei , e j ) = sqrt ( ∑ | hik − hjk |2
(2)
k =1
The similarity of two records (including similarity of numerical attribute and similarity of character attribute)
Sim(ei, ej ) = Sim N (ei, ej ) + Sim P (ei, ej )
(3)
5) The Center of Cluster:A cluster are represented by its cluster center .In the HCA algorithm, the algorithm Count ( ) is used to compute the cluster center. The center of a cluster is composed of the center of numerical attributes and character attribute. Let P= (PN + PS), and P= (P1, P2,…., Pm) where PN is the center of numerical attribute, the Ps is the center of character attribute, n
P N i = (1 / n) ∑ hji
(4)
j =1
i = 1,2,...., p( p <= m) The hji is the numerical attribute. The Ps is a frequent character attribute set which consists of q (q = m-p) most frequent character attribute.
Page 379
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6) The Initial Center of Clustering: In the beginning of clustering, confirm two initial center of clustering by the algorithm Search ( ). Algorithm：Search_m (E,l). Import：data set E,the number of sampling l Output：initial center m1,m2. (1) Sampling E, get S1,S2,..,Sl (2)For i=1 to l do mi＝Count_m(Si)； (3)For i=1 to l do m＝Count_m(mi)； // m=center {m1,m2, m3……..ml} (4) m1=m , m2=max (Sim (m, mi))； 7) Detection Rule: A method is proposed to detect anomaly that does not either depend on the population ratio of the clusters. In our labeling method, assume that center of a normal cluster is highly close to the initial cluster centers vh which are created from the clustering. In other words, if a cluster is normal, the distance between the center of the cluster and vh will be small, otherwise it will be large. For each cluster center Cj (1 ≤ j ≤ k), calculate the maximum distance to vh. Then calculate the average distance of the maximum distances. If the maximum distance from a cluster to vh is less than the maximum average distance, finally label the cluster as normal. Otherwise, label as attack. Fig.1 shows how we can detect the anomaly instances.
Fig. 1 Assigning Scores to IDS
B. Anomaly Detection with C4.5 Decision Tree In this section, a brief introduction of the classification algorithms used in the IDS, i.e., the C4.5 algorithm for building decision trees.
Velammal College of Engineering and Technology, Madurai
C4.5 Algorithm: A classifier, which is a function (or 1) model) that assigns a class label to each data item described by a set of attributes, is often needed in these classification tasks. There are quite a few machine learning approaches for generating classification models, among which decision tree learning [2] is a typical one. As an example, C4.5 in [2] builds a decision tree where each internal node denotes a test on an attribute, each branch represents an outcome of the test, and leaf nodes represent classes or class distributions. The top-most node in a tree is the root node. The tree is a model generated by the classification algorithm. In order to classify an unknown sample, the attribute values of the sample are tested against the decision tree. A path is traced from the root to a leaf node that holds the class prediction for that sample. The C4.5 algorithm builds a decision tree, from the root node, by choosing one remaining attribute with the highest information gain as the test for the current node. In this paper, a later version of the C4.5 algorithm, will be used to construct the decision trees for classification. The specific algorithm is given below. The reader is referred to [2] for further details. Algorithm: Generate_decision_tree. Generate a decision tree from the given training data. Input: training samples, represented by discrete/continuous attributes; the set of candidate attributes, attribute-list. Output: a decision tree Method: (1) create a node N; (2) if samples are all of the same class, C, then (3) return N as a leaf node labeled with the class C; (4) if attribute-list is empty then (5) return N as a leaf node labeled with the most common class in samples; (6) select test-attribute, the attribute among attribute-list with the highest information gain; (7) label node N with test-attribute; (8) for each known value ai of test-attribute (9) grow a branch from node N for the condition test-attribute = ai; (10) let si be the set of samples in samples For which test-attribute = ai; (11) if si is empty then (12) attach a leaf labeled with the most common class in samples; (13) else attach the node returned by Generate_decision_tree (Si, attributelist).
Page 380
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2) Attribute Selection: The information gain measure used in step (6) of above C4.5 algorithm is used to select the test attribute at each node in the tree. Such a measure is referred to as an attribute selection measure or a measure of the goodness of split. The attribute with the highest information gain is chosen as the test attribute for the current node. This attribute minimizes the information needed to classify the samples in the resulting partitions. Such an information-theoretic approach minimizes the expected number of tests needed to classify an object and guarantees that a simple (but not necessarily the simplest) tree is found. Information Gain: Imagine selecting one case at 3) random from a set S of cases and announcing that it belongs to some class Cj. The probability that an arbitrary sample belongs to class Cj is estimated by
Pi =
freq(cj , s ) |s|
This represents the difference between the information needed to identify an element of T and the information needed to identify an element of T after the value of attribute X has been obtained. Thus, it is the gain in information due to attribute X. Gain Ratio Criterion:The notion of information gain 4) introduced earlier tends to favor attributes that have a large number of values. For example, if we have an attribute D that has a distinct value for each record, then Info(D,T) is 0, thus Gain(D,T) is maximal. To compensate for this, it was suggested in [2] to use the following ratio instead of gain. Split info is the information due to the split of T on the basis of the value of the categorical attribute D, which is defined by n | Ti | |T | SplitInfo( D) = − * log 2 i (9) | T | | T | i =1 And the gain ratio is then calculated by
∑
(5)
Where | S | denotes the number of samples in the set S and o the information it conveys is −log2 Pi bits. Suppose probability distribution P = {p1, p2,…, pn} is given then the information conveyed by this distribution, also called the entropy of P, is well known as If a set T of records is partitioned into disjoin n
Info( P) = ∑ − Pi log 2 Pi
(6)
i =1
If a set T of records is partitioned into disjoint exhaustive classes C1 ,C2 , ,Ck on the basis of the value of the categorical attribute, then the information needed to identify the class of an element of T is Info(T) = Info(P) , where P is the probability distribution of the partition (C1 ,C2 , ,Ck). First partition T on the basis of the value of a non-categorical attribute X into sets T1 ,T2 , …,Tn , then the information needed to identify the class of an element of T becomes the weighted average of the information needed to identify the class of an element of Ti , i.e. the weighted average of Info(Ti) , n |T | Info( X , T ) = ∑ i * Info(Ti ) |T | i =1
(7)
GainRatio( D, T ) =
Gain( D, T ) SplitInfo( D, T )
(10)
The gain ratio, expresses the proportion of useful information generated by split, i.e., that appears helpful for classification. If the split is near-trivial, split information will be small and this ratio will be unstable. To avoid this, the gain ratio criterion selects a test to maximize the ratio above, subject to the constraint that the information gain must be large, at least as great as the average gain over all tests examined. HEURISTIC +C4.5 METHOD FOR ANOMALY DETECTION Heuristic -based anomaly detection method are first applied to partition the training pace into k disjoint clusters C1, C2 . . . Ck. Then, an C4.5 decision tree is trained with the instances in ach Heuristic cluster. The Heuristic method ensures that each training instance is associated with only one cluster. However, if there are any subgroups or overlaps within a cluster, the C4.5 decision tree trained on that cluster refines the decision boundaries by partitioning the instances with a set of if-then rules over the feature space. The decisions of the Heuristic and C4.5 decision tree methods are combined to give a final Decision. The overall architecture of the cascade intrusion detection system is shown in fig 2. For combining the Heuristic and C4.5 decision tree methods, this paper proposes a rule called Nearest-consensus rule.
IV.
The information gain, Gain(X ,T ) , is then defined as
Gain( X , T ) = Info(T ) − Info( X , T )
(8)
Velammal College of Engineering and Technology, Madurai
Page 381
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1. TPR or recall is the percentage of anomaly instances correctly detected, 2. FPR is the percentage of normal instances incorrectly classified as anomaly, 3. “Precision” is the percentage of correctly detected anomaly instances over all the detected anomaly instances,
Training instances
Cluster
1
d1
<
DT 1
Cluster…… 2
d2 <
Cluster n
The experiment demonstrates the clustering can detect the unknown attacks and generate the new intrusion pattern
….. dn
DT 2……
The classification of normal and abnormal classes has been represented in the form of a matrix called Confusion matrix as given in Table I.
DT k
automatically. TABLE I
Fig. 2 Overall Architecture of Cascade Anomaly Detection System
CONFUSION MATRIX
A. Nearest-Consensus Rule For clusters C1,C2 . . .Cn formed by applying the heuristic clustering method on the training instances, r1, r2 . . . rn be the centroids of C1, C2, . . .Ck, respectively. Calculate the distances (d1, d2 ,…. dn ) among centroids. In the Nearest-consensus rule, the decision of the nearest candidate cluster in which there is consensus between the decisions of the Heuristic and the C4.5 decision tree methods. EXPERIMENTS AND RESULTS V. In this section discuss experiment with the heuristic clustering using the 10% subset of KDD-99 data. The data in the experiment was acquired from the 1998 DARPA intrusion detection evaluation .They set up an environment to acquire raw TCP/IP dump data for a local-area network(LAN) simulating a typical U.S.Air Force LAN. More than 200 instances of 58 attack types were launched against victim UNlX and Windows NT hosts in tree weeks of training data and two week of test data. For each TCP/IP connection, 41 various quantitative and qualitative features were extracted. Attacks fall into four main categories: DOS: denial of service R2L: unauthorized access from a remote machine U2R: unauthorized access to local super user (root) Probing: surveillance and other probing This dataset was acquired from the 1998 DARPA intrusion detection evaluation program consists of 11000 connections and 22 types of intrusions in the test set. The test set consists of 9 subtests .In heuristic clustering the relations between categorical and continuous features are handled naturally, without any forced conversions (kmeans) between these two types of features. The decision tree on each cluster refines the decisions by learning the subgroups within the cluster.
Actual \ Normal
Predicted
Abnormal
Class Normal
True positive (TP)
False negative (FN)
Abnormal
False positive (FP)
True negative (TN)
The detection rate is computed using the equations
DetectionRate =
TN * 100 ( FN + TN ) FP * 100 (TP + FP)
(12)
( FP + FN ) * 100 (TP + FP + FN + TN )
(13)
(11) FalsePositiveRate =
ErrorRate =
The experimental results obtained from the experiments conducted on 10% of the KDD dataset using Cascading without nearest consensus rule is given in Table II. The results obtained from heuristic clustering based average intrusion detection accuracy is 85.05 percent at a false-positive-rate of 10.08 percent on 10% subset of the training dataset KDDcup1999. The results obtained from C4.5 based average intrusion detection accuracy is 83.21 percent at a false-positive-rate of 9.28 percent. The experimental results obtained from three experiments conducted on 10% of the KDD dataset using Cascading with Nearest Consensus (NC) rule is given in Table III.
The following measures are used for evaluating the performance
Velammal College of Engineering and Technology, Madurai
Page 382
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
TABLE II
Iteration
RESULTS FOR CASCADE ANOMALY DETECTION WITHOUT NC RULE
Iteration
Training
Detection
False
Error
instances
Rate
Positive
Rate
Training
Detection
False
Error
Instances
Rate
Positive
Rate
Rate
Rate
0
415
90.9%
2.8%
3.4%
1
820
85.7%
2.5%
2.9%
0
415
78.1%
5.0%
7.2%
2
1200
80.0%
4.4%
5.4%
1
820
80.0%
4.4%
7.0%
3
3000
80.0%
6.4%
7.5%
2
1200
79.4%
4.2%
7.3%
4
5200
83.3%
2.4%
3.1%
3
3000
77.7%
6.8%
10.1%
5
6130
85.7%
1.9%
2.7%
4
5200
82.0%
6.3%
9.3%
6
7117
83.3%
3.1%
3.9%
5
6130
80.5%
5.9%
8.7%
7
8045
77.7%
6.1%
7.3%
6
7117
80.4%
6.4%
9.4%
8
9300
87.5%
2.8%
3.6%
7
8045
79.5%
6.7%
10.5%
9
11245
85.7%
5.3%
5.4%
8
9300
82.6%
5.5%
7.6%
9
11245
80.4%
6.4%
10.7%
95.00%
90.00%
From Table II, it is observed that the average detection rate is 83.98%, the average false positive rate is 3.77% and the average error rate is 4.52% for intrusion detection by cascading without Nearest Consensus rule using KDD dataset. From Table III, it is observed that the average detection rate is 89.98%, the average false positive rate is 3.77% and the average error rate is 4.52% for intrusion detection by cascading without Nearest Consensus rule using KDD dataset. 8.00%
85.00% Detection rate
7.00%
6.00%
80.00%
5.00% FPR 75.00% 4.00%
3.00% 70.00% 0
1
2
3
4
5
6
7
8
9 2.00%
iteration without NC rule
with NC rule 1.00%
Fig. 3 Performance of the heuristic + C4.5 method without Nearest Consensus ( NC-Rule) TABLE III RESULTS FOR CASCADE ANOMALY DETECTION WITH NC RULE
Velammal College of Engineering and Technology, Madurai
0.00% 0
1
2
3
4
5
6
7
8
9
Iteration without NC rule
with NC rule
Page 383
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig. 4 Performance of the heuristic + C4.5 method with Nearest Consensus (NC-Rule)
From Fig.3, Fig.4 it is observed that the cascading with NC has better performance in terms of Detection rate and false positive rate. CONCLUSIONS AND FUTURE WORKS VI. This paper Network intrusion detection with heuristic clustering and C4.5 incorporates the idea of applying data mining techniques to intrusion detection system to maximize the effectiveness in identifying attacks, thereby helping the users to construct more secure information systems. The main advantage of our algorithm is that the relations between categorical and continuous features in KDD are handled naturally, without any forced conversions (k-means) between these two types of features. C4.5 decision tree built on each cluster learns the subgroups within the cluster and partitions the decision space into finer classification regions; thereby improving the overall classification performance. For future work, we need to verify performance of the clustering algorithm over real data and make a new benchmark dataset for intrusion detection, because KDD Cup 1999 dataset was generated in the virtual network and the attacks included in it are greatly old-fashioned. Future works carried out for evaluating the performance of anomaly detection methods with new clustering, labeling, cascading various data mining methods with dataset drawn from various application domains.
security “, IEEE Trans.Syst.,Man and Cybern., vol.35, no.2,pp.302,Apr.2005. [6] M. Thottan and C. Ji, “Anomaly Detection in IP Networks,” IEEE Trans. Signal Processing, vol. 51, no. 8, pp. 2191-2204, May.2003. [7] C. Kruegel and G. Vigna, “Anomaly Detection of Web-Based Attacks,” Proc. ACM Conf. Computer and Comm. Security, Oct. 2003. [8] Z. Zhang, J. Li, C.N. Manikopoulos, J. Jorgenson, and J. Ucles,“HIDE: A Hierarchical Network Intrusion Detection System Using Statistical Preprocessing and Neural Network Classification,” Proc. 2001 IEEE Workshop Information Assurance, pp. 85-90, June 2001. [9] S.T. Sarasamma, Q.A. Zhu, and J. Huff, “Hierarchical Kohonen Net for Anomaly Detection in Network Security,” IEEE Trans.Systems, Man, and Cybernetics-Part B, vol. 35, no. 2, pp.450,Apr. 2005. [10] J. Gomez and D.D. Gup ta, “Evolving Fuzzy Classifiers for Intrusion Detection,” Proc. 2002 IEEE Workshop Information Assurance, June 2001. [11] A. Ray, “Symbolic Dynamic Analysis of Complex Systems for Anomaly Detection,” Signal Processing, vol. 84, no. 7, pp. 1115-1130, 2004. [12] N. Ye, S.M. Emran, Q. Chen, and S. Vilbert, “Multivariate Statistical Analysis of Audit Trails for Host-Based Intrusion Detection,” IEEE Trans. Computers, vol. 51, no. 7, pp. 810-820, 2002. [13] H.S. Javitz and A. Valdes, “The SRI IDES Statistical Anomaly Detector,” Proc. IEEE Symp. Security and Privacy, Vol.87,no.7,pp. 316326, May 1991. [14] J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas, “On Combining Classifiers,” IEEE Trans. Pattern Analysis and Machine Intelligence,vol. 20, no. 3, pp. 226-239, Mar. 1998. [15] L.I. Kuncheva, “Switching between Selection and Fusion in Combining Classifiers: An Experiment,” IEEE Trans. Systems, Man, and Cybernetics, vol. 32, no. 2, pp. 146-156, Apr. 2002. [16] R.P. Lippman, D.J. Fried, I. Graf, J. Haines, K. Kendall, D.McClung, D. Weber, S. Webster, D. Wyschogrod, R.K. Cunningham, and M.A. Zissman, “Evaluating Intrusion Detection Systems:The 1998 DARPA Off-Line Intrusion Detection Evaluation,” Proc.DARPA Information Survivability Conf. and Exposition (DISCEX ’00), pp. 12-26, Jan. 2000. [17] The third international knowledge discovery and data mining tools competition dataset KDD99-Cup, http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html, 1999.
ACKNOWLEDGEMENT This research is supported by Bose of Anna University for his support in completion of this work. Special thanks to Sam D Raj Kumar for their efforts in evaluating the system. Finally I am also very thankful to the reviewers for their detailed reviews and constructive comments, which have helped to significantly improve the quality of this paper. REFERENCES [1] S.R.Gaddam, V.V. Phoha, and K.S. Balagani, “KMeans+ID3: A Novel Method for Supervised Anomaly Detection by Cascading K-Means Clustering and ID3 Decision Tree Learning Methods ,” IEEE Trans.Knowledge and Data Eng, vol.19, no.3,pp.345, Mar.2007. [2] Zhi-Xin Yu; Jing-Ran Chen; Tian-Qing Zhu, “A novel adaptive intrusion detection system based on data mining”, Proceedings of Fourth International Conference on Machine Learning and Cybernetics Volume 4, 1821 Aug. 2005. [3] Jungsuk SONG, Kenji OHIRA, Hiroki TAKAKURA, Nonmembers, Yasuo OKABE,and Yongjin KWON, “ A Clustering Method for Improving Performance of Anomaly-Based Intrusion Detection System ,” IEICE Trans. Inf. & Syst., vol 91–d, no.5,pp.350, May 2008. [4] A. Lazarevic, A. Ozgur, L. Ertoz, J. Srivastava, and V. Kumar, “A Comparative Study of Anomaly Detection Schemes in Network Intrusion Detection” , Proceedings of SIAM International Confonference on Data Mining, May 2003. [5] Suseela T.Sarasamma , Quiming A.Zhu , and Julie Huff , “ Hierarchial KMeans+ID3: A Novel Method for Supervised Anomaly Detection by Cascading K-Meaohenen net for anomaly detection in network
Velammal College of Engineering and Technology, Madurai
Page 384
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Membrane Computing - an Overview R.Raja Rajeswari #1, Devi Thirupathi *2 #
Assistant Professor of Computer Science, M.V.M Govt Arts College(W), Dindigul, Tamil Nadu, India. 1
[email protected] *
Head Incharge, Department of Computer Applications, Bharathiar University, Coimbatore 641 046, Tamil Nadu, India. 2
[email protected]
Abstract— Membrane Computing is an emerging natural computing paradigm that emulates membrane operations on an imaginary membrane platform to solve computational problems. P Systems form the base for Membrane computing and there are three types of P Systems, Cell like P Systems, Tissue like P Systems and Neural like P Systems. Transition P Systems or P Systems, in general belong to the first category, Cell like P Systems which simulate a living cell for computing. This research paper gives an introduction to P Systems based on a preliminary survey of existing literature. Keywords— Membrane Computing, P Systems, Membrane Structure, Evolution Rules, Generative Device.
I. NATURAL COMPUTING Nature has always fascinated mankind through all its processes. Inspired by Nature, many inventions have happened: from aeroplanes that simulated birds to computers which simulate brain. Now after the buzz of Information and Communication technologies, this is the age of Nano-technology which closely resembles Nature’s technology. Nano-technology is engineering of nano particles, i.e. particles of size 1 to 100 nm (1 nm = 10 -9 m). Examples of Nano particles are atoms, Zinc oxide particles and DNA molecules. If nature is closely observed, it can be inferred that any natural process is a nano process at basic level. (e.g.) Raindrops are formed using nano particles called condensation nuclei. And a pumpkin under SEM (Scanning Electron Microscopy) will fascinate anyone with marvellous nano structures which cannot be artificially manufactured. Thus Nature has been the source of inspiration behind any man made technology.
Computer scientists were also fascinated by Nature which led to the budding of a new branch in computer science called Natural Computing or Biologically Inspired Computing. Natural computing dates back to the evolution of Neural networks and Genetic algorithms which emulate brain and genetic operators in computer science respectively, long back. Recent trends in Natural Computing saw the use of Swarm Intelligence in Soft computing and the budding of a new branch called Molecular computing. Molecular computing is a branch of Computer Science which is instigated by the biomolecular interactions in living organisms. There are two forms of molecular computing: (i) DNA computing and (ii) Membrane computing. DNA Computing DNA Computing [6] is computing with DNA strands to solve computational problems. Its history dates back to 1994, when the Computer Scientist Adleman solved the famous NP - complete problem, Hamiltonian Path Problem using DNA strands in a test tube. The advantages of DNA Computing are its massive parallelism and eco-friendly nature, which reduces the environmental hazards posed by conventional computers to nature and hence comes close to Green Computing. Its main draw backs are huge requirements of DNA and errors due to DNA operations, when it comes to practical implementation. But theoretically DNA Computing has developed well and has given solutions to NP Complete class problems of Computer Science. D.
E.
Membrane Computing The Second form of Natural computing, Membrane Computing does the computing in an imaginary membrane
Velammal College of Engineering and Technology, Madurai
Page 385
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
platform. It simulates the molecular interactions inside the cell and also the membrane operations. An initial survey on existing literature [1] infers the following results. Gheorge Paun [7] who initiated this field in 1998 defined Membrane computing as “Membrane computing is a branch of Natural computing which abstracts computing models from the structure and the functioning of living cells, as well as from the organization of cells in tissues or other higher order structures”. The next section discusses the P Systems which form the basis for Membrane Computing as defined in [8], [9] and [10]. II. P SYSTEMS P Systems form the basic models of Membrane Computing and there are three main types of P systems: 1 Cell like P systems, 2 Tissue like P systems and 3 Neural like P systems. The first type is developed with cell and its basic ingredients as the base. Tissue like P systems simulate several one membrane cells evolving in a common environment. Neural like P systems are similar to Tissue like P systems but they have also a state which controls the evolution. And cell like P systems are the ones which are addressed simply as P systems or Transition P systems in general. There are many variants of this P system like P systems with Symport / Antiport rules. But this research paper attempts only to analyze P Systems or Transition P Systems briefly. The three fundamental features of cells which are used in this computing model are: Membrane Structure, Multisets of chemical compounds and Evolution Rules which are analogous to Functions, Variables and Statements in a Programming Language. A. Membrane Structure The fundamental ingredient of a P system is the membrane structure as depicted following as a Venn diagram in Fig 1. 2 1
4
5
region 3
elementary membrane environment
Velammal College of Engineering and Technology, Madurai
Fig.1 Example
A membrane structure is a hierarchically arranged set of membranes, contained in a distinguished external membrane. Several membranes can be placed inside the skin membrane. A membrane without any other membrane inside it is said to be elementary. Each membrane determines a compartment, also called region, the space delimited from above by it and from below by the membranes placed directly inside, if any exists. Usually the membranes are identified by labels from a given set of labels. The symbolic representation of the above said membrane structures is 1 [ 2[ ] 2 3 [ ] 3 4 [5[]5]4]1. The number of membranes in a membrane structure is called the degree of the membrane structure. B. Evolution Rules In the basic P system, each region contains a multiset of symbol objects which correspond to the chemicals swimming in a solution in a cell compartment. The objects evolve by means of evolution rules [4] which are also localized, associated with the region of the membrane structure. And there are three main types of rules: 1 Multiset – rewriting rules (evolution rules) 2 Communication rules and 3 Rules for handling membranes. And Transition P systems use only the first type rules. They are of the form u Æ ν where u and ν are multisets of objects. Also we can add target indications like here, in, out to the rules indicating here is the same region, in goes immediately into a directly lower membrane and out indicates that the object has to exit the membrane, thus becoming an element of the same region surrounding it. (eg) aab Æ (a, here) (b, out) (c, here) (c, in) After using this rule in a given region of a membrane structure, two copies of a and one b are consumed and one copy of a, one of b and two of c are produced; one a and one c remain in the same region, new copy of b exits the membrane and one c enters into one of child membranes. The above said rules with atleast two objects in the L.H.S. are called cooperative rules. Rules of the form ca Æ cν, where c is a catalyst are called catalytic rules. Rules of the form a Æ v where a is an object and v is a multiset are called non cooperative. Rules of the form u Æ νδ, where δ denotes the action of membrane, dissolves the corresponding
Page 386
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
12 2 1 w1 = aac w2 = a R1 = { r1 : c---> (c, in2), r2 : c .---> (b, in2)} P1 = { r1 > r2 } R2 = {a --> (a, out), ac --> δ) P2 = Ǿ
membrane. Also in a P system rules are used in the maximally parallel manner, non deterministically choosing the rules and the objects. A sequence of transitions constitutes a computation. A computation is successful if it halts, it reaches a configuration where no rule can be applied to the existing objects, and the output region i0 still exists in the halting configuration. If the output region specified is an internal region, then it is called an internal output. And the objects present in the output region in the halting configuration is the result of the computation. If io = 0, then count the objects which leave the system during the computation and this is called as external output. A possible extension of the definition is to consider a terminal set of objects, T and to count only the copies of objects from T, present in the output region. Thus basically a P system computes a set of numbers. And a system is said to be propagating if there is no rule which diminishes the number of objects in the system. III. DEFINING TRANSITION P SYSTEMS A transition P system of degree n, n>1 is a construct π = (V, µ, w1, .... wn, (R1, P1) .....(Rn, Pn), i0), where : i) V is an alphabet; its elements are called objects µ is a membrane structure of degree n, with ii) membranes and the regions labeled in a one to one manner with elements in a given set. wi, 1 ≤ i ≤ n are strings from V* associated iii) with the regions 1, 2 .... n of µ; Ri, 1 ≤ i ≤ n, are finite sets of evolution rules iv) over V associated with the regions 1, 2… n of µ; Pi is a partial order relation over Ri 1 ≤ i ≤ n specifying a priority relation among rules of R i. An evolution rule is a pair (u,v) which denotes u Æ v where u is in V and v = v’ or v = ν’δ where v’ is a string over (V x { here, out} ) U (V x { inj /1 ≤ j ≤ n } ) and δ is a special symbol not in V. The length of u is called the radius of the rule u Æ ν v) i0, is a number betwen 1 and n which specifies the output membrane of π Example : A P system of degree 2 can be defined as follows: π = (V, µ, w1, w2, (R1, P1) , (R2, P2) , 1), V = { a, b, c, d} µ= [[ ] ]
Velammal College of Engineering and Technology, Madurai
Generally a P system can be used for solving two types of tasks: as a generative device – start from an initial configuration and collect all sets of natural numbers describing the multiplicities of objects from an output membrane or ejected from the membrane system at the end of a successful computation; and as a decidability device: introduce a problem in an initial configuration, in a specific encoding and look for the answer after a specified number of steps. The next section introduces an example for this [3]. IV. P SYSTEM AS A GENERATIVE DEVICE Consider the P system of degree 4. Π=(V, M, w1, w2, w3, w4, (R1, P1), (R2, P2), (R3, P3), (R4, P4), 4) V = { a, b, b, c, f}, M=[ [ [ ] [ ] ] ] 1 23 34 4 2 1 w1 = λ, R1 = Ǿ, P 1 = Ǿ w2 = λ R2 = {b’ ---> b, b ---> b (c, in4)}, P2 ={ r1> r2), r1 : ff --> af, r2 : f --> a δ }, w3 = af, R3 = { a--> ab’, a --> b’ δ, f --> ff}, P3 = Ǿ w 4 = Ǿ , R 4 = Ǿ , P4 = Ǿ
af a ---> ab' a ---> b' δ8 f ---> ff
b' --> b b --> b c(c, in 4 ) ff --> af > f --> a8δ
4 3
2 1
Page 387
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.2 Initial configuration
Initially w1 = Ǿ, w2 = Ǿ as shown in Fig.2. Hence start with µ3. Here the variables objects are a, f and the rules that can be applied are ……. 1 aÆ ab’ ……. 2 a Æ b’δ f Æ ff ……. 3 Using rules 1 and 3 parallely after n steps n>= 0, ab’n and 2n occurrences of f can be obtained. Now apply rules 2 & 3. That results in b’n+1 and 2n+1 occurences of f. And Membrane3 is dissolved and the objects pass onto Membrane2. At this moment rules of Membrane3 becomes inactive and rules of Membrane 2 are active. Applying rules b’Æb, ffÆaf parallely yields bn+1 and the number of f occurrences gets divided by two. In the next step, from bn+1, n+1 occurrences of c are introduced in membrane 4 using the rule b-Æb(c,in 4). At the same time, the number of f occurrences is again divided by two. This can be continued and at each step further n+1 occurrences of c are introduced in the output membrane. This can be done n+1 steps: n times when the rule ffÆaf and once when the rule fÆaδ is used. In this moment, Membrane2 is dissolved and no further step is possible. The obtained configuration is [1a2n+1bn+1[4 c(n+1)2 ]4 ]1 Consequently N(π)={m2/ m ≥ 1} Thus the above P System collects in its output membrane different values of m2, m>=1. The words “there’s plenty of room at the bottom” of Richard Feynman for Nano science, suit Membrane Computing also.
ACKNOWLEDGEMENT I record my sincere thanks to Prof. Kamala Krithivasan, Department of Computer Science and Engineering, IITM, Chennai for kindling my interest in Membrane Computing. Also I express my regards to Dr.Mrs.A.Pethalakshmi, Head, Department of Computer Science,M.V.M Govt Arts College(W), Dindigul who constantly encouraged all my research activities.
REFERENCES [10] P Systems webpage. [Online]: http://ppage.psystems.eu [11] Bernardini.F, “Membrane Systems for Molecular Computing and Biological Modelling”, Ph.D thesis, Department of Computer Science, University of Sheffield, Sheffield, 2005. [12] Krishna.S.N, “Languages of P Systems”, Ph.D thesis, Department of Mathematics, IIT Madras, Chennai, 2001. [13] Madhu.M, “Studies of P Systems as a model of cellular computing”, Ph.D thesis, Department of Computer Science and Engineering, IIT Madras, Chennai, India, 2003. [14] Oswald.M, “P Automata”, Ph.D thesis, Faculty of Computer Science TU Vienna, Vienna, Austria, 2003 . [15] Paun.Gh et al, DNA Computing New Computing Paradigms, Texts in Theoretical Computer Science, 1998. [16] Paun.Gh, “Computing with membranes”, Journal of Computer System Sciences, 2000. [17] Paun.Gh et.al.,“A guide to membrane Computing”, Theoretical Computer Science, 2002. [18] Paun.Gh, “Introduction to Membrane Computing”, Proc. Membrane Computing Workshop, 2004. [19] Paun.Gh, “Membrane Computing. An Introduction”, Springer, Berlin, 2002.
V. CONCLUSIONS Through this research paper, a preliminary review of membrane computing has been done. And membrane computing can be applied in computer science to simulate distributed computing environment and in solving NPComplete Class of problems. Also there are many variants [5] of this basic P system and some variants are used in modeling operating system. With the advent of Cloud Computing and its huge parallel processing power, P systems can facilitate development of a parallel processing language. Also in BioNano technology, Membrane Computing is used to model biological processes [2]. Still there are many avenues where membrane computing can be applied.
Velammal College of Engineering and Technology, Madurai
Page 388
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Entrustment Based Authentication Protocol For Mobile Systems R.Rajalakshmi1 R.S.Ponmagal2 1. Senior Lecturer, CSE, MAM College of Engineering 2. Lecturer, CSE, Anna University Tiruchirapalli
ABSTRACT - The aim of the paper is to describe the development of an efficient entrustment-based authentication protocol for mobile communication systems .A backward hash Chain Method is Proposed to use to ensure the authentication messages in offline authentication process in Visited Location Register (VLR), which includes the reduction in Computational Cost. The Entrustment-Based Protocol is claimed to provide non-repudiation in on-line and off-line authentication. The efficient Protocol preserves the security properties of the original scheme and reduces the computational cost in Mobile Station (MS).
authentication processes. This weakness is that any legal VLR can forge authentication messages without the help of the mobile user. The Entrustment protocol achieves non-repudiation in both on-line and off-line authentication processes, and thus avoids the weakness described above.
1. INTRODUCTION Mobile systems provide mobile users with global roaming services. To support that, numerous authentication approaches employ the public-key system to develop their protocols. A private authentication protocol to prevent the home location register(HLR) from eavesdropping on communications between the roaming station (MS) and the visited location register (VLR).Due to hardware limitations, MS cannot support heavy encryption and decryption, and therefore wastes a lot of time in exponential computations. Lee and Yeh [5] proposed a delegation based authentication protocol to solve the problems of data security, user privacy, computational loads and communicational efficiency in PCSs. Their protocol also adopted the public-key system to achieve the security requirements. To increase the communicational efficiency, and save authentication time, their protocol employs off-line authentication processes; such that VLR does not need to contact HLR frequently, and can rapidly re-authenticate MS. Therefore, compared with, the protocol of Lee and Yeh not only has a lower computational load for MS than related approaches such as GSM and MGSM [5][6][7], but also provides greater security. Though the protocol of Lee and Yeh exhibits non – repudiation in on-line authentication process, it still has a weakness in off-line
2. METHODOLOGY OF ENTRUST MENT BASED AUTHENTICATION PROTOCOL The enhanced protocol employs a backward hash chain to provide authentication and non- repudiation in off-line authentication processes. Setup: HLR and MS have their private/public key pairs (x, v) and (σ,K), respectively. The key pair (σ,K) is also stored in the MS’s SIM card. Besides, MS generates random number n1, pre-computes a hash chain h(1)(n1), h(2)(n1), . . . , h(n+1)(n1) and stores them in its database, where h(1)(n1) = h(n1) and h(i+1)(n1) =h(h(i)(n1)) for i = 1, 2, . . ., n. 2.2 On-line authentication Step 1. MS→VLR: K MS sends K to VLR. Step 2. VLR→MS: n2, IDV VLR generates a random number n2 and sends n2 and IDV to MS. Step 3. MS→VLR: r, s, K,N1, IDH, IDV MS generates a random number t, picks N1 =h(n+1)(n1) stored in its database, signs N1, n2 and IDV , and sends r, s,K,N1, IDH and IDV to VLR, where r = gt (mod p) and s = σ × h(N1_n2_IDV ) + t × r (mod q). Step 4.VLR→HLR: [N1_n2_K]KHV IDH, IDV If VLR successfully verifies the received messages by checking gs = vKK)h(N1_n2_IDV )rr (mod p), then sends [N1_n2_K]KHV , IDH and IDV to HLR. Otherwise, VLR rejects MS’s request. Step 5. HLR→VLR: [[N1, n3, IDV ]σ_n2_l_C1]KHV , IDH, IDV HLR decrypts [N1_n2_K]KHV and obtains K. If he successfully searches the corresponding σ in its database according to K, then he computes C1 = h(N1_n2_n3_σ) and l = N1, where n3 is a random number, and sends [[N1, n3, IDV ]σ_n2_l_C1]KHV , IDH and IDV to VLR. Step 6. VLR→MS: [N1, n3, IDV ]σ, IDVVLR decrypts [[N1, n3, IDV ]σ_n2_l_C1]KHV and obtains [N1, n3, IDV ]σ, n2, l and C1. Then he checks n2 and l, sets the
Velammal College of Engineering and Technology, Madurai
Page 389
Key Words: - Mobile station, security, computational cost, non- repudiation.
VLR,
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
current session key SK = C1 used by VLR and MS, and forwards [N1, n3, IDV ]σ and IDV to MS. Finally, MS decrypts [N1, n3, IDV ]σ, checks N1 and computes the current session key SK = C1. In the on-line authentication process, VLR connects with HLR when MS accesses the network through VLR and demands authentication. HLR also leaves secure authentication tokens with MS and VLR in this process. The off-line authentication process means that VLR need not contact HLR frequently, and can rapidly authenticate MS by verifying the secure token when MS accesses the network through VLR again. A newly token used for next authentication is simultaneously generated by using the old token. 2.3 Off-line authentication MS→VLR: [h(n−i+1)(n1)]Ci MS picks h(n−i+1)(n1) stored in his database and sends [h(n−i+1)(n1)]Ci to VLR for i = 1, 2,…., n, Where a predefined constant n is the limited times of off-line authentications, On receiving the authentication message from MS, VLR checks whether h(h(n−i+1)(n1)) and l are equal, updates l = h(n−i+1)(n1) and computes the session key Ci+1 = h(l,Ci). He also updates the count i = i + 1 and checks i ≤ n. For example, let n = 5. After the on-line authentication, MS stores h(4)(n1), h(3)(n1), . . . , h(1)(n1) in his/her database; VLR obtains the authentication token l = h(5)(n1); and C1 is the session key shared between MS and VLR. Subsequently, in the first off-line authentication, on receiving the authentication message [h(4)(n1)]C1 from MS, VLR checks h(h(4)(n1)) =?l and updates l = h(4)(n1) for the second offline authentication. Finally, MS and VLR have the common session key C2 = h(l,C1). Accordingly, MS and VLR are able to use the same method to execute the subsequent offline
illustrates the concept of delegation. In a business corporation, the manager uses his private key to sign a document and his staff can verify the document based on his public key. If the manager cannot sign a document because the is away on business, he can delegate his signature authority to his trustworthy assistant to sign the document without giving the assistant his private key. His staff verifies that the document is still based on his public key. This authorized signature technique is called proxy signature. This new type of digital signature gave us the inspiration for our model. The assistant is authorized to sign the document when the manager is absent, but the staff can still use the manager’s public key to verify the document. This implies that, even if the staff can distinguish the signature of the assistant from that of the manager, the staff cannot know the real identity of the assistant. The manager cannot deny the signature if a dispute arises. Of course, the manager should have the ability to identify a dishonest assistant. The previous example can be applied to our model. HLR gives MS the power to sign and VLR can verify the signature based on the public key of HLR. VLR can only verify the legality of MS but it does not know the real identity of MS. The model can provide user privacy and non_ repudiation features, and key management is easier than in the pure public-key system model since only the public key of HLR should be managed. Fig. 2 illustrates the concept of our model.
authentications.
Fig 1. Communication systems 3. CONCEPT OF DELEGATION In order to introduce our scheme, it is necessary to briefly describe the concept of delegation. Our method is inspired by the Proxy signature, which is the delegation of the power to sign messages. The following example
Velammal College of Engineering and Technology, Madurai
3.1. User Identity Privacy GSM, MGSM and public-key based protocols provide weak user identity privacy since MS must deliver his real identity to the network for authentication. In our protocol, the real identity of MS is never transmitted over the entire network for authentication purposes. Because we use pseudonym generated by HLR in the registration phase to represent the identity of MS in the network, no one except HLR can obtain any information about the identity of MS. Even VLR can only verify the legality of MS based on the public key of HLR, nothing about the identity of MS but only implies that MS is authorized by HLR. Hence, our scheme provides stronger user identity privacy than GSM, MGSM and public-key based protocols. 3.2. Non - repudiation GSM and MGSM are based on the secret-key system, so they cannot provide the feature of non - repudiation. No doubt, public-key based systems can greatly benefit by
Page 390
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
the non - repudiation feature of the public-key cryptosystem. In our scheme, each MS gets a different pair key from HLR in the registration phase. The key implies the authorization from HLR. This authorization makes VLR transfer his trust in HLR to the requested legal pseudonym MS. Because only HLR has the ability to authorize MS to sign on his behalf, HLR cannot deny this in the event a disputation occurs. Of course, HLR has the ability to identify the misused MS. Thus, our scheme can also provide the feature of non - repudiation. 3.3. Mutual Authentication between MS and VLR GSM and MGSM only provide the mechanism for VLR to authenticate MS, and the public-key based protocols can provide mutual authentication services. In our scheme, it is easy for VLR to authenticate MS by verifying the proxy signature made by MS using the proxy key authorized by HLR. If MS is authenticated by VLR. On the other hand, MS can authenticate VLR by decrypting the message received in Step 6) of the on-line authentication phase to get and checking whether is the same as what he sent to VLR in Step 3) of on-line authentication phase. Because HLR is trustworthy, only the legal VLR can decrypt the message received from HLR to get the correct. There is no way for an attacker to pretend to be a legal MS or VLR. Besides, without knowing the secret keys and, impersonating HLR is impossible. Thus, our protocol can provide mutual authentication service between MS and VLR. Furthermore, MS gets a proxy pair key from HLR over a secure channel in the registration phase. The relation between the pair key and the corresponding real identity of the MS are protected in a secure database located in the HLR. In our scheme, MS uses the proxy key to sign and VLR is responsible to verify the MS according to (3). If a MS-A acquires MS-B’s pseudonym and impersonates B to ask authentication request, (3) cannot hold since MS A has no idea about B’s proxy key. Therefore, this impersonating attack cannot succeed. In the off-line authentication process, MS generates the message and sends it to VLR. VLR decrypts the message to get. It is very difficult to compute according to, since is a one way hash function which is relatively easy to compute, but significantly harder to reverse [14]. If any attacker tries to replay this message to pass the authentication process, he cannot succeed since VLR will find out that the value does not equal to which is updated to. If any attacker tries to forge this message to pass the authentication process, he cannot succeed since the message is encrypted by a session key which is different from time to time and only known to MS and VLR. For security considerations, it is not reasonable to do off-line authentication all the time. Hence, a predefined constant should be set to a reasonable constraint on the times to do off-line authentication. 3.4. Key Management
Velammal College of Engineering and Technology, Madurai
In public-key-system-based protocols, the verification of MS is based on the public key of MS. However, it is not easy to do so in practice. Because there are many mobile users in the system, the complexity of the public-key infrastructure (PKI) will be introduced into such protocols. In our protocol, HLR authorizes MS to sign the message and VLR is merely needed to verify MS based on the public key of HLR. The number of HLR is much less than that of MS, so the complexity of PKI is dramatically reduced. Besides, theoretically, the key of HLR must be more strictly defined and protected than that of MS, and it should be a long-term key that can be used without being frequently updated. Key management will become easier than in the protocols based on public-key systems. The key management of GSM and MGSM is also easy since secret key is a long-term key kept permanently in the SIM card. If HLR has to change its public/private key pair for some security reason, he should generate a new proxy key pair for each user and send the key to the corresponding user securely. This situation is the same as when CA changes its public/private key pair [16]; the old certificates, which CA signed before, should be collected back and destroyed, and the new certificates must be generated for the users. There seems to be no good solution to avoid this ugly situation. However, compared to the original public-key-based protocols, the key management of our protocol is much easier since the number of HLR’s public key is much less than that of MS’s in public-key-system-based protocols and the public key of HLR is not necessarily updated frequently.
4. ANALYSIS BETWEEN EXISTING SYSTEM AND PROPOSED SYSTEM Previous Scheme: Token i and tokeni+1 are independent.HN can forge token i. Have not non-repudiation Charge Problem:Mobile users deny has used services and refuse to pay. Overcharge mobile users for services that he did not request. Proposed Scheme: All token i are chained by backward hash-chain and are decided by MS. HN cannot forge token I;Have non-repudiation ;PreCompute and reduce the computational cost in MS. 5. CONCLUSION The weakness of the delegation based authentication protocol which cannot exhibit non-repudiation in offline authentication processes. The proposed Entrustment authentication protocol provides Online and Offline authentication In the Entrustment based authentication protocol; MS uses a backward hash chain to guarantee
Page 391
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
that no-one can forge the authentication messages to trick HLR and provides security.
REFERENCES [1] H.-Y. Lin, “Security and authentication in PCS," Comput. Elect. Eng.,vol. 25, no. 4, pp. 225-248, 1999. [2] A. Merotra and L. Golding, “Mobility and security management in the GSM system and some proposed future improvements,” Proc. IEEE, vol. 86, pp. 14801496, July 1998. [3] M. Long, C.-H. Wu, and J. D. Irwin, “Localised authentication for internet work roaming across wireless LANs,” IEE Proc. Commun., vol. 151,no. 5, pp. 496500, Oct. 2004. [4] T.-F. Lee, C.-C. Chang, and T. Hwang, “Private authentication techniques for the global mobility network,” Wireless Personal Commun. vol. 35, no. 4, pp. 329-336, Dec. 2005 [5] W.-B. Lee and C.-K. Yeh, “A new delegation-based authentication protocol for use in portable communication systems,” IEEE Trans. Wireless Commun., vol. 4, no. 1, pp. 57-64, Jan. 2005. [6] M. Rahnema, “Overview of the GSM system and protocol architecture," IEEE Commun. Mag., pp. 92100, Apr. 1993.
Velammal College of Engineering and Technology, Madurai
Page 392
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Modeling A Frequency Selective Wall For Indoor Wireless Environment Mrs. K.Suganya1 , Dr.N.Suresh Kumar2 , P.Senthil Kumar3, 1 [email protected] 1. Assistant Professor III, Velammal College of Engg and Tech, Madurai 2.Principal, Velammal College of Engg and Tech, Madurai 3. Research Scalor , Bharath University Chennai.
ABSTRACT: People today are demanding a communications infrastructure that is both secure and mobile. Meeting both these requirements cause considerable challenge. Currently, mobile communication networks and systems are designed on the basis of detailed analysis of RF coverage and capacity requirements. Security and privacy issues can be addressed through good design, but ‘eavesdropping’ remains a real vulnerability An emerging technology named FSS (frequency selective surface) is increasingly being proposed as an answer to the deployment of secure wireless systems for indoor wireless environments, taking advantage of innovative techniques in building design and the use of attenuating materials. FSS’s could be deployed to allow certain frequencies to propagate into a room, while reflecting other frequencies. In this paper we analyzed a Frequency Selective Surface (FSS) attached onto existing common construction material at certain distance that transforms it into frequency selective filtering wall which isolates indoor wireless system from external interference.
INTRODUCTION The interference between the co-existing indoor wireless systems is becoming an important issue. Unwanted interference not only degrades system performance but also compromise security. Interference mitigation can be achieved using advance signal processing or antenna design but in indoor environments, a simpler and more effective approach may be to physically modify the propagation environment. More specifically, a building wall might be transformed into frequency selective (FS) filter which filters
cost FSwall can be created by covering a building wall with a custom designed frequency selective surface (FSS). The custom designed FSS
discussed here is constructed using conducting aluminum foil square loops, with loop dimensions shown fig.1. A bandstop response can be obtained with attenuation of about 50dB greater than that of an uncovered wall and the resonant frequency. This could permit a substantial interference reduction and thus much better signal to interference ratio (SIR), and accordingly a better system performance. However [1] also indicates that a response of a FS structure is highly dependent on substrate (i.e., the wall board) and the illumination angle.
I.
out undesired interference (e.g., 2.4GHz WLAN signals), but still allows other desired signals (e.g., cellular telephone, TV & radio broadcasting) to pass through as usual. This is the ultimate research reported in this paper [2]. In [1] it has been demonstrated that a simple and low
Velammal College of Engineering and Technology, Madurai
BFSS BWALL CFSS DFSS DWALL
AAIR
BAIR
CAIR
DAIR
AFSS AWALL CWALL
Fig.1. the matrix cascading technique, as applied to the FSS problem in this paper. Each layer is characterized by an ABCD matrix. The angle sensitivity is, to some extent a function of the FSS pattern and is under the control of the FSS design engineer. However the properties of building walls are not. For practical reason, it would be ideal if the wall and FSS cover could be modeled independently, and cascaded to
Page 393
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
obtain the combined performance. Modeling a FSS with a simple equivalent circuit has been investigated in [4].
II. FACTORS INFLUENCING THE FSS PERFORMANCE AND DESIGN The performance and behavior of the FSS filters depends on the following factors. The conductivity of the FSS conductor. The geometry of the FSS element (shape, with of conductive strip lines, proximity of conductive strip lines, thickness of conductor). The permittivity of the FSS substrate. The period of the FSS array. The number of FSS arrays when these are employed in a cascade. The electrical distance between the FSS arrays in cascade configurations. The choice of element types in hybrid FSS configuration. The finite number of periods and the metallic frames surrounding the FSS window. III. EQUIVALENT CIRCUIT METHOD Equivalent circuit modeling is the preferred modeling technique used in this research. Although it can only be used for linear polarization and is limited to certain element shapes, it offers the designer a quick and simple way to evaluate possible FSS solutions without intensive computing. EC modeling is ideal for use as an initial tool in FSS design. Each layer of square loop FSS can be represented by an equivalent circuit with an inductive component in series with a capacitive component’s. The FSS is modeled by separating the square loop elements into Vertical and horizontal strips (gratings). For TE wave incidence, the vertical strips Contribute to the inductive impedance in the equivalent circuit, where the value L is calculated according to equation 1. XL/Z0=ωL=d/p*F (p, 2s,λ,θ)………. ...(1) Similarly, the horizontal gratings correspond to the capacitive component and the C value is calculated according to equation 2. BC/ZO=ωC=d/p*4*F (p, g,λ,θ)*εeff……..( 2) p
s
d
g
c c
L
Equivalent circuit model for the square loop FSS. The field F is given by, F(p,w,λ,θ)=p/λ*cosθ[ln(cosec(πw/2p))+ G (p, w,λ,θ)] G (p, w,λ,θ)] = {0.5*(1-β²)²[(1-(β²)/4) (A+ + A_)+4β² A+ A_]} {[1(β²/4)] +β²[(1+ (β²/2) (β^4/8)] (A++A_) + 2β^6 A+ A_} A± ={1/[1±(2psin θ/λ)-(p cos θ/λ) ²]ˆ½}-1 β=sin (ωπ/2p) It was suggested that with sufficient spacing (comparable to a wavelength), the mutual influence between a wall and its FSS cover can be neglected. This paper considers the minimum spacing required between the FSS and the wall in order that they may be modeled as acting essentially independently.
IV. SIMULATION AND CASCADING MATRICES The measurements presented in this paper were simulated at S band frequency. The frequency responses of FS structure were obtained at various illumination angles. The transmission (S21) and reflection (S11) coefficient of both the FSS cover and existing wall are obtained. Simulations were preceded with different air spacing between the FSS and wall surface (0, 10 and 20mm). The FSS and wall should act sufficiently independent for simple wave propagation. V. INTERACTION BETWEEN FSS AND WALL. The simulations of FS wall performance at 0° signal incidence with air spacing is shown in fig.3. When the air spacing is 0mm fr for FS wall is at 2GHz from the designed fr at 2.4GHz which suggest that the overall response was significantly influenced by the wall structure conversely when the air is increased to 10mm or 20mm, fr did not shift from 10GHz. Therefore, 10mm air spacing provides sufficient independence to ensure insignificant interaction between FSS and the wall. VI. RESULTS The simulation carried with 10mm air spacing at 0º incidence is obtained at 2.15 GHz with 50dB attenuation.
Fig .2. Square loop FSS and the appropriate equivalent circuit.
Velammal College of Engineering and Technology, Madurai
Page 394
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.6. FSS with 0degree incidence and 10mm air spacing Fig.3. FSS alone in air At other incident angles discrepancy in fr value occurs. This suggests that at 0mm intervening air spacing, the FSS response is greatly influenced by the wall. The FSS and the wall interact with each other in a complicated manner. Such interaction detunes the performance of FSS and prevents the FS wall from being modeled by cascading the effects of the FSS and the wall individually. By increasing the air spacing to 10mm the response is obtained for the predicted frequency. At 0º incidence the
resonance occurs at 2.4GHz. In other words, the FSS characteristics are preserved, with the FSS remaining unaffected by the difficult-to-model dielectric properties of the wall. This indicates that the interaction between the FSS and the wall is almost insignificant with 10mm air spacing. More specifically, the air spacing allows the FSS and the wall to be treated independently, and to be represented by the FSS and the wall matrices. Accordingly, the proposed matrix cascading technique can be adopted for modeling such a FS wall. Increasing the air spacing to 20mm, yields no additional performance benefit over a 10mm spacing. The FSS matrix is multiplied directly with the wall matrix. The focus has been placed on determining the minimum air spacing required, so that the FSS and the wall can act independently without influencing the ‘fundamental’ response of each other. Provided that the wall is not highly reflective, the impedance transformation caused by the 10mm air section is regarded as insignificant in the investigation considered here.
VII.
Fig.4 FSS with 0degree incidence and 0mm air spacing
Fig.5. FSS with 0degree incidence and 5mm air spacing
Velammal College of Engineering and Technology, Madurai
APPLICATION OF FSS. Selective shielding of the electromagnetic interference from high power microwave heating machines adjacent to wireless communication base-stations. Selective shielding of frequencies of communication in sensitive areas (military installations, airport, police etc.). Protection from harmful electromagnetic radiation especially in the 2-3GHz band arising externally (wireless communication base stations) or internally (microwave ovens) in the domestic environment, schools, hospitals etc. Control of radiation at unlicensed frequency bands (e.g. Bluetooth applications,2.45GHz). Pico cellular wireless communications in office environments such as the Personal Handy-phone System in offices whereas to improve efficiency each room needs to prevent leakage of radio waves into another room. This implies that windows, floor and ceiling need to be shielded. Isolation of unwanted
Page 395
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
radiation. FSS windows can be incorporate in trains to prevent mobile phone frequencies. Note: that in the above applications one wishes to prevent certain frequency bands of Electromagnetic radiation to be transmitted whereas others are required to pass (frequencies related to emergency services for example). Hence the use of a broadband shielding material is not an option.
Fig8. 5mm 45deg
VIII. CONCLUSION AND FUTURE SCOPE It is concluded that 10mm air spacing is sufficient to provide independence between the FSS cover and the wall. This allows engineers to focus on designing a FSS with desired response without knowledge of the specific properties of each building/office wall. The wall material may affect the absolute attenuation but not the frequency selectivity. However, the spacing required could be reduced down to almost 0mm with an appropriate FSS design and careful choice of dielectric substrate to provide the spacing rather than air. The future works are concerned with simulating the FSS surface with various structures.
REFERENCES
Fig7. 5mm 15deg
Fig7. 5mm 30deg
Velammal College of Engineering and Technology, Madurai
[1] G.H.H.Sung, K.W. Sowerby and A.G.Williamson, “The impact of frequency selective surfaces applied to standard wall construction materials.” Proc.IEEE 2004 Antennas Propag. [2] A.Newbold, “Designing buildings for the wireless age.” Compat. Contn. Eng. J., vol.15. [3] G.H.H.Sung, K.W. Sowerby, “Modeling a lowcost frequency selective wall for wireless friendly indoor environments”.IEEE Antennas and wireless propagation letters.vol.5.2006. [4] G.H.H.Sung, K.W. Sowerby and A.G.Williamson, “Equivalent circuit modeling of a frequency selective plasterboard wall.” In Proc.IEEE 2005 Antennas Propag.Symp., vol.4A. 2005. [5] D.M.Pozar, Microwave Engineering. AdisonWesley.1990. [6] “An investigation into the feasibility of designing frequency selective windows employing periodic structure” C.Mias, C.Oswald and C.Tsakonas- The Nottingham Trent University.
Page 396
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Multi-Agent Based Personalized e-Learning Environment 13 2
T. Vengattaraman1, A. Ramalingam2, P. Dhavachelvan3, R.Baskaran4, Department of Computer Science & Engineering, Anna University, Chennai, India.
Department of Master in Computer Applications, Sri Manakula Vinayagar Engineering College, Puducherry, India. 4
Department of Computer Science, Pondicherry University, Puducherry, India.
Abstract— Personalizing e-learning services becomes more
important and it is needed for learners when e-learning takes place in an open and dynamic environment. It is of great importance to provide personalized e-learning services which can automatically adapt to the interests and levels of learners. In this paper, we propose the personalization of e-learning services with multiple feedback measures and adapting at individual level also at group level using multi agent architecture. Multiple feedback measures are collected from the user and stored in user profile. The feedbacks are adapted using an adaptation algorithm and thereby providing personalized e-learning content to the individual user and group of users. VI. INTRODUCTION E-learning is being viewed as an important activity in the field of distance and continuing education. It offers obvious advantages for e-learners making access to educational resource at any time or at any place. With the development of e-learning technologies, learners can be provided more effective learning environment to optimize their learning. Web personalization is an advanced stage in the e-learning systems evolution and it alleviates the burden of information overhead by tailoring the information presented based on an individual user’s needs. The learners have different learning styles, objectives and preferences, fact that leads to different efficiency and effectiveness from individual to individual [1]. Hence all the learners can't be treated in a uniform way. So there is a need for personalized system which can automatically adapt to the interests and levels of learners. A promising approach towards the personalization of e-learning services is user profiling. User profile includes interests, levels and learning patterns can be assessed during the learning process. Based upon the profile, personalized learning resource could be generated to match the individual preferences and levels. User profiling finds its importance in recommendation systems. Typically a recommender system compares the user profile to some reference characteristics to find similarity and differences and thereby provide suggestions to the users. For personalized search, user profiles can be
Velammal College of Engineering and Technology, Madurai
used for re-ranking the results returned to the user. Comparing to recommender systems and personalized search, user profile is important and it is needed in elearning as it is a continuous process. Feedback is the information the user receives from the system as the result of his or her action. It occurs in elearning not only in the assessment process, but can be provided to a student during navigation through learning materials, communication and collaboration with other students. Even the alerts and remainders that often appear in the e-learning system can be considered as feedback [2]. Mostly e-learning systems are personalized using either implicit or explicit feedbacks. In Implicit profiling, the user's behavior is tracked automatically by the system whereas in explicit profiling each user is asked to fill in a form when visiting the website. In our approach we are not following either of these methods. Instead we combine multiple feedback measures to get accurate and more complete user profiles for personalization. The learners with common interests and levels can be grouped, and multiple feedbacks of one person can serve as the guideline for information delivery to the other members within the same group. By adapting multiple feedbacks, personalized effective e-learning content can be delivered at the individual level and at the group level. Individual adaptation means that feedback is adapted to each student based on their characteristics. Group adaptation means adapting the feedback to common characteristics of a group of users. RELATED WORK AND LITERATURE SURVEY VII. Many approaches uses semantic Web or Web service technologies to provide dynamic as well as personalized support for learning objectives. Few approaches [3][4] are concerned with bridging learning contexts and resources by introducing semantic learning context descriptions. This allows the adaptation to different contexts based on reasoning over the provided context ontology, but does not provide solutions for building complex adaptive learning applications by reusing distributed learning functionalities. Some proposals [5] follows the idea of using a dedicated personalization web service which makes use of semantic learning object descriptions to identify and provide
Page 397
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
appropriate learning content. Integration of several distributed learning services or service allocation at runtime is not within the scope of this approach. The related research on a Personal Reader Framework (PRF) introduced in [6] and [7] allows mediation between different services based on a”connector service”. The work described in [8] utilize semantic web as well as web service technologies to enable adaptation to different learning contexts by introducing a matching mechanism to map between a context and available learning data. However, neither it considers approaches for automatic service discovery nor it is based on common standards. Hence, the reuse and automatic allocation of a variety of services or the mediation between different metadata standards is not supported [9]. Whereas the majority of the described approaches enable context-adaptation based on runtime allocation of learning data, all of them do not enable the automatic allocation of learning functionalities neither it does enable the integration of new functionalities based on open standards. Nevertheless, all approaches do not envisage mappings between different learning metadata standards to enable interoperability not only between learning contexts but also between platforms and metadata standards. VIII.
APPROACHES TO PERSONALIZATION OF ELEARNING SERVICES
Many approaches to personalization of e-learning services consider re-authoring of existing learning material. Mechanisms that facilitate the reuse of learning content increase the life span of the e-learning content but they are too expensive. To reduce the implementation overhead of the designer it is imperative to facilitate the maximum reuse of learning resources. The fulfilment of the individual needs of each user – learning personalization, educational content re-usability on large scale – content reuse, assurance of the communication between e-learning systems as well as with other human resources management systems – interoperability are the main objectives of the researches in this domain. The following subsections explain about the research works carried out by different authors in personalization of e-learning systems. A. Personalized E-learning system Blochl et al [10], proposed an adaptive learning system which can incorporate psychological aspects of learning process into the user profile to deliver individualized learning resource. The user profile is placed in multidimensional space with three stages of the semantic decisions: cognitive style, skills and user type. However,
Velammal College of Engineering and Technology, Madurai
both the means to acquire user's feedback and the algorithms to update user profile have not been addressed in the presentation. SPERO is a [12] based on the IEEE Learning Technology Systems Architecture (LTSA). It could provide different contents for the foreign language learners according their interests and levels. The problem of SPERO system is that it is largely using questionnaires and e-surveys to build user profiles, which costs the users too much extra work. B. Recommendation Systems User profiling is the key process of recommendation systems, which collect user feedback for items in a given domain and assess user profiles in terms of similarities and differences to determine what to recommend. Depending on underline technique, recommendation systems can be divided into collaborative filtering-based content-based [11] and hybrid approaches. Classified by means to acquire feedback, they can be categorized as explicit rating, implicit rating and no rating needed [11] systems. In fact, user's feedbacks are so important that only very few content-based recommendation systems require neither explicit rating nor implicit rating. For example, Surflen is a recommendation system using data mining techniques to assess the association rules on web pages through user’s browsing history without the feedbacks. However, it's hard to find user's exact interests just based on the browsing history, since it always happens that users open a page they don't like or just by mistake. This problem becomes even more severe in the situation that the system is sparsely used. Many methods of user profiling are heavily depending on the user feedbacks to construct user profiles. The feedback can be assessed explicitly by rating, or implicitly by the user behaviors such as print and save. The major drawback of existing e-learning platform is personalization with implicit feedbacks and explicit feedbacks. The proposed approach is trying to improve the personalized content delivery to the e-learners. C. Feedback Adaptation Individual adaptation means that feedback is adapted to each student and his/her individual (combination of) characteristics [12]. For example, the individual characteristics could include the user’s knowledge of the subject being studied (knowledge of the main concepts, formulas, etc) and the number of mistakes the user makes during the testing. The time and the way of feedback presentation could be personalized to these individual characteristics. For example, if the user has started to make some mistakes more often the system can present the feedback more often and include more detailed explanations in the feedback (compared to the feedback given to a user who only occasionally makes a mistake). The information that is presented in the feedback can be
Page 398
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
also personalized by relating to concepts that are already mastered by the user. Group adaptation supposes that the system adapts the feedback to common characteristics of a group of users. For example, they can be grouped according to their learning style. Immediate feedback could be presented in a brief form for active users, while detailed elaborated feedback could be presented to reflective learners. Another example of the group adaptation of feedback is personalization of the feedback to students who has passed the same courses before. The feedback could include references to the previous course (if the student has passed it) or the detailed explanation (in the case the material might be unknown to the user as s/he has not taken the course). IX. PROPOSED APPROACH In our approach, we propose the personalization of elearning services with multiple feedbacks and adapting feedbacks at individual level and group level using multiple agents. Knowledge about a user inferred from user interactions with the e-learning systems is stored in user profile and they are used to adapt offered e-learning resources and guide a learner through them. The learning material is accessible as other information sources on the web. Hypertext interlinks related pieces of information and allows the user to browse through the information space. The links are provided either explicitly, encoded by authors of the pages, or they generated automatically based on the test. Generating links automatically based on user profiles is an attractive option. However, it creates challenges as well. The data which are collected about the user can be classified into two types: static data and dynamic data. The static data are those which are not altered during the student-system interaction. The dynamic data are that changes according to the student learning progress and with the system interaction. The static data comprises five different parts. They are personal, personality, cognitive, pedagogical and preference data. Each one is an aggregation of student characteristics, which are not usually changed during an elearning session. Personal data comprises the biographical information about the student, and can be easily obtained from the course enrolment form. These information are: Student name • Student’s professional activities • • List of degree and qualifications The personality data models the student characteristics that represent the type of person the student is. These
Velammal College of Engineering and Technology, Madurai
characteristics can be inferred from personality tests. The attributes of personality data are: Personality type • • Concentration skills –based on the average time spent in the learning contents. Collaborative work skills based on the • participation in group works Relational skills based on the interactions with • students and teacher. The cognitive data models the student characteristics that represent the type of cognition the student possess. These characteristics can e inferred from cognitive tests. Based on the cognitive data, the contents can be tailored to meet the student needs. The cognitive data are Cognitive style • • Level of experience the student possesses in using the e-learning system • Student experience in using computers The pedagogical data defines the student characteristics that deal directly with the learning activity. This data intends to model the student’s behavior in learning situations, comprising two personal properties Learning style • • Learning approach The preference data stores a set of student preferences regarding the system customization. Most of the preferences are gathered from the student, but some of them are defined by the system administrator. The attributes are: Preferred presentation format • Preferred language for content display • • Web design personalization • Video speed Sound volume • The dynamic data comprises two sets of data. They are performance data and the student knowledge data. The performance data comprises level of motivation, confidence for e-learning, ability to formalize and comprehend course concepts, global performance level by the student in the course, level of effort spend by the student in the course, and portfolio that stores all the results obtained by the student in the current course. Data is constantly being gathered in order to keep an updated data and this data can be collected from the student-system interaction. The student knowledge data comprises of the concepts referred in the course, student’s progress regarding the knowledge concepts describes the knowledge that the student must possess for the current course and must possess until the end of the course. This set of data also gathers information about the student’s progress during the course sessions. This data is also gathered from the studentsystem interaction. User profile which includes user preferences contains three versions of profile data. They are:
Page 399
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
• Initial login profile • Learner’s skill profile Learner’s type profile • Initial profile formed on first time login and based on the query answered by the learner. It can be dynamically changed over a period of time, because of every action of learner is monitored and updated in learner profile. The feedbacks which are stored in user profile can be adapted to provide personalized e-learning service to the users. For adapting feedback, a feedback adaptation algorithm is proposed. X. IMPLEMENTATION The development of e-learning platform can be done with java. Basically an e-learning platform consists of login validation service module, registration service module, etest service module. Apart from this, in our project we are adapting multiple feedbacks for providing personalized elearning content to the individual user and group of users. For this we are going to propose an algorithm for adaptation of feedbacks. Our proposed system consists of profiler service, learner service, monitoring service, feedback service, content service, adaptation service. The fig. 1 depicts the system architecture using multiple agents. C. Learner service The learner service consists of login service, registration service, pre assessment test service. 1) The login service: The login service module consists of on-line registration, validation. Before taking online course, the student has to register and do the validation. The validation part consists of checking the values with the database and validating it. After validation, the student can take online course. 2) The registration service: The registration service module consists of registering the form if the user is a new user. Usually the user has to get registered with the website before taking the online course. 3) The pre assessment service: The pre assessment test module provides the on line test to the e-learners in order to check for the level of the learners in the subjects. The test marks will be calculated and his efficiency will be rated based on the marks. D. The profile service The profile service consists of two tasks. They are assessment of expertise level, providing guideline for delivery. The levels of learners are determined by the average preferences. The webpage user has read on any topic which could have different levels in terms of beginning, intermediate and advanced. The information
Velammal College of Engineering and Technology, Madurai
delivery to individual user and group of users is done by adaptation algorithm. E. Feedback extraction The feedbacks are extracted to make a final assessment of user preference. The feedbacks which are employed in system including reading time, scroll, print/save and relational index. Reading time: return 1 if user read pg longer than • Φt, where Φt is a predefined threshold; 0 otherwise. Scroll: return 1 if the number of user scrolls • (either mouse or keyboard pageup/pagedown) on pg is greater than pies, where Φs is a predefined threshold; 0 otherwise. Print/save: return 1 if user prints/saves pg; 0 • otherwise.
System Interface (GUI)
Profiler service (Agent)
Learner Service (Agent)
Content Service (Agent)
Monitoring service (Agent) Profile Database
Feedback service (Agent)
Adaptation service (Agent)
• elational index: return 1 if keywords of pg appear in user's chatting history ch more than Φt times, where Φt is a predefined threshold; 0 otherwise. Fig. 1 System Architecture
F. Feedback service Feedback service consists of collecting multiple feedbacks such reading time, no of scroll, no of print/save and relational index on chatting history and storing it in user profile.
Page 400
R
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
G. Adaptation service The feedbacks which are maintained in user profile can be adapted and provided to the individual user and group of users. The adaptation service consists of adaptation algorithm which adapts the contents and provides the personalized content to the e-learners. H. Adaptation algorithm The web pages are organized by the topics which are structured in e-content ontology. Each topic is attached with several keywords. XI. EXPERIMENTAL RESULTS A prototype system has been implemented and multiple feedback measures are recorded for feedback extractor. From the feedback we obtained the learning performance curve over various iterations and the fig.2 shows the improvements. For system training purpose, we ask a group of students to do the following experiments: Step 1: Select a topic such as “Data Structures” • and “Object Oriented Programming”, let the students indicate their levels on it in terms of beginning, intermediate, or advanced. • Step 2: Evaluate the performance. • Step 3: Prepare proficiency profile for each student and store it in database. • Step 4: Customize the learning content based on the proficiency profile. Fig. 2
measures are collected from the user and stored in user profile. The feedbacks are adapted using an adaptation algorithm and thereby providing personalized e-learning content to the individual user and group of users. The experimental results show that there is a great improvement in the learning curve. REFERENCES [20] Zhang Yin, Zhunag Yueting, WU Jiangqin, “Personalized Multimedia Retrieval in CADAL Digital Library”, Advances in Multimedia Information Processing - PCM 2008, Volume 5353/2008, pp703-712, 2008. [21] Okkyung Choi, SangYong Han, “Personalization of Rule-based Web Services”, Sensors 2008, 8(4), pp-2424-2435, 2008. [22] Knight, C., Gašević, D., & Richards, G., “An Ontology-Based Framework for Bridging Learning Design and Learning Content”, Journal of Educational Technology & Society, 9 (1), pp-23-37, 2006. [23] M. Baldoni, C. Baroglio, V. Patti, and L. Torasso, “Using a rational agent in an adaptive web-based tutoring system”, In Proc. of the Workshop on Adaptive Systems for Web- Based Education, 2nd Int. Conf. on Adaptive Hypermedia and Adaptive Web-based Systems, pp- 43-55, Malaga, Spain, 2002. [24] M. Baldoni, C. Baroglio, I. Brunkhorst, N. Henze, E. Marengo and V. Patti, “A Personalization Service for Curriculum Planning”, ABIS 2006 - 14th Workshop on Adaptivity and User Modeling in Interactive Systems, Hildesheim, pp-17-20, October 9-11 2006. [25] Henze, N., “Personalized e-Learning in the Semantic Web. Extended version of 4”, International Journal of Emerging Technologies in Learning (iJET), Vol. 1, No. 1, pp-82-97, 2006. [26] Nicola Henze, Peter Dolog, and Wolfgang Nejdl, “Reasoning and Ontologies for Personalized E-Learning”, Educational Technology & Society, Vol. 7, Issue 4. pp-82-97, 2004. [27] Schmidt, A., Winterhalter, C., “User Context Aware Delivery of ELearning Material: Approach and Architecture”, Journal of Universal Computer Science (JUCS), vol.10, no.1, pp-28-36, January 2004. [28] Simon, B., Dolog., P., Miklós, Z., Olmedilla, D. and Sintek, M., “Conceptualising Smart Spaces for Learning. Journal of Interactive Media in Education”, Special Issue on the Educational Semantic Web, ISSN:1365-893X [http://wwwjime. open.ac.uk/2004/9], 2004. [29] Martin Blöchl, Hildegard Rumetshofer, Wolfram, “Individualized E-Learning Systems Enabled by a Semantically Determined Adaptation of Learning Fragments”, Proceedings of the 14th International Workshop on Database and Expert Systems Applications, pp-640, 2003. [30] Xiaobin Fu, Jay Budzik, Kristian J. Hammond, “Mining navigation history for recommendation”, Proceedings of the 5th international conference on Intelligent user interfaces, United States, pp-106 - 112, 2000. [31] Xin Li and S. K. Chang, “A Personalized E-Learning System Based on User Profile Constructed Using Information Fusion”, The eleventh International Conference on Distributed Multimedia Systems (DMS'05), Banff, Canada, pp. 109-114, Sep.2005.
Learning Improvement Graph
The experimental results show improved learning curve. We used scoring in the range of 1 to 5 for each student; the overall performance of the students improves in the subsequent iterations. CONCLUSION XII. This paper proposes the personalization of e-learning services with multiple feedback measures and adapting at individual level also at group level. Multiple feedback
Velammal College of Engineering and Technology, Madurai
Page 401
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Identification in the E-Health Information Systems Ales Zivkovic University of Maribor FERI Maribor [email protected] Abstract Fast development of health informatics promise to offer new services supported with information technology. Different stakeholders in the public health care systems are already using partially integrated solutions in order to exchange electronic health records. Patient identification is one of the key elements in the future electronic health systems. While technology provides several solutions for patient identification, processing of personal data related to health is strictly regulated. The paper addresses some of the most common issues in regard to privacy, discuss technical opportunitiesin the identification problem space and presents current identification solution in Slovenia.
1. Introduction The use of most information systems depends on user identification. The purpose of identification is representation of the user or another system to the information system being used. Based on the identification, the information system will allow or deny access to the system or one of its parts. With the integration of information systems, high availability of e-services and ubiquitous solutions, the importance of identification mechanisms and efficient authorization schemes is also becoming more and more important. The identification mechanisms should guarantee first level security to the information systems as well as the user, while the authorization prevents unauthorized access to the information, data and services. The costs of public health care schemes are substantially increasing and governments are calling for new strategies [1]. One of the solutions is open integrated electronic health system that is able to exchange the information between all stakeholders in the public health care system. The key stakeholders are:hospitals, medical centers, health centers, insurance companies, government, patients, employers, pharmacies, drugstores, pharmaceutical companies, health care professionals (i.e. doctors, nurses, pharmacists) and others. The central part of the electronic health system is electronic health record (EHR), the electronic document that contains the comprehensive data of the past and present physical and mental state of health of an individual. The EHR is highly sensitive in regard to personal data protection as well as confidentiality, safety and accuracy. The identification plays an important part in fulfilling these requirements. The three basic principles of identification are [2]:
Velammal College of Engineering and Technology, Madurai
Something you have - the user poses something • that is then used for identification since other people do not have this. For example EHR smart card, hardware key generator andpublic key certificate. Something you know - the identification element • is secret information that is known only to the user of the system. For example password, paraphrase and Personal Identification Number (PIN). Things you are - the principle is based on the • physical characteristics or behavior of the user that can be used for identification. For example voice recognition, fingerprint, biometric signature, palm and hand print recognition, iris and retina recognition. The most common principle in use today is "something you know" since passwords and PIN numbers are widely used for computer and system login while ATM machines and mobile phones require PIN identification. The principle is simple to implement, easy to use and inexpensive. The only disadvantage is password vulnerability. Therefore it is better to implement and use other identification mechanisms. The paper is organized as follows. The second section gives a brief overview of the standards and regulations in health informatics and identification. In section three, technical solutions for user identification are described and compared. Section four gives an overview of the recent identification solution in Slovenian public health system. Section five concludes the paper.
2. Standards and Regulations There are many standards, laws and directives related to protection of personal data, security, identification, health information systems and biometrics. The aim of this section is to point out some of the related standards and regulations and provide the overview of the legal framework in Slovenia. The ISO/DIS 21091:2008 Health informatics - Directory services for security, communications and identification of professionals and patients [3] standard is aimed at providing key requirements for secure communication of the health care professionals in conducting clinical and administrative functions. The standard extends the public key infrastructure with additional requirements for its efficient use within the health care transactions. The health care directory includes the information about valid certificates, certificate revocation lists (CRL), identification
Page 402
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
of individual roles within the health care system, trust chains and secure communication over the public networks. The directory services are using the X.500 framework. The specification also defines the directory security managementframework where the following standards should be considered for protection of the directory infrastructures: ISO 22600-2, ISO/IEC TR 13335-1 and Control Objectives for Information and Related Technologies (COBIT) [4]. According to the Technical specification CEN/TR 15872 Health informatics - Guidelines on patient identification and cross-referencing of identities [5], the accuracy of identification and information is essential requirement within the healthcare information system. The individual needs to be uniquely identified in order to ensure appropriate and relevant care. Currently the patient may have several identifiers (i.e. domestic health care identifier, international insurance identifier while traveling, temporary identifiers for various medical services outside the public health care system) corresponding to different geographical locations, different health care organizations and services. Consequently, different identifiers increase the risk of identification error and potentially compromise the patient's safety. The quality of identification ensures that health care professionals have access to patient information, facilitating closer coordination and continuity of care and improving medical service in terms of preventive as well as follow-up activities. In the technical report, the solution for providing continuity of medical care and patient data exchange that is based on the reliable consolidation of different identities in the patient identifier domain. It also provides solutions for cross-referencing of identities when the medical professionals needs to access all patients data that resides in different information systems managed by different health organizations. The main legal instrument of EU data protection law Directive 95/46/EC [6], defines health information as sensitive data. Sensitive data cannot be processed unless the subject of the data (the patient in case of the EHR) gives explicit consent, or another exemption applies. Article 8 of the Directive expressly provides that Member State laws may prohibit the processing of certain sensitive data, irrespective of the individual's consent. Without obtaining consent, organizations may only process the following data: Data necessary for exercising the organization's • obligations or rights related to employment. Data necessary to protect the vital interests of the • individual when the individual is physically or legally incapable of giving consent. Data necessary for the establishment, exercise or • defense of legal claims. Data necessary for the purposes of preventive • medicine, medical diagnosis, the provision of care or treatment or the management of health-care services, and where those data are processed by a health professional
Velammal College of Engineering and Technology, Madurai
subject to professional secrecy or by another person subject to an equivalent obligation of secrecy. Data that were clearly made public by the • individual itself. In addition, the processing of sensitive data generally requires prior approval from national data protection authorities. In Slovenia, the law requires the approval of the Information Commissioner. In Italy a detailed security policy document is required, and specific technical requirements must be met. In Spain, the processing of health-related data triggers a requirement for more rigorous security measures under Royal Decree 994/1999. After the Member States have implemented the exceptions differently and inconsistently, theEuropean Commission Report on the implementation of the Data Protection Directive (95/46/EC [6]) recognized the problem and the Data Protection Working Party, an independent European advisory body on data protection and privacy, issued the Working Document on the processing of personal data relating to health in electronic health records (EHR).The document provides guidelines on the interpretation of the applicable data protection legal framework for EHR systems, presents the general principles and recommendations on eleven topics namely respecting self determination, identification and authentication of patients and health care professionals, authorization for accessing EHR in order to read and write in EHR, use of EHR for other purposes, organizational structure of an EHR system, categories of data stored in EHR and modes of their presentation, international transfer of medical records, data security, transparency, liability issues and control mechanisms for processing data in EHR. In 2004 Slovenia adopted the Personal Data Protection Act [7] that determines the rights, responsibilities, principles and measures to prevent unconstitutional, unlawful and unjustified encroachments on the privacy and dignity of an individualin the processing of personal data. The enforcement of the law is under the supervision of the Information Commissioner of Republic of Slovenia. In 2008 the commissioner issued two important guidelines [8]: Guidelines for safeguarding the personal data in 1. the Hospital Information Systems (HIS) and Guidelines regarding the use of biometric 2. technologies. In the first guideline the commissioner points out that the logging of user interaction could be at different levels: (1) change log, (2) data access log and (3) full audit trail. While the law does not require the audit trail, technically it is possible to implement it within the HIS. The second level is mandatory for sensitive data and must be implemented regardless of the circumstances. The use of group login accounts is not advisable and should be used only in special cases (i.e. emergency team). However, for these exceptions additional measures should take place
Page 403
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
making clear who were the members of the emergency team in the particular time frame. The second guideline first explains the difference between identification and authentication. When implemented with biometric techniques, both approaches are treated equally and may be used only when it is impossible to fulfill the same aim with other means. The use must be regulated. In the public sector the biometric solutions may be used for entering working premises or work registration. In private sector, the use is allowed only when higher security measures are required for protecting people or trade secrets. It can be used only for company employees with prior written notice. Before introducing the biometric, the company must acquire approval from the government supervisory office. According to the Slovenian and EU laws the biometric data are always treated as sensitive personal data.
3. Technical Solutions Something you know (usernames and passwords, PIN) is the most widely used identification principle for accessing IT resources today. The administrator or the system itself, using the algorithm, sets the username and password for the new user. The user can later partially (only password) or fully (username and password) change it. This identification principle has several drawbacks: Reliability of identification - when the system • receives the correct username and password it only knows that the identification data matches the data in the database. It can not guarantee that the person that entered the data to the system is the right person. The data could be stolen or the user gives it to someone else in order to perform the task for him. Insufficient control - even if the company • introduced the password policy, there are many problems when used in practice (i.e. the password length, complexity, the password change frequency, storing the passwords). The compromised password detection - usually it • is quite difficult to detect that the password was compromised. The attacker could use the stolen data for days, months or even years before anybody founds out. The compensation control would be to change the password regularly to prevent long time exposure. The system could also try to recognize unusual logins (after the premises were closed, weekends). 3.1 Biometric identification The biometric identification could be divided into two subgroups: • Physiological biometry - includes the methods that use physiological properties of the person like fingerprints, iris and retina recognition and palm geometry. The variability of the properties in this group is very low. Behavioral biometry - uses person's behavioral • patterns. In this group are voice, typing and handwriting. The biggest problem of the methods in this group is high
Velammal College of Engineering and Technology, Madurai
variability of the property being measured for identification. The recognition accuracy is also lower in comparison to the physiological biometry. The variability could be due to the changes a person is going through (i.e. illness, stress). The advantage is better user acceptance. The biometric identification has several advantages: the biometric properties are unique, • • the property can not be transferred to another person, it can not be forgotten, lost or stolen, • • it is difficult to copy or forge, • it can be used when person is unconscious and it is hard for someone to hide or change the • property. Therefore, the use of biometric identification could be more appropriate in healthcare systems for some users (for example children, senior users and people with disabilities) and special use cases (for example emergency). 4. Identification in the Slovenian Health System In the year 2000 the Health Insurance Institute of Slovenia (ZZZS) [9] introduced the Slovene Health Insurance Card (KZZ) [9] as the official document applied in the implementation of the rights deriving from the compulsory and voluntary health insurance in Slovenia. Slovenia was the first country to introduce an electronic card at a national scale within the EU. The common EU member countries objective is to introduce an electronic document applicable within a country and across its borders. The Slovenian card is issued, free of charge, to every person upon the first regulation of the compulsory health insurance status in Slovenia and is used for the identification of the policy owner, the patient, as well as the health service providers. The card is made of plastic, measures 8.5 x 5.5 centimeters and has the chip with the following data: thecard owner data, • • thedata about the mandatory insurance policy, thedata about the voluntary health insurance, • • the data about the personal doctor, • the data about the medical accessories, the data about the owner decision regarding the • organ donation in case of death, the data about the issued medicine and • • the payer's data. The access to the data is granted only together with the authorized professional card owned by the medical staff (doctors, nurses, physicians, chemist). In October 2008 the ZZZS issued the new KZZ card that will replace the old KZZ card.
Page 404
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Figure 1: The new Slovene Health Insurance Card (KZZ) with the reader on the right In addition to the data on the old KZZ card, the new KZZ card includes the digital certificates issued by the ZZZSCA and has additional space for two additional personal certificates (2x2048 bits). The default certificate issued by the ZZZS is used for accessing the health insurance on-line system, while the additional personal certificates could be used for accessing other resources on the web. Figure 1 shows the new card (left) and the card reader (right).
Figure 2: The software for the management of the new Slovene Health Insurance Card On Figure 2 the software for the managements of the KZZ card is presented. The screenshot shows the available space on the card, the certificates status,the card type and the data stored on the card. While the old card was used as the data carrier, the new card is used as the identification element the key for accessing the health data. This change enables on-line access to the patients' data using digital certificates. In the future the card will be integrated with the Electronic Health Record (EHR), e-prescription and authorizations within the health information systems (i.e. patients waiting lists) as well as integration with the identity card. 5. Conclusion The identification is becoming one of the most important elements of the future IT-based solutions. The paper discusses legal as well as technical issues of the identification in the health care domain. Slovenia was the first EU country that introduced the electronic health
Velammal College of Engineering and Technology, Madurai
insurance card on the national scale. The second-generation card was recently put into production enabling better security and introduction of new e-health services. While the biometric identification might be better solution for the healthcare domain, the strong national and EU regulations prevents IT solution providers to use the technology in the Health Information Systems (HIS). Consequently, the Health Insurance Institute of Slovenia decided to use smartcard solution together with the digital certificates. In the future the integration with the identity card is anticipated.
References 1. Working Document on the processing of personal data relating to health in electronic health records (EHR), WP131, 2007 Ramesh Subramanian, Computer Security,Privacy 2. and, Politics-Current Issues Challenges and Solutions, IRM Press,2008 ISO/DIS 21091:2008 Health informatics 3. Directory services for security, communications and identification of professionals and patients ISACA, Control Objectives for Information and 4. related Technology(COBIT), version 4.1, IT Governance Institute, 2007 CEN/TR 15872 Health informatics - Guidelines 5. on patient identification and cross-referencing of identities Directive 95/46/EC of the European Parliament 6. and of the Council, http://ec.europa.eu/justice_home/fsj/privacy/docs/95-46ce/dir1995-46_part1_en.pdf (accessed March 19, 2010) Personal Data Protection Act (Slovenia), 7. http://ec.europa.eu/justice_home/fsj/privacy/docs/impleme ntation/personal_data_protection_act_rs_2004.pdf (accessed March 19, 2010) Information Commissioner (Slovenia), 8. http://www.ip-rs.si/ (accessed March 19, 2010) Health Insurance Institute of Slovenia, 9. http://www.zzzs.si/indexeng.html
Page 405
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Privacy Preserving Distributed Data Mining Using Elliptic Curve Cryptography M.Rajalakshmi#1, T.Purusothaman*2 #
Department of CSE/IT, Senior Grade Lecturer, Coimbatore Institute of Technology, Coimbatore,Tamilnadu, India 1
[email protected]
*
Department of CSE/IT, Assistant Professor,Government College of Technology, Coimbatore, Tamilnadu, India 2
[email protected]
Abstract— Association rule mining is one of the most investigated fields of data mining. When it occurs on distributed data, privacy of participating parties becomes great concerns. Distributed association rule mining is an integral part of data mining aimed at extracting useful information hidden in distributed data sources. As local frequent itemsets are globalized from various data sources, sensitive information pertaining to the individual data source needs high protection. Different privacy preserving data mining approaches for distributed environment have been proposed but they all suffer from privacy, computation complexity and communication overhead. The proposed method finds global frequent itemsets in a distributed environment with minimal communication among sites and ensures higher degree of privacy with Elliptic Curve Cryptography(ECC). The experimental results shows that ECC is more secure than RSA cryptography and generates global frequent itemsets without affecting mining performance and confirms optimal communication among sites Keywords— distributed data mining, privacy, secure multiparty computation, frequent itemsets, sanitization, cryptography
1. Introduction Major technological developments and innovations in the field of information technology have made it easy for organizations to store a huge amount of data within its affordable limit. Data mining techniques come in handy to extract valuable information for strategic decision making from voluminous data which is either centralized or distributed [1], [11]. The term data mining refers to extracting or mining knowledge from a massive amount of data. Data mining functionalities like association rule mining, cluster analysis, classification, prediction etc. specify the different kinds of patterns mined. Association Rule Mining (ARM) finds interesting association or correlation among a large
Velammal College of Engineering and Technology, Madurai
set of data items. Finding association rules among huge amount of business transactions can help in making many business decisions such as catalog design, cross marketing etc. A best example of ARM is market basket analysis. This is the process of analyzing the customer buying habits from the association between the different items which is available in the shopping baskets. This analysis can help retailers to develop marketing strategies. ARM involves two stages i) Finding frequent itemsets ii) Generating strong association rules
1.1 Association Rule Mining: Basic concepts Let I = {i1,i2…im} be a set of m distinct items. Let D denote a database of transactions where each transaction T is a set of items such that T ⊆ I. Each transaction has a unique identifier, called TID. A set of item is referred to as an itemset. An itemset that contains k items is a k-itemset. Support of an itemset is defined as the ratio of the number of occurrences of the itemset in the data source to the total number of transactions in the data source. Support shows the frequency of occurrence of an itemset. The itemset X is said to have a support s if s% of transactions contain X. The support of an association rule XÆY is given by Support = (Number of transactions containing XUY)/(Total number of Transactions) where X is the antecedent and Y is the consequent. An itemset is said to be frequent when the number of occurrences of that particular itemset in the database is larger than a user-specified minimum support. Confidence shows the strength of the relation. The confidence of an association rule is given by, Confidence = (Number of transactions Containing XUY) (Total number of Transactions containing X)
Page 406
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
1.2 Distributed Data Mining In the present situation, information is the key factor which drives and decides the success of any organization and it is essential to share information pertaining to an individual data source for mutual benefit. Thus, Distributed Data Mining (DDM) is considered as the right solution for many applications, as it reduces some practical problems like voluminous data transfers, massive storage unit requirement, security issues etc. Distributed Association Rule Mining (DARM) is a sub-area of DDM. DARM is used to find global frequent itemsets from different data sources distributed among several sites and interconnected using a communication network. In DARM, the local frequent itemsets for the given minimum support are generated at the individual sites by using data mining algorithms like Apriori, FP Growth tree, etc.[1], [11]. Then, global frequent itemsets are generated by combining local frequent itemsets of all the participating sites with the help of distributed data mining algorithm [5], [2]. The strong rules generated by distributed association rule mining algorithms satisfy both minimum global support and confidence threshold. While finding global frequent itemset, local frequent itemsets at individual sites need to be collected. Due to that the participating sites know the exact support count of itemsets of all other participating sites. However, in many situations the participating sites are not interested to disclose the support counts of some of their itemsets which are considered as sensitive information. Hence the privacy of sensitive information of the participating sites is to be preserved [8]. In such cases, classical data mining solutions cannot be used. Hence, Secure Mutiparty Computational (SMC) solutions can be applied to maintain privacy in distributed association rule mining [16]. The goal of SMC in distributed association rule mining is to find global frequent itemset without revealing the local support count of participating sites to each other. The subsequent sections of the paper are organized as follows. Firstly, related existing works are reviewed. Secondly, the proposed approach and its performance evaluation are discussed. Lastly, a suitable conclusion and future work for maintaining privacy is attempted.
2. Related Work While practicing data mining, the database community has identified several severe drawbacks. One of the drawbacks frequently mentioned in many research papers is about
Velammal College of Engineering and Technology, Madurai
maintaining the privacy of data residing in a data source [6], [20]. Privacy preserving data mining provides methods for finding patterns without revealing sensitive data. Numerous research works are underway to preserve privacy both in individual data source and multiple data sources. There are two broad approaches for privacy-preserving data mining namely data sanitization [21] and Secure Multiparty Computation (SMC) [10], [16]. A few representative works in SMC is discussed in this section. The concept of Secure Multiparty Computation (SMC) was introduced in [22]. In many applications the data is distributed between two or more sites, and for mutual benefit these sites cooperate to learn the global data mining results without revealing the data at their individual sites. The basic idea of SMC is that this computation is secure if at the end of the computation no party is unaware about the other participating sites except its input and the results. The secure computation protocols are presented in the form of combinational circuit [9]. The idea is that the function F to be computed is first represented as a combinational circuit and then the parties run a short protocol to securely compute every gate in the circuit. Every participant gets corresponding shares of the input wires and the output wires for every gate. Here the size of the protocol depends on the size of the circuit, which depends on the size of the input. This is inefficient for large inputs as in data mining. In paper [18] Vaidya et al. proposed a method for privacy preserving association rule mining in vertically partitioned data. Each site holds some of the attributes of each transaction. An efficient scalar product protocol was proposed to preserve the privacy among two parties. This protocol did not consider the collusion among the parties and was also limited to boolean association rule mining. In [14] Kantarcioglu et al. proposed a work to preserve privacy among semi-honest parties in a distributed environment. It works in two phases assuming no collusion among the parties. But in some real life situations collusion is inevitable. The first phase identifies the global candidate itemsets using commutative encryption (RSA algorithm) which is computationally intensive and the second phase determines the global frequent itemsets. This work only determines the global frequent itemsets but not their exact support counts. An algorithm using Clustering future (CF) tree and secure sum is proposed to preserve privacy of quantitative association rules over horizontally partitioned data [17] and fixing of proper threshold value for constructing the CF tree is not easy. The efficiency of the algorithm is
Page 407
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
unpredictable since it depends on the threshold value chosen to construct the CF tree. The present analysis proposes a novel, computationally simple and secure multiparty computation algorithm for semi-honest parties.
One basic condition for any cryptosystem is that the system is closed, i.e. any operation on an element of the system results in another element of the system. In order to satisfy this condition for Elliptic curves it is necessary to construct non-standard addition and multiplication operations.
3. Problem Definition
To encrypt P, a user picks an integer, k, at random and sends the pair of points (k*BP, P+k*PUBKEY)
Let S1, S2… Sn. be the set of participating sites where n>2. Let D1,D2…Dn be the data sources of sites S1, S2… Sn respectively which are geographically distributed and let I = {i1,i2…im} be a set of items. Each transaction T in Di such that T ⊆ I, where i=1 to n. Li be the local frequent itemset generated from a participating site Si and G be the global frequent itemset. To generate the global frequent itemset G, each site needs to send its respective support counts of its local frequent itemsets to the other participating sites. The intended goal of this proposed approach is to discover the global frequent itemsets without revealing the sensitive information of all the participating sites.
4. Proposed Approach A new approach is proposed in this paper to estimate the global frequent itemsets from multiple data sources while preserving the privacy of the participating sites. All sites in the mining process are assumed as semihonest. The semi-honest parties are honest but try to learn more from received information. Any site can initiate the mining process. To protect sensitive information of each participating site, elliptic curve cryptography and randomization are applied.
4.1 Overview of Elliptic Curve Cryptography (ECC) Many cryptographic algorithms and protocols have been proposed to encrypt and decrypt data. Among the algorithms, ECC is proved as more secured cryptographic technique. Also it is computationally efficient than RSA and Diffie-Hellman [23]. The use of elliptic curve groups over finite fields as a basis for a cryptosystem was first suggested by Koblitz [24]. The elliptic curve approach is a mathematically richer procedure than standard systems like RSA. The basic units for this cryptosystem are points (x,y) on an Elliptic curve, E(Fp), of the form Y2 =x3+ax+b, with x,y,a,b~Fp ={1,2,3 ....... p-2,p-1} Where Fp is a finite field of prime numbers.
Velammal College of Engineering and Technology, Madurai
To decrypt this message you multiply the first component by the secret key, s, and subtract from the second component, (P+ k*PUBKEY ) - s*(k*BP) = P+k*(s*(BP)) - s*(k*BP) =P We then reverse the embedding process to produce the message, m, from the point P. This system requires a high level of mathematical abstraction to implement.
4.2 Secure Mining of Support Counts Our proposed method finds the global frequent itemsets without disclosing the support counts of individual sites. Algorithm 1 given below computes global candidate itemsets in a secured manner using ECC technique. Algorithm 2 computes the globally frequent itemsets using random number addition.
Algorithm 1 Input: n local frequent itemsets. Each set, denoted by Ai (1≤ i ≤ n), belongs to each party. s is the Threshold support. Si denotes the local storage of each party. A1 denotes the commoner, i.e. i=1.LL is the count of locally frequent itemsets received so far by the commoner. Initially LL=0. CLL is the total count of locally frequent itemsets. Output: UiAi (1≤i≤n), that is the set union of locally frequent itemsets (Global candidate itemsets). 1. Compute the locally frequent itemsets at each site. Make an entry in the local storage, Si of each party. 2. Encryption traversal: Encrypt the locally frequent itemsets using the • public key of the commoner say pub1. • Send the encrypted message to the next party, which encrypts the locally frequent itemsets using its public key and forwards the result to the next site and the process continues until it reaches the commoner again. 3. Decryption traversal: • Decrypt the message using the private key of the commoner say priv1. • Send the decrypted message to the next party, which decrypts the message further using its private key and forwards the result to the next site and the process continues until it reaches the commoner again.
Page 408
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
4. Increment LL If any received subset is not present in the local storage Si Then add to Si 5. The process ends when LL equals CLL at the commoner. The above algorithm meets the requirement that no party can determine any itemset’s owner. Initially, each of the participating sites generates their locally frequent itemsets using any of the frequent itemset generation algorithms like Apriori, FP growth tree algorithm, etc. The generated locally frequent itemsets at each site, denoted as Ai, are added to the local storage, Si of respective sites. The purpose of the local storage is to avoid the infinite looping of any itemset before reaching the commoner. This is accomplished by forwarding any received itemset by verifying its presence in the local storage. Thus the local storage ensures that any itemset will pass through all the participating sites only once, that is at the worst case. The count of locally frequent itemsets generated at each of the participating sites is calculated to determine the end of the Algorithm 1. For this, the commoner chooses a random number to be added with its count of locally frequent itemsets. This value is sent to the next participating site, which adds its count of locally frequent itemsets to the received value and forwards the result to the next site. The process is repeated until all the participating sites have added their counts of locally frequent itemsets and when the value reaches back the commoner, it subtracts the random number to get the count of all locally frequent itemsets. It is denoted as CLL. The commoner maintains the count of received itemsets (LL). On receiving any forwarded itemset, the commoner increments LL and if it is not present in the local storage, an entry is made. If it is present already then the received itemset is discarded. Thus the commoner removes the redundancies. When the other sites receive any forwarded itemset, it is verified with the respective local storage and if an entry is present, then the itemset is directly forwarded to the commoner, else an entry is made and the itemset is forwarded to random destinations. Thus we avoid the repeated looping of any itemset. The end of Algorithm 1 is when the commoner’s count of received itemsets (LL) reaches CLL. The commoner’s local storage contains the union of the locally frequent itemsets of all the participating sites. The following algorithm privately computes the globally supported itemsets [5].
Algorithm 2 Input: Global candidate itemsets. Number of participating sites n >= 3. The site initiate the mining process can be the commoner.
The commoner calculates the global database size 1. as follows: • Add any random number R to the local database size. • Send the value to the next party, which adds its local database size to the received value and forwards the resulting value to the next site and the process continues until it reaches the commoner again. The commoner subtracts R from the received • value to get the global database size. • 2. For each item i in S1 do Add a random number to the local support count of I and send it to the next site. The next site adds its support count for i to the count and sends it to the next site and this process continues till the count reaches the commoner. End for. The global support percentage of an itemset, X is calculated as summation of local support over all the participating sites, 1≤i≤n divided by the summation of the database size (DB) over all the participating sites, 1≤i≤n. The commoner securely calculates the global support count of each itemset in the union of locally frequent itemsets. For each itemset in the union, the commoner the next site. The process is repeated until it reaches back the commoner. The commoner calculates the global support percentage for every itemset in the union. Those itemsets whose support is greater than or equal to the threshold support form the set of globally frequent itemsets.
5 Performance Evaluation Distributed environment with three participating sites was simulated to evaluate the performance of the proposed algorithm. At each site, a P-IV, 2.8 GHz, 2 GB RAM machine is run on Windows operating system. The proposed method is implemented using Java and MSAccess. Three different data sources of size 10K, 50K, and 100K are synthetically generated to study the performance of the proposed work. The local frequent itemsets of the participating sites which are the inputs for the algorithm is generated by using the FP tree algorithm for supports varying from 10% to 30%.
5.1 Timing analysis
Output: Global frequent itemsets.
Velammal College of Engineering and Technology, Madurai
Page 409
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Graph in Fig. 1 shows that the time taken to encrypt and decrypt the local frequent itemsets using existing RSA D|=100k,|I|=12,ATL=9; No. of sites = 3
60
140
50
120 100
40
Time (in sec.)
Tim e (in Sec.)
|D| = 10k, |I| = 10, ATL=7; No. of sites = 3
RSA 30
ECC
20 10
80
RSA
60
ECC
40 20
0
0
10
20
30
10
Cryptographic Technique and Our proposed ECC Technique. The result shows that the ECC cryptographic technique is more efficient than the existing cryptographic technique. Fig. 1.1 Time complexity for Encryption and Decryption of 10K data source
30
Fig 1.3 Time complexity for Encryption and Decryption of 100K data source
5.2 Accuracy analysis Accuracy is one of the main objectives of privacy preserving distributed data mining. Table 1 shows the number of global frequent itemsets generated by both methods that use traditional RSA method and proposed ECC method. Both are found to be identical.
|D|=50k,|I|=10,ATL=7; No. of sites = 3 140 120 Time (in Sec.)
20 Support (%)
Support (%)
100 80
RSA
60
ECC
40 20 0 10
20
30
Support (%)
Fig. 1.2 Time complexity for Encryption and Decryption of 50K data source TABLE 1
Accuracy comparison of traditional RSA and proposed ECC methodology
Data set 10K 50K 100K
Support
No. of Sites
10% 20% 10% 20%
3 3 3 3
10% 20%
3 3
Total No. of Global Frequent Itemsets RSA ECC 1023 1023 320 320 231 231 175 175
Velammal College of Engineering and Technology, Madurai
60 10
60 10
Accuracy 100% 100% 100% 100% 100% 100%
Page 410
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6 Conclusion and Future Work Several existing algorithms for preserving privacy in a distributed environment have been analyzed and an efficient algorithm for the same has been proposed and implemented. This proposed method uses a mathematically rich procedure of cryptography namely the elliptic curve cryptography. The proposed model believes that the participating sites are semi honest. We can extend the proposed model to work even for dishonest parties. The proposed model works for only homogeneous databases. We can improve the system to work even for heterogeneous databases where the attributes of the participating sites are different.
References [32] R.Agrawal, R.Srikant, “ Fast Algorithms for Mining Association Rules, “ in proceedings of the 20th VLDB Conference Santiago, Chile,1994, pp.487-499. [33] M.Ashrafi, D.Taniar, K.Smith, ” ODAM : An Optimized Distributed Association Rule Mining Algorithm,” IEEE Distributed Systems Online, 2004, 5th ed., Vol. 3. [34] M.Ashrafi, D.Taniar, K.Smith, “ Privacy-PreservingDistributed Association Rule Mining Algorithm.” in proceedings of International Journal of Intelligent Information Technologies, 1st ed. Vol.1, 2005, pp. 46-69. [35] M.Atallah, E.Bertino, A. Elmagarmid, M.Ibrahim, V.S.Verykios, “Disclosure limitation of sensitive rules,” in Proceedings of the IEEE Knowledge and Data Exchange Workshop (KDEX'99). IEEE Computer Society, 1999, pp. 45-52. [36] D. Cheung, V.Ng, A.Fu, Y.Fu, “Efficient Mining of Association Rules in Distributed Databases,” IEEE Transactions on Knowledge and Data Engineering. 1996, 8th ed., Vol. 6, pp. 911-922. [37] C.Clifton, D.Marks, “Security and Privacy Implications of Data Mining,” in Proceedings of the ACM SIGMOD Workshop on Data Mining and Knowledge Discovery, 1996, pp.15-19. [38] C. Clifton, “Secure Multiparty Computation Problems and Their Applications: A Review and Open Problems,” in Proceedings of the Workshop on New Security Paradigms, Cloudcroft, New Mexico, 2001. [39] C.Clifton, M. Kantarcioglu, J.Vaidya, “Defining privacy for data mining. Book Chapter Data Mining, Next generation challenges and future directions, 2004. [40] O.Goldreich, S, Micali, Wigderson, “How to play any mental game - a completeness theorem for protocols with honest majority,” in 19th ACM symposium on the theory of computing, 1987, 218-229. [41] O.Goldreich, “Secure Multiparty Computation (Working Draft)”, 1998. [42] J.Han, M. Kamber, ”Data Mining: Concepts and Technique”, Morgan Kaufmann Publishers, 2001. [43] J. Han, J. Pei , Y.Yin, R. Mao, “Mining Frequent Patterns without Candidate Generation : A Frequent Pattern Approach,” IEEE Transactions on Data Mining and Knowledge Discovery, 8th ed., Vol.1, 2004, pp.53-87. [44] Inan.A., Saygyn.Y., Savas.E., Hintoglu.A.A, & Levi.A, “ Privacy preserving clustering on horizontally portioned data,” in proceedings of
Velammal College of Engineering and Technology, Madurai
the 22nd International Conference on Data Engineering Workshops (ICDEW'06), 2006. [45] M. Kantarcioglu, M., C.Clifton, “Privacy-Preserving Distributed Mining of Association Rules on Horizontally Partitioned Data,” IEEE Transactions On Knowledge And Data Engineering, 2004, 16th ed. Vol.9. [46] G.Lee, C.Chang, A.L.P.Chen, “ Hiding sensitive patterns in association rules mining,” in Proceedings of the 28th Annual International Computer Software and Applications Conference (COMPSAC'04), 2004. [47] Y.Lindell, B. Pinkas, ” Secure Multiparty Computation for PrivacyPreserving Data Mining, “ in the Journal of Privacy and Confidentiality, 2009, 59-98. [48] W.Luo, “An Algorithm for Privacy-preserving Quantitative Association Rules Mining,” in Proceedings of the 2nd IEEE International Symposium, 2006. [49] J.Vaidya, C.Clifton, “Privacy Preserving Association Rule Mining in Vertically partioned data,” in Proceedings of the SIGKDD ’02 Copyright ACM, 2002. [50] J.Vaidya, C.Clifton, “Privacy-Preserving Data Mining: Why, How, and When, “ IEEE Security and Privacy, 2004. [51] S.Verykios, E.Bertino, I.Provenza, Y.Saygin, Y.Theodoridis, “State- of- the -Art in Privacy Preserving Data Mining,” in ACM SIGMOD Record, 33rd ed., Vol 1., 2004, pp. 50-57. [52] S.Wang, Y. Lee, S.Billis, A.Jafari, “Hiding Sensitive Items in Privacy Preserving Association Rule Mining” in IEEE International Conference on Systems, Man and Cybernetics, 2004. [53] A.C. Yao, “ How to generate and exchange secret,” in Proceedings of the 27th IEEE symposium on Foundations of Computer Science. 1986, pp. 162-167. [54] M. Brown, D. Hankerson, J. Lopez, and A. Menezes, “Software Implementation of the NIST Elliptic Curves Over Prime Fields, ” IEEE computer society, 2004. [55] Andrew Burnett, Keith Winters, and Tom Dowling, “A Java implementation of an elliptic curve Cryptosystem- Java programming and practices,” 2002. [56] Mugino Saeki ,“Elliptic curve cryptosystem, ” McGill University, Montreal, February 1997.
Page 411
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Harmonics In Single Phase Motor Drives And Energy Conservation Mustajab Ahmed Khan#1, Dr.A.Arunagiri #2 #
EEET Department, Yanbu Industrial College, Royal Commission ,Kingdom of Saudi Arabia. 1
[email protected] [email protected]
2
Abstract— The modern electric power systems that include non-linear loads may experience power quality problem. Nonlinear loads draw current that cause power quality problems such as harmonic distortion. Harmonic problems are mainly due to substantial increase of non- linear loads such as the Switched mode power supplies(SMPS),Uninterrupted power supplies, Variable speed drives using power electronic devices or microprocessor and power electronic controllers in AC/DC Transmission links. In this paper Electronic regulators, resistive type regulators for single phase induction motors used in fans are analyzed for harmonic analysis and energy saving. A proposed capacitive type regulator is also analyzed and compared with the existing regulators.
Keywords— DAS, Triplens , Harmonics, Regulators,FFT.
1.
INTRODUCTION
The harmonics and power saving schemes study in this paper are based on single phase induction motor in fan loads. A fan regulator is a means of regulating the fan speed in order to achieve the desired air velocity. It is interfaced between the single phase supply and the single phase induction motor which drives the fan blades and effects change in speed . At present the electronic type regulators are increasingly being used in place of the resistive type regulator on account of power savings. compactness, step less speed regulation . However, electronic regulators operate at low power factor and generate high level of harmonics. While the low power factor operation will mean extra burden on the supply system for the supply of reactive power, the harmonics generated cause overheating of the motor which could lead to reduction in fan life. Further, electronic regulator cause a typical humming sound, disturbance in TVs, tape recorder and interference with the telecommunication networks. It is therefore necessary to examine the fan regulators critically from two angles namely, it's energy consumption potential and its side effects. With this in mind, the capacitive type regulator was developed and its performance determined
Velammal College of Engineering and Technology, Madurai
and compared with other regulators like the resistive type and electronic type.
2.
HARMONICS
Harmonics are integral multiples of the fundamental frequency. Devices causing harmonics are present in all industrial, commercial and residential, installations. Harmonics are caused by non-linear loads. A load is said to be non-linear when the current it draws does not have shape as the supply voltage. Devices comprising power electronics circuits are typical non-linear loads. Such loads are increasing frequently and their percentage in overall consumption is growing steadily. Examples include industrial equipment (Welding machines, arc furnaces. Induction furnaces rectifiers),variable-speed drives for Asynchronous and DC motors, office equipment (PCs, photocopy machines, fax machine), household appliances (television sets, microwave ovens, fluorescent lighting, etc.) and UPS. 2.1
Effects of Harmonics:
In distribution systems, the flow of harmonics reduces power quality and consequently causes a number of problems: overloads on distribution systems due to the 1. increase in the rms current. 2. overloads on neutral conductors due to the summing of third-order harmonics and triplens created by single-phase loads. overloads, vibrations and premature ageing of 3. generators, transformers, motors, etc., transformer hum. overloading and premature ageing of capacitors in 4. power factor correction equipment. distortion of the supply voltage, capable of 5. disturbing sensitive loads.
Page 412
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
6. disturbances on communications networks and telephone lines. The harmonics most frequently encountered (and consequently the most troublesome) on three-phase distribution systems are the odd-order harmonics (3rd, 5th, 7th, etc.).Utilities monitor harmonic orders 3, 5, 7, 11 and 13. It follows that conditioning of harmonics is imperative up to order 13 and ideally should include harmonics up to order 25. 2.2 Selection of measurement devices:
The conventional resistive type regulator controls the speed by suitably dropping the voltage across the resistance in series with the motor thus causing considerable power loss. 3.2
Electronic regulator:
Digital analyzers, based on recent technology, provide sufficiently accurate measurements for the indicators presented above. Other measurement devices were used in the past.
The solid state fan regulator is triac based. It operates on the principle of the firing angle control of the triac which consumes negligible power. Delayed firing angle of the triac causes chopping of the current to reduce the voltage across the motor and hence reduced the speed. This feature which results in energy conservation. However, the non linear operating nature of the electronic circuitry is the main causes of harmonic generation .
2.2.1
3.3
Functions of Digital Analyzers:
Capacitive type:
The microprocessors used in digital analyzers calculate the values of the harmonic indicators (power factor, crest factor, distortion power, THD), offer a number of additional functions (correction, statistical detection, management of measurements, display, communication, etc.), when they are multi-channel devices, provide simultaneously and nearly in real time the spectral breakdown of voltage and current.
It is a step type regulator like the resistive type . It's power consumption is near to zero at all speeds and there is no waveform distortion. The system operates near unity power factor or slightly leading power factor . This type of regulator can be used for single phase motors of exhaust fans, mixers and grinders .
2.2.2 Operating principle of Digital Analyzers and dataprocessing techniques:
A sample of different types of regulators was chosen for evaluation. The following parameters of the different fan regulators where measured by using the data acquisition system(DAS), software from LabVolt, laptop, and the comparative analyses of the results were done.
Analogue signals are converted into a series of digital values. On the basis of the digital values, an algorithm implementing the Fast Fourier Transform (FFT) calculates the amplitude and the phases of the harmonics over a large number of observation time windows. Most digital analyzers measure harmonics up to the 20th or 25th order for calculation of the THD. Processing of the various values calculated using the FFT algorithm (smoothing, classification, statistics) can be carried out by the measurement device or by external software.
3.4
Methodology of the evaluation:
current drawing by the regulator and the fan 1. combination . power consumption by the regulator . 2. 3. voltage and current harmonics (THD(V),THD(I)) Power factor of the circuit. 4.
On the basis of analysis results, it may be necessary to: Derate any future equipment installed. 1. Quantify the protection and harmonic-filtering 2. solution that must be installed. Compare the values measured to the reference 3. values of the utility (harmonic- distortion limits, acceptable values, reference values).
3.5 3.5.1
Measurement and Analysis: Electronic control of the Single phase motor fan: Figure 1:Electronic type regulator.
PRINCIPLE OF OPERATION OF DIFFERENT 3. TYPES OF REGULATORS 3.1
Resistive type:
Velammal College of Engineering and Technology, Madurai
Page 413
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
α
V1
100.5 110. 7 90 110. 6 75.7 110. 6 60.5 110. 6 45 110. 6 30.2 110. 6 15 110. 6 7 110. 6 0 10.6 1
V2
W2
cosφ THD THD N (V) (I) (rpm)
I1
W1
0.11 9 0.13 1 0.14 3 0.14 3 0.14 7 0.14 7 0.14 7 0.14 6 0.14 9
7.6 67.9 7.45 0.76 0 6 9.65 77.5 0.84 9.6 2 12.3 89.3 12.1 0.91 2 7 6 3 12.3 89.3 12.1 0.91 2 7 6 3 15.3 104. 15.2 0.98 0 9 0 2 15.9 108. 15.7 0.99 1 4 3 16.1 109. 15.9 0.99 5 4 7 8 16.2 110. 16.0 0.99 3 1 7 9 16.4 109. 16.4 0.99 6 9 0 9
2.9
66.8 730
3
62.4 850
3
54
1200
2.9
43.5 1500
2.6
31.3 1715
3.6
19.9 1805
4.2
10.3 1835
3.6
6.3 1860
3.3
4.9 1880
Table 1: Measurements of the Electronic type regulator. (V1-Supply Volts,I1-Input Current ,W1-Input Power, W2-Motor Power,V2-Motor Volts)
Figure 3. Current harmonic spectrum for (I1)( α=45 deg)
Figure 4. Voltage harmonic spectrum(obtained from DAS).
3.5.2 Resistive type:
Figure 2.Current and voltage waveform.( α=45deg) Figure 5: Resistive type regulator.
Velammal College of Engineering and Technology, Madurai
Page 414
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Table 2. Measurements of the Resistive type regulator
V1
I1
110 109.7 110.3 110.7 110.7
W1
0.1 0.11 0.12 0.13 0.15
V2
Cos(φ) %THD %THD N (V) (I) (rpm) 6.25 0.998 1.6 5.1 722 7.8 0.998 1.6 4.5 870 9.8 0.997 1.1 3.9 1100 11.95 0.997 2.2 4.3 1360 16.64 0.987 2.8 4.1 1880
W2
10.8 64 11.85 71 13.34 81.24 14.61 90.7 16.63 110.6
Figure 8:plot of THD(V) Vs Speed 3.5.3 Capacitor regulator for fan:
Figure 9: plot of THD(I) Vs Speed
Figure 6 : Capacitive type regulator
V1
I1
W1
V2
W2
120 120 120 120 120 120
0.09 0.124 0.133 0.136 0.141 0.152
5.45 10.65 12.33 12.72 14.46 17.2
61 86 94 94 102.8 113
5.45 10.65 12.33 12.72 14.46 17.2
Cos(φ) %THD (V) 0.5 1.5 0.675 1.4 0.778 1.5 0.781 1.7 0.857 1.2 0.942 1.6
%THD (I) 5.6 4.1 3.8 5.6 3.4 4.7
N(rpm) 690 1278 1450 1545 1700 1940
Table 3: Measurements of the capacitive type regulator
Figure 10: plot of input Power Vs Speed 4
Figure 7 :plot of power factor Vs Speed
Velammal College of Engineering and Technology, Madurai
CONCLUSION
The results of the evaluation are summarized as : It is found the power loss in the resistive type is of the order of 5 Watts per fan . Capacitive or electronic type can have power saving of about 4.5 watts. per unit. As the fan load is huge, the overall power saving will be enormous. Capacitive type regulator has low level of harmonic in both voltages (1-2%) and current which is less than 5%. The electronic regulator has a level of THD about 5% for voltage and (6to 60)% for current .Capacitive and electronic type regulators have better energy saving potential, but an electronic regulators is a source of
Page 415
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
harmonic. Capacitive type regulators has no noise problems which exist in electronic regulators at low speed. REFERENCES [57] Francis De La Rosa, “Harmonics and Power Systems”,2006, CRC Press. [2] Effects of harmonics on equipment, IEEE Transactions, April 1993,Vol 8. [3] George Wakelah,2000” Harmonics in power systems, their causes,effects and mitigation”, Department of electrical & comp.engg-,Mexico University. [4] Roger C Dugan,Electrical power Systems Quality, McGraw-Hill,New York,1996,pp.130-133.
Velammal College of Engineering and Technology, Madurai
Page 416
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Improvement towards efficient OPFET detector Jaya V. Gaitonde, Rajesh B. Lohani Department of Electronics and Telecommunications, Goa University
Goa Engineering College, Farmagudi-Ponda-Goa-India [email protected], [email protected]
Abstract—OPFET(Optical Field Effect Transistor) is an useful device for optical communication and as photo detector. This paper discusses some performance issues of the device under various conditions. Keywords—OPFET, photodetector, switching time, Noise Equivalent Power(NEP), Signal to Noise Ratio(S/N).
I . INTRODUCTION At present, optical fiber communication plays an important role in cable communication technology for wideband, multimedia and high-speed applications [3] [5]. In order to be able to manufacture wireless terminals for optical fiber links at reasonable cost, good agreement must be achieved between the photo-detector and the millimeter-wave circuit, as well as small size and low weight. When a GaAs MESFET device is optically illuminated, absorption phenomena take place at the gate–drain and gate–source regions, which induce both photoconductive and photovoltaic effects. The performance of a GaAs MESFET can be significantly enhanced by scaling down the device geometry. The radiation is allowed to fall on the semitransparent Schottky gate of the device. When light is turned ‘on’ the parameters such as threshold voltage, channel charge, channel current, channel conductance and gate to source capacitance reach the steady state value at a lesser time than that when light is turned off. The device performance is greatly improved by shortening the gate length. The OPFET performance is also dependent on the NEP(Noise Equivalent Power) and signal to noise ratio(S/N)[4]. It is understood that the received optical signal generates electron-hole pair in the semiconductor resulting in the modulation of the channel conductivity due to photoconductive effect and channel conductance through a development of a forward photovoltage due to photovoltaic effect. The electrical parameters such as threshold voltage, drain-source current, gate capacitances and switching response affect the OPFET’s performance. The diffusion process introduces less process-induced damage compared to ion-implantation, which suffers from current reduction due to a large number of defects introduced by ion-implantation process.
Velammal College of Engineering and Technology, Madurai
The photocurrent peak value, peak-time and discharge-time are some of the performance factors. The frequency response of OPFET is dependent upon transit time and the RC time constant of the device. II.
THEORY
The schematic structure of OPFET is as shown in fig. 1a.
Fig. 1 a. Cross-sectional view of an OPFET [2].
The schematic structure of the ion-implanted GaAs OPFET with back illumination is shown in Fig.1b and 1c for the two cases.
Fig.1. b Schematic structure of the device with fiber inserted partially into the substrate [8]. Fig 1 c Schematic structure of the device when the fiber is inserted up to the substrate-active layer interface [8].
Page 417
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.!. d Schematic structure of the opaque gate device [5]. Fig1.e Schematic structure of the device which is illuminated everywhere on the surface [5].
With the improvement on the process technology, a short gate-length MESFET device can be fabricated. The radiations are allowed to fall on the schottky gate as well as between the spacing between the source-gate & gate- drain , allowing more absorbtion. The gate-controlling capability will be reduced by the penetration of the electric field from the sidewall at both sides of the gate, and the threshold voltage of a short gate-length device will be influenced by the drain bias. The drain current Ids has been calculated by numerically integrating the charge in the channel region given by (1) Ids= εoεrZµn/d[Vgs-Vth-V(x)]E(x) where εo and εr are the permittivity of the air and GaAs , d is the gate length, Z is the device width, µn is the field dependent mobility of electrons in the channel length, Vgs is the gate-to-source voltage, V(x) is the potential distribution in the channel, E(x) is the electric field in x direction. The RC time constant has been obtained by RC= Cgs/gm where Cgs is the gate to source capacitance and gm is the transconductance. The frequency response of a GaAs MESFET depends upon the transit time and the RC time constant of the device. The transit time is the finite time required for the carriers to travel from the source to the drain. For short channel devices, the transit time is usually small compared to the RC time constant resulting from the input gate capacitance and the transconductance. In the presence of illumination,
Velammal College of Engineering and Technology, Madurai
the RC time constant is affected significantly by the incident optical power. The responsivity of the device has been obtained from R(λ)= Iph/Popt where Iph is the photoresponse current and Popt is the incident optical power. λ is the operating wavelength. The photocurrent gain M is given by M= Iph/IL where Iph is the photoresponse current and IL is the primary current. The photoresponse current is the difference between the drain current in the illuminated condition and that in the dark condition. The primary current IL is given by IL= qηPoptA/hv where q is the electronic charge, η is the quantum efficiency, Popt is the incident optical power, A is the illuminated device area, h is the Planck’s constant and v is the frequency of incident radiation. The threshold voltage Vth of an optically biased MESFET is given by Vth= Vth0 – sech( k1Lg/2)A1s where Vth0 is the threshold voltage for a long channel optically biased MESFET given by Vth0=Vbi-Vop-q/e[(Nd-Rtp’/b)(b2/2)-Qtn’((b+1/α)exp(-αb)1/α)] where A1s is the first term of the Fourier coefficient for excess sidewall potential at the source side of the gate, Nd is the uniform donor impurity concentration in the channel, R is the surface recombination rate, b is the thickness of the active layer, Q is the incident photon flux density, tn’ is the lifetime of the electron, tp’ is the lifetime of the hole, α is the absorption coefficient per unit length, k1 is the eigen value of the Green’s function in the gate region, and Lg is the gate length. Vbi is the built-in potential at the Schottky gate contact, and Vop is the photovoltage developed at the Schottky junction due to illumination. The switching time of an OPFET depends on the active channel thickness and corresponding impurity flux density of the diffusion process. The switching time t is computed for different active layer thickness expressed by the following equation t=[L(QBn-Vop-Vgs)1/2]/[Vs[(qNdavga2/2ε)1/2-(QBn-VopVgs)1/2]] (2) Finally, the switching time t as a function of impurity flux density Qdiff can be expressed by following equation: t=[L(QBn-Vop-Vgs)1/2]/[Vs[(Qdiffq(D1t1)1/2/2ε(Л)1/2)1/2(QBn-Vop-Vgs)1/2]] (3) where q is the electronic charge, L is the channel length, Ndavg is the average channel doping concentration, a is the active layer thickness, Qdiff is the impurity flux density during diffusion process, D1 is the diffusion constant for diffusion process, t1 is the diffusion time, Vgs is the gate to source voltage, QBn is the indium tin oxide Schottky barrier height, Vop is the photo induced voltage, Vs is the saturation velocity.
Page 418
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
When a GaAs MESFET device is optically illuminated, absorption phenomena take place at the gate–drain and gate–source regions, which induce both photoconductive and photovoltaic effects. In front illumination, the radiation is allowed to fall on the semi-transparent Schottky gate of the device. For opaque gate, the light is incident through the gate-drain and gate-source spacings. In the case of back illumination, two cases are considered: one in which the fiber is inserted partially into the substrate and the other, in which the fiber is inserted upto the active layer—substrate interface. The later case represents improved absorption in the active layer of the device. In Fig.1b ,the fiber is inserted partially into the substrate so that the absorption takes place in both substrate and active region. In Fig1c, the fiber is inserted up to the junction of the substrate and the active layer where photo-absorption takes place in the active region only. The drain-source current flows along the x-direction and the illumination is incident along the y-direction of the device. Electron-hole pairs are generated due to absorption of photons in the neutral substrate region, the active layer-substrate depletion region, the neutral channel region and the schottky junction depletion region. The optically generated electrons move toward the channel and contribute to the drain-source current when a drain-source voltage is applied while the holes move in the opposite direction. When these holes cross the junction a photo-voltage is developed. This voltage being forward biased reduces the depletion width of both the junctions.
The current is enhanced due to absorption more in the active channel region.
Fig. 2. switching time τ versus Active layer thickness for dark and optical illumination condition.
Equation (2) is used to compute the switching time τ as a function of the device active layer thickness a for dark and optical illumination conditions. The result illustrated in Fig. 2 shows that the switching time under dark condition abruptly decreases with respect to increasing device active layer thickness. The switching time is also reduced under optical illumination condition. Therefore it is observed that the transition from dark to illuminated condition of the OPFET significantly changes the switching time for certain device active layer thicknesses.
The drain-source current is significantly enhanced for the device when the fiber is inserted up to the active layer substrate junction than the case where finite substrate effect is taken into account. Under back illumination the number of carriers generated at the active layer-substrate depletion region is more than that in the gate depletion region. For opaque gate, two photo-voltages are developed: one across the Schottky junction due to generation in the side walls of the depletion layer below the gate and the other across the channel-substrate junction due to generation in the channel-substrate depletion region. There is no generation of electron-hole pairs just below the gate. When the device is illuminated over the entire surface, i.e. on the gate and between the spacing of gate-source and gate-drain, two photo-voltages are developed, one at the Schottky junction and one at the active-layer substrate junction.
III.
Fig 3 switching time τ versus impurity flux density .
The characteristics of switching time as a function of impurity flux density have been determined by using the equation (3) and the plot is illustrated in Figure3. The switching time is shown to be abruptly decreasing initially for increasing impurity flux densities and still decreasing but less dramatically with respect to continued increase in impurity flux density. The plot indicates that the switching characteristics change with low, moderate and high impurity deposition on the silicon during diffusion process. Most of the optically generated carriers should originate within the Schottky depletion region (gate depletion) for both high quantum efficiency and high switching speed, and in addition, the depletion region must not be so wide
RESULTS AND DISCUSSIONS
Velammal College of Engineering and Technology, Madurai
Page 419
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
that the transit time of the carriers limits the frequency response. It is also seen from the two graphs that the switching time required for back illumination is less compared to other cases. This is because for the back illumination case, there is enhanced absorption in the active region. IV.
Microwave and optical technology letters / vol. 26, no. 4, August 20 2000 [8]. Nandita Saha Roy and B. B. Pal “Frequency-Dependent OPFET Characteristics Improved Absorption under Back Illumination.” Journal of Lightwave Technology, Vol. 18, No. 4, April 2000.
with
[9]. Nandita Saha Roy, B. B. Pal, and R. U. Khan “Analysis of GaAs OPFET with Improved Optical Absorption under Back Illumination.” IEEE Transactions on Electron Devices, Vol. 46, No. 12, December 1999.
CONCLUSION
Some methods of increasing the performance of OPFET are studied. It is found that the performance of the device is increased by reducing the switching time with the increase in the active layer thickness and the impurity flux density. Also, the device with back illumination, having reduced switching time may be considered as an useful device for the design of high speed optical detector and radio frequency optical switch in communication. ACKNOWLEDGEMENT Authors thanks Dr. B.B. Pal , ITBHU, and Mr. Vivek Kamat, DTE Goa for providing constant help and encouragement and necessary guidance. REFERENCES [1] Shubha Rani Saxena, R. B. Lohani, R. U. Khan, B. B. Pal. “Generalized dc model of GaAs optical field effect transistor considering ion-implanted profile.” [2]. S.N. Chattopadhyay, N. Motoyama, A.Rudra, A.Sharma, S. Sriram, C.B. Overton and P. Pandey. “Optically controlled silicon MESFET modeling considering diffusion process.” Journal of semiconductor technology and science, vol.7, no.3, September, 2007. [3]. K. Balasubadra1, A. Arulmary, V. Rajamani, K. Sankaranarayanan. “Two dimensional numerical modeling and simulation of a uniformly doped GaAs MESFET photodetector.” Journal of Optical Communications 29 (2008) 4 [4]. P. Chakrabarti, Vinayak Jha, Pankaj Kalra, and Gaurav Gupta. “ Noise modeling of an optically controlled MESFET(OPFET).” Microwave and optical technology letters / vol. 33, no. 2, April 20 2002 [5] Nandita Saha Roy, B. B. Pal, and R. U. Khan, “Frequency-Dependent Characteristics of an Ion-Implanted GaAs MESFET with Opaque Gate Under Illumination”, Journal of lightwave technology, vol. 18, no. 2, February 2000. [6]. M.A. Alsunaidi_, M.A. Al-Absi “ Influence of electrode parameters on the performance of optically controlled MESFETs.” M.A. Alsunaidi, M.A. Al-Absi / Optics & Laser Technology 40 (2008) 711–715 [7]. M. Madheswaran, V. Rajamani, and P. Chakrabarti“ Quasi-twodimensional simulation of an ion-implanted GaAs MESFET photodetector.”
Velammal College of Engineering and Technology, Madurai
Page 420
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Enhancing Temporal Privacy and SourceLocation Privacy in WSN Routing by FFT Based Data Perturbation Method R.Prasanna Kumar#1, T.Ravi*2 #
Department of Computer Science and Engineering, Anna University Jothipuram, Coimbatore 1
[email protected]
*
KCG College of Technology Karapakkam, Chennai 2
[email protected]
Abstract— Wireless Sensor Network (WSN) is an emerging
technology that shows great promise for various futuristic applications both for mass public and military. The sensing technology combined with processing power and wireless communication makes it lucrative for being exploited in abundance in future. The inclusion of wireless communication technology also incurs various types of security threats. Although the content of sensor messages describing “events of interest” may be encrypted to provide confidentiality, the context surrounding these events may also be sensitive and therefore should be protected from eavesdroppers. An adversary armed with knowledge of the network deployment, routing algorithms, and the base-station (data sink) location can infer the temporal patterns of interesting events by merely monitoring the arrival of packets at the sink, thereby allowing the adversary to remotely track the spatiotemporal evolution of a sensed event. One of the most notable challenges threatening the successful deployment of sensor systems is privacy. Although many privacyrelated issues can be addressed by security mechanisms, one sensor network privacy issue that cannot be adequately addressed by network security is sourcelocation privacy. Keywords— WSN, FFT, temporal privacy, source location privacy, perturbation
INTRODUCTION Many data mining applications deal with privacy sensitive data. Financial transactions, health-care records, and network communication traffic are some examples. Data mining in such privacy-sensitive domains is facing growing concerns. Therefore, we need to develop data mining techniques that are sensitive to the privacy issue. This has fostered the development of a class of data
Velammal College of Engineering and Technology, Madurai
mining algorithms that try to extract the data patterns without directly accessing the original data and guarantees that the mining process does not get sufficient information to reconstruct the original data. This paper considers a class of techniques for privacy preserving data mining by value perturbation perturbing the data while preserving the underlying probabilistic properties. It explores FFT based perturbation-based approach , a well-known technique for masking the data using transformation noise. This approach tries to preserve data privacy by adding transformation noise, while making sure that the noise still preserves the “signal” from the data so that the patterns can still be accurately estimated. 1. TEMPORAL AND SOURCE LOCATION PRIVACY BY FFT: A BRIEF OVERVIEW 1.1 DATA PERTURBATION
Data perturbation approaches fall into two main categories, which we call the probability distribution approach and the value perturbation approach. The probability distribution approach replaces the data with another sample from the same (estimated) distribution or by the distribution itself. On the other hand, the value perturbation approach perturbs the values of data elements or attributes directly by some additive or multiplicative noise before it is released to the data miner. Some randomized methods for categorical data may also be classified under this category. One of the main problems of the traditional additive perturbation and multiplicative perturbation is that they perturb each data element independently, and therefore the similarity between attributes or observations which are considered as vectors in the original data space is not well preserved. Many distance/similarity based data mining applications are thus hurt. 1.2 LOCATION PRIVACY
Page 421
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Location privacy is an important security issue. Loss of location privacy can enable subsequent exposure of identity information because location information enables binding between cyberspace information and physical world entities. For example, web surfing packets coming out of a home in a Mesh network enable an eavesdropper to analyse the surfing habits of one family if the source location of those packets can be determined. In a wireless sensor network, location information often means the physical location of the event, which is crucial given some applications of wireless sensor networks. A wireless sensor network can be a low duty cycle network. Often, traffic has a strong correlation with a certain event at certain time. This gives big advantages to an eavesdropper since he does not need sophisticated techniques to discriminate traffic among different events. In this paper, we study the source location privacy problem under the assumption of one single source during a specific period. However, we need to point out that such a scenario can happen in a real wireless sensor network. 1.3 TEMPORAL PRIVACY Temporal privacy amounts to preventing an adversary from inferring the time of creation associated with one or more sensor packets arriving at the network sink. In order to protect the temporal context of the packet’s creation, it is possible to introduce additional, random delay to the delivery of packets in order to mask a sensor reading’s time of creation. Although delaying packets might increase temporal privacy, this strategy also necessitates the use of buffering either at the source or within the network and places new stress on the internal store-and-forward network buffers. 1.4 FAST FOURIER TRANSFORM A fast Fourier transform (FFT) is an efficient algorithm to compute the discrete Fourier transform (DFT) and its inverse. We use a Fast Fourier Transform (FFT) based method for data perturbation, and compare it with the Singular Value Decomposition (SVD) based method. The experimental results show that the FFT based method can obtain similar performance as SVD based method in preserving privacy as well as maintaining utility of the data, however, the computational time used by the FFT based method is much less than the SVD based method. We conclude that the FFT based method is a very promising data perturbation method. The most well known FFT algorithms depend upon the factorization of N, but (contrary to popular misconception) there are FFTs with O(N log N) complexity for all N, even for prime N. Many FFT algorithms only depend on the fact that is an Nth primitive root of unity, and thus can be applied to analogous transforms over any finite field, such as number-theoretic transforms.
2. RELATED WORK There exists a growing body of literature on privacy sensitive data mining. These algorithms can be divided into several
Velammal College of Engineering and Technology, Madurai
different groups. One approach adopts a distributed framework. This approach supports computation of data mining models and extraction of “patterns” at a given node by exchanging only the minimal necessary information among the participating nodes without transmitting the raw data. Privacy preserving association rule mining from homogeneous [9] and heterogeneous [19] distributed data sets are few examples. The second approach is based on data-swapping which works by swapping data values within same feature [3]. There is also an approach which works by adding random noise to the data in such a way that the individual data values are distorted preserving the underlying distribution properties at a macroscopic level. The algorithms belonging to this group works by first perturbing the data using randomized techniques. The perturbed data is then used to extract the patterns and models. The randomized value perturbation technique for learning decision trees [2] and association rule learning [6] are examples of this approach. Additional work on randomized masking of data can be found elsewhere [18]. This paper explores the third approach [2]. It points out that in many cases the noise can be separated from the perturbed data by studying the spectral properties of the data and as a result its privacy can be seriously compromised. Our primary contribution is to provide an explicit data perturbation method, based on Fast Fourier Transformation. 3. DATA PERTURBATION METHODS The desired perturbation methods must preserve the privacy, and at the same time, must keep the utility of the data after the perturbation [14]. We propose a FFT based method for data perturbation in this section, and also review the SVD based method. FFT BASED DATA PERTURBATION 3.1 We propose a FFT based method for data perturbation in this section, and also review the SVD based method. In order to perform FFT (Fast Fourier Transform), a given matrix A, whose size is M x N, must be first transformed so that the width and height are an integer power of 2. This can be achieved in two ways, scaling the matrix up to the nearest integer power of 2 or zero padding to the nearest integer power of 2. The second option was chosen in this paper to facilitate comparison with the original matrix. Suppose the A M N. padded new matrix is A A with size of M M xN There is a DC (Dominant Component) corresponding to zero frequency in Fourier Transform, which represents the average value across the whole matrix. The DC is located at F(O, O). However, in order to perform the filtering more easily on the transformed matrix, it is more easy to move the DC to the center of the transformed matrix ie. F(M/2, N/2). This can be achieved by multiplying each entry (x, y) of the original A A with size matrix by eπi(x±Y) . Now the new matrix matrix is A M N of M M xN N. After the DC is centred, the forward Fast Fourier Trans form is performed to get the transformed matrix F. With the transformed matrix F, low-pass filtering or high-pass filtering
Page 422
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
is performed to distort the transformed matrix in the frequency domain to get a filtered matrix FFF. Then, the inverse Fast Fourier Transform is performed on the filtered matrix FFF to B obtain a distorted matrix B. Because this distorted matrix B B is bigger in size than the original given matrix A, we only pick B B, which is denoted the up-left corner of the distorted matrix B by B so that the resulting matrix has the same size M x N as the original matrix. 3.2 SVD BASED DATA PERTURBATION Singular Value Decomposition (SVD) is a popular method in data mining and information retrieval. SVD based methods have been proven to be efficient in keeping both data privacy and data utility when used as data perturbation methods , however, the time complexity of performing SVD perturbation is o(n3). This limits the size of the data that SVD based methods can be applied on. 4. PERFORMANCE METRICS In this section, we describe the performance metrics used in this paper to compare the data perturbation methods.
4.1 PRIVACY PRESERVATION METRICS This measure is given by Var(X -Y) where X represents a single original attribute and Y the distorted attribute. This measure can be made scale invariant with respect to the variance of X as S = Var(X - Y)/Var(x). The above measure to quantify privacy is based on how closely the original values of a modified attribute can be estimated. We use PM to denote privacy metric in this paper, where PM is the average of the S for all attributes in a data.
4.2 DATA PERTURBATION METRICS RP is used to denote the average change of rank for all the attributes. RK represents the percentage of elements that keep their ranks of magnitude in each column after the perturbation. The metric CR is defined to represent the change of rank of the average value of the attributes. Similarly as RK, CK is defined to measure the percentage of the attributes that keep their ranks of average value after the perturbation. 4.3 UTILITY METRICS The data utility metrics assess whether a data keeps the performance of data mining techniques after data perturbation, e.g., whether the distorted data can maintain the accuracy of the data mining techniques such as classification, clustering, etc. In this paper, we choose the accuracy in Support Vector Machine (SVM) classification as the data utility metric.
test whether the classification algorithm can keep its performance on distorted data. The above data perturbation methods are also compared on their ability to protect the privacy of data. COMPARISON OF PERTURBATION METHODS
5.1
FFT
AND
SVD
BASED
DATA
Table I shows the performance of two data perturbation methods: FFT and SVD. The performance of these two methods varies with different Rank, CutOff or Cutratio parameters. Here we choose Rank to be 10 for SVD, Cutoff = 0.95 and Cutratio = 0.9 for FFT. The parameters are chosen so that the SVM classification can obtain the highest accuracy. The two methods have similar capability in preserving data privacy. They have the same value for PM, CP, and CK. Both of the methods obtain the perfect value for PM, and CK, which means they are powerful in keeping the privacy of data. FFT is a little better than SVD for RP, and SVD is a little better than FFT in RK. The accuracy of classifying the original data is 0.8125. We can get the same accuracy by using FFT. The privacy accuracy of applying SVD is slightly lower than FFT, which is 0.8114. Next we compare the time cost in perturbing the original data. It costs 8.79 seconds for SVD but only 3.24 seconds for FFT. As we have shown in previous sections, FFT excels SVD in time cost. 5.2 INFLUENCE OF CUTRATIO AND CUTOFF IN FFT In this section, we analyze the influence of parameters on the performance of FFT based data perturbation method. We use two parameters for FFT, CutOff and CutRatio. Figure 1 shows the influence of cut ratio on the some of the privacy and utility measures. Here we keep the parameter CutOff to be 0.95. CP, CK and PM are not affected by CutRatio for the data in use and CK and PM always maintain the perfect value. With the decrease of CutRatio, RK steadily goes down and the Accuracy of classification reduces too. However, RP increases and obtains the highest value when CutRatio = 0.7 and decreases afterwards. If we keep the value of CutRatio unchanged and decrease CutOff , the Cutoff are similar to what we have described. 5.3 INFLUENCE OF RANK IN SVD Figure 2 illustrates the influence of Rank on the performance in the SVD-based methods. Like with FFT, CP, CK and PM are not affected by rank to make sense of data. and PM are not affected by Rank for the data in use. With the decrease of Rank in SVD, Accuracy and RK decreases while RP increases. This means using smaller Rank can improve the ability of SVD in preserving privacy, as the data is more distorted. However, using smaller Rank may also sacrifice the utility of the data as the accuracy of classification may decrease.
5. EXPERIMENTS AND RESULTS We have conducted experiments to test the performance of the SVD based and FFT based data perturbation methods. SVM classification are applied on the original data as well as the data distorted by each method to
Velammal College of Engineering and Technology, Madurai
Page 423
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
with Singular Value Decomposition (SVD) based perturbation method. The experiments show that FFT based method is similar to SVD based method in preserving privacy and keeping utility of data. However, the computational time used by the FFT based method is much less than the SVD based method. We conclude that the FFT based method is a very promising data perturbation method.
REFERENCES [1] Hillol Kargupta and Souptik Datta Qi Wang and Krishnamoorthy Sivakumar On the Privacy Preserving Properties of Random Data Perturbation Techniques [2] R. Agrawal and R. Srikant. Privacy-preserving data mining.
Accuracy VS Cut ratio
Accuracy VS Rank
Cut ratio Vs RP
Rank Vs RP
Cut ratio Vs Rk
Rank Vs Rk
Fig 1: Influence of parameters in FFT SVD
Fig 2: Influence of parameters in
XIII. TABLE I COMPARISON OF FFT AND SVD BASED DATA PERTURBATION METHODS.
Data
P M
RP
RK
CP
C K
SV D FFT
1
918.982 6 921.536 5
0.003 6 0.010 0
5.2308 2 5.2308 2
2
In Proceeding of the ACM SIGMOD Conference on Management of Data, pages 439–450, Dallas, Texas, May 2000. ACM Press. [3] V. Estivill-Castro and L. Brankovic. Data swaping: Balancing privacy against precision in mining for logic rules. In Proceedings of the first Conference on Data Warehousing and Knowledge Discovery (DaWaK-99), pages 389 – 398, Florence, Italy, 1999. Springer Verlag. [4] A. Evfimevski, J. Gehrke, and R. Srikant. Limiting privacy breaches in privacy preserving data mining. In Proceedings of the ACM SIMOD/PODS Conference, San Diego, CA, June 2003. [5] A. Evfimevski, R. Srikant, R. Agrawal, and J. Gehrke. Privacypreserving mining of association rules. In Proceedings of the ACM SIKDD Conference, Edmonton, Canada, 2002. [6] S. Evfimievski. Randomization techniques for privacy preserving association rule mining. In SIGKDD Explorations, volume 4(2), Dec 2002. [7] S. Janson, T. L. , and A. Rucinski. Random Graphs. Wiley Publishers, 1 edition, 2000. [8] D. Jonsson. Some limit theorems for the eigenvalues of a sample covariance matrix. Journal of Multivariate Analysis,
0
Accurac y 0.8114
Tim e 8.79
0
0.8125
3.24
XIV. CONCLUSIONS XV. In this paper we propose a Fast Fourier Transform (FFT) based data perturbation method and compare its performance
Velammal College of Engineering and Technology, Madurai
[9] M. Kantarcioglu and C. Clifton. Privacy-preserving distributed mining of association rules on horizontally partitioned data. In SIGMOD Workshop on DMKD, Madison, WI, June 2002. [10] H. Kargupta, K. Sivakumar, and S. Ghosh. Dependency detection in mobimine and random matrices. In Proceedings of the 6th European Conference on Principles and Practice of Knowledge Discovery in Databases, pages 250–262. Springer, 2002. [11] K. Liu, H. Kargupta, and J. Ryan. Random projection and privacy preserving correlation computation from distributed data. Technical report, University of Maryland Baltimore County, Computer Science and Electrical Engineering Department, Technical Report TR-CS-03-24, 2003.
Page 424
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[12] D. G. Manolakis, V. K. Ingle, and S. M. Kogon. Statistical and Adaptive Signal Processing. McGraw Hill, 2000. [13] M. L. Mehta. Random Matrices. Academic Press, London, 2 edition, 1991. [14] V. S. Verykios, E. Bertino, I. N. Fovino, et al.
State-of-the-art in privacy preserving data mining. SIGMOD Record, 33:50-57, 2004. [15] S. J. Rizvi and J. R. Haritsa. Maintaining data privacy in association rule mining. In Proceedings of the 28th VLDB Conference, Hong Kong, China, 2002. [16] J. W. Silverstein and P. L. Combettes. Signal detection via spectral theory of large dimensional random matrices. IEEE Transactions on Signal Processing, 40(8):2100–2105, 1992. [17] G. W. Stewart. Error and perturbation bounds for subspaces associated with certain eigenvalue problems. SIAM Review, 15(4):727–764, October 1973. [18] J. F. Traub, Y. Yemini, and H. Woz’niakowski. The statistical security of a statistical database. ACM Transactions on Database Systems (TODS), 9(4):672–679, 1984. [19] J. Vaidya and C. Clifton. Privacy preserving association rule mining in vertically partitioned data. In The Eighth ACM SIGKDD International conference on Knowledge Discovery and Data Mining, Edmonton, Alberta, CA, July 2002.
Velammal College of Engineering and Technology, Madurai
Page 425
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Theoretical Nvestigation Of Size Effect On The Thermal Properties Of Nanoparticles K.Sadaiyandi# 1 and M.A.Zafrulla Khan#2 #1
Department of Physics, Velammal College of Engineering and Technology Madurai 625 009, India #2 Department of Physics, Udaya School of Engineering, Vellamodi Ammandivilai, Nagercoil - 629 001, India #1
[email protected], #2 [email protected]
Abstract: In the present work the size dependent Melting Temperature (Tm), Debye Temperature ( θ D ) and Mean Square Displacements (MSDs) of Tungsten nanoparticles are determined. First, the size dependent melting temperatures and Debye temperatures are calculated by following liquid drop model. The size dependent lattice contraction is properly accounted following Qi and Wang. Secondly, the size dependent melting temperatures and Debye temperatures for Tungsten nanoparticles are calculated from the recently reported molecular dynamic simulation studies on specific heat capacities. The calculated values by the two methods are agreeing well, which supports the determination of size dependent melting temperature and Debye temperature by the liquid drop model by accounting lattice contraction. The procedure is extended to determine the MSDs of Tungsten nanoparticles for sizes ranging from 2nm to 30nm at 300K. Key Words: Nanotungsten, melting temperature, Debye temperature, Mean Square Displacement, liquid drop model, lattice contraction.
are high density, hardness, melting temperature, elastic modulus and conductivity in conjunction with the low thermal expansion. The combination of these unique properties explains the diverse applications of tungsten ranging from home lighting to thermonuclear fusion first wall protection [4, 5]. With nanoscale tungsten powders available at a reasonable cost, its usage will increase greatly and a new approach is required to balance the size dependent advantages against the temperature dependent limitations. Therefore, it is of great importance to understand the thermal stability of tungsten nanoparticles for their applications at higher temperatures. The Debye Temperature ( θ D ) of nanocrystals is an essential physical quantity to characterize many material properties such as thermal vibrations of atoms and phase transitions. Also the Einstein temperature and volume expansion coefficient are related with θ D . Hence, if the size
The study of nanocrystalline materials is an active area of research in physics, chemistry and materials sciences [1,2]. While numerous techniques are known to produce nanostructures, it is much more difficult to determine the properties of such small collections of atoms. Different physical properties such as mechanical strength, plasticity, melting, sintering and alloying ability, diffusivity, chemical reactivity and the mode of crystal growth have been found depend upon particle size. Size dependent melting point depression of small particles has been studied for many years both theoretically and experimentally [3]. The technical advantages of the low melting temperatures of small particles are (i) the ability to fuse nanoparticles to form a film at a relatively low temperature, (ii) possibility of soldering at relatively low temperatures using nanoparticles, (iii) possibility of controlling the growth process of nanoparticles by controlling the deposition or substrate temperature. Tungsten along with its alloys and compounds, occupies a unique position in material science. The material properties that make tungsten attractive to the metals industry
effect on θ D is known, the size effect on other related properties can be easily determined. As experimental and theoretical investigations in this direction are very less, it is planned to determine the size effect on the Debye Temperature ( θ D ) and Mean Square Displacements (MSDs) of tungsten nanoparticles. Melting temperature (Tm) of nanogold with particle sizes in the range 20 to 50 Å is studied experimentally and theoretically by Buffat and Borel [6]. They proposed a phenomenological model to determine the melting temperature of particles with different sizes. Also Jiang. et. al [7] introduced a simple model, free of any adjustable parameter, for the size dependent melting of nanocrystals. Recently Nanda et. al. [8] derived an expression for the size dependent melting of low dimensional systems following the liquid drop model [9]. Although the results from [6,7,8] are in good agreement with experiments they have not properly accounted the size dependent lattice contraction. Recently Sadaiyandi [10] calculated the melting temperatures and Debye temperatures of Au, Ag and Al nanoparticles from liquid drop model proposed by Nanda et. al [8] by properly accounting the size dependent lattice contraction. As the idea
Velammal College of Engineering and Technology, Madurai
Page 426
I.
INTRODUCTION
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
proposed by Sadaiyandi [10] is simple and successful, it is planned to determine the melting temperatures of tungsten nanoparticles by the same method. From the calculated values of melting temperatures, the size dependent Debye temperatures and MSDs are determined by the Lindemann theory [11]. Recently Amitava Moitra et al [12] reported the size dependent specific heat capacities of tungsten nanoparticles at different temperatures by molecular dynamic simulation. From the reported values of specific heat capacities the size dependent Debye temperatures and melting temperatures are determined by the idea proposed by Reisland [13]. The calculated values are in good agreement with the values arrived from the first method [10] which supports the determination of melting temperatures of nanoparticles by liquid drop model by considering the lattice contraction.
II.
LIQUID DROP MODEL FOR SIZE DEPENDENT MELTING One of the successes of the liquid drop model lies in providing an intuitive explanation for the phenomenon of spontaneous fission of some nuclei. Atomic clusters and nanoparticles being finite systems, their properties are dominated by the surface atoms, there fore their binding energy can be effectively represented by the volume and surface dependent term as in the liquid drop model. From this point of view the melting of atomic clusters and nanoparticles can be understood by scaling the cohesive energy to the melting temperature. Based on the liquid drop model [9], Nanda et al [8] derived an empirical relation for the size dependent melting temperature of nanoparticles (Tm) in terms of bulk melting temperature (Tmb) as,
Tm = Tmb −
6v 0 γ 0.0005736d
Tm β = 1− Tmb d
(1)
(2)
where v0 is the molar volume, d is the diameter of the particle and γ represents the coefficient of surface energy of the material. Here the particle shape is assumed to be spherical. For a nanomaterial with general shape, the size dependent melting, in general, can be written as
Tm β = 1− Tmb zd
(3)
with z = 1, 3/2, and 3 for nanoparticles , nanowires, and thin films. In equation (3), d represents the diameter in case of nanoparticles and nanowires, whereas it represents the thickness in case of thin films. Similar expressions for size dependent melting for spherical nanoparticles has also been derived from thermodynamic arguments [6, 14] and from a model based on
Velammal College of Engineering and Technology, Madurai
surface phonon instability [15]. Using the known values of vo,
γ and Tmb, the values of β for different elements are estimated by Nanda et al [8] and are tabulated in Table-1. Nanda et al [8] compared their results with the theoretical predictions of Jiang et al [16] and found that both the models [8,16] are consistent for Pb thin films. 1) III.
SIZE DEPENDENT LATTICE PARAMETERS It is proved experimentally and theoretically that the lattice parameters of nanoparticles depend on the particle size [17-21]. For an isolated nanoparticle, the lattice constants are often measured to contract. Montano et al [22] measured the nearest neighbor distance of isolated silver particles with sizes from 2.5 nm to 13 nm and found a noticeable contraction in the nearest neighbor distance. Also, the model developed by Jiang et al [23, 24] based on the Laplace – Young equation supports the size induced lattice contraction. Recently Qi and Wang [25] developed a simple model to account for the size and shape dependent lattice parameters of metallic nanoparticles. They have assumed that a nanoparticle is formed in the three steps. First a particle is taken out from an ideal bulk crystal without changing structure. Secondly the surface tension contracts the particle little elastically and thirdly a nanoparticle is formed in equilibrium. By minimising the sum of the increased surface energy and the elastic energy, Qi and Wang obtained a formula to calculate the lattice parameters of metallic nanoparticles as
Δa 1 =− a 1 + Kd where,
K=
α G γ
(4)
(5)
here, a - is the lattice constant , α – is the shape factor and is equal to one for a spherical particle and G is the shear modulus. Qi and Wang [25] found that the values obtained by them for Au, Ag and Al are in good agreement with the experimental values [26-28]. Also Molecular dynamic simulation studies on the size and shape dependent lattice parameters of unsupported small Pt nanoparticles [29] conclude that for transitional metallic nanoparticles, size is the main factor, which affects the lattice parameter.
IV. SIZE DEPENDENT MELTING TEMPERATURE CONSIDERING LATTICE CONTRACTION In the present work, the size dependent lattice constants for the tungsten nanoparticles are TABLE I SURFACE ENERGY ( γ ), BULK MELTING TEMPERATURE (Tmb), BULK MOLAR VOLUME (vo) AND ( β ) FOR TUNGSTEN [8]
Page 427
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
γ
J/m2 at 298 K 2.753
Tmb K
vo x 10-6 m3
3680
9.53
β
nm
0.772
determined by equation (4). As the model proposed by Qi and Wang [25] assumes that a particle is taken from a bulk crystal without changing the structure, for the formation of nanoparticle, one can imagine that a nanoparticle can have the same structure as that of its bulk counter part. Also the contraction in the lattice constant will lead to the change in molar volume of the nanoparticle [10]. Hence from the corrected values of lattice constants obtained from equation (4), new molar volumes are determined for sizes ranging from 2nm to 30nm. The size dependent melting temperatures for tungsten nanoparticles are determined from equation (1), by incorporating the volume contraction and the results along with the results obtained from specific heat calculations are shown in fig-1.
disturb them and the melting process initiates. Quantitative calculations based on the model are not easy, hence Lindemann offered a simple criterion that the melting might be expected when the root mean square displacement exceeds a certain threshold value ( namely when the amplitude reaches at least 10% of the nearest neighbor distance). Although experiments confirm that the Lindemann criterion is not very accurate, it lends strong support to the idea that the magnitude of vibrational amplitude dominates the phenomenon of melting. Also Lindemann model for vibrational melting is adequate for simplest structures i.e. assemblies of closed packed atoms. Further, there are some discrepancies regarding the Tm dependence of θ D . One opinion is that θ D varies linearly with Tm [31] and the other suggests a square root dependence of θ D on Tm according to Lindemann Criterion of melting [11]. Since there are more supporting evidences for the square root dependence [16], the size dependent θ D and MSDs are determined by the Lindemann theory. Following Lindemann theory, Reisland [13] derived the expressions for the Debye temperature and mean square displacement for a bulk crystal as
M elting temperatu re (K)
4000
⎛ ⎜ T θ D = C Lind ⎜ m2 ⎜ mv 0 3 ⎝
3600 3200 Liquid Drop Model From Sp.ht
2800 2400 2000 0
0.2
0.4
0.6
1/size (nm)
Fig.1 Size dependent melting temperature for nanotungsten. It can be seen that the melting temperatures determined from the idea proposed by Sadaiyandi [10] and the melting temperatures calculated from specific heat data [13] are agreeing well.
u2 =
9h 2 T mk B θ 2D
1
⎞2 ⎟ ⎟ ⎟ ⎠
(6)
(7)
where m is the atomic mass and CLind is the Lindemann constant and is equal to 200 for nonmetals and 137 for metals. T is the temperature at which the MSDs are calculated. Studies on the lattice structure of nanocrystalline samples by means of XRD and Massbauer spectroscopy showed that the lattice structure of the nanometer sized crystallites deviates from the equilibrium state. The deviation may be classified as (i) distorted lattice structures in pure elements and stoichiometric line compounds , and (ii) formation of metastable phase below a critical crystalline size. The lattice distortion in various nanocrystalline materials is manifested by a significant change in lattice parameter [32] and corresponding change in molar volume. Hence the expression for θ D of a bulk crystal can be used for a nanocrystalline material by substituting the size dependent molar volume and melting temperature. Determination of size dependent θ D from specific
V. DEBYE TEMPERATURE AND MEAN SQUARE DISPLACEMENTS The forces between the atoms are reflected in the Debye temperature and it is useful to have this as a reference to characterise a crystal. The Debye temperature is a measure of the vibrational response of the material and therefore intimately connected with properties like the specific heat, thermal expansion and vibrational entropy [30]. Lighter inert gas solids melt below their θ D while the other crystals remain solids above it. The first theory explaining the mechanism of melting in the bulk was proposed by Lindemann [11], who used vibration of atoms in the crystal to explain the melting transition. The average amplitude of thermal vibrations increases when the temperature of the solid increases. At some point the amplitude of vibration becomes so large that the atoms start to invade the space of their nearest neighbors and
heat measurements by Herr et al [33] showed that θ D decreases significantly for nanocrystalline materials. The depressed Debye temperature in nanocrystalline sample implies a decrease in the cohesion of atoms in the nanocrystallites, which agrees well with the measured grain
Velammal College of Engineering and Technology, Madurai
Page 428
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VI. SIZE DEPENDENT MELTING TEMPERATURE AND DEBYE TEMPERATURE FROM SPECIFIC HEAT MEASUREMENTS The specific heat capacity Cv of a solid according to Debye approximation is given as [13], x
4
where
x =
1.3E-02 1.2E-02 1.1E-02 1.0E-02 9.0E-03 8.0E-03 7.0E-03 6.0E-03 0
x
x e C v = 9 Nk B x ∫ x dx 2 0 (e − 1) 3
1.4E-02
M S D x 10 -22 m 2
size dependence of lattice parameter [34]. In the present work the size dependent Debye temperatures are determined from equation (6) by substituting the newly calculated melting temperatures and the size dependent MSDs are determined from equation (7) for tungsten nanoparticles. The results along with the results obtained from specific heat calculations are shown in fig-2 and fig-3.
0.1
0.2
0.3
0.4
0.5
0.6
1/size (nm)
(8)
Fig.3 Size effect on the MSDs of tungsten nanoparticles at 300K. The MSD values are increasing with respect to the reduction in size. Also it can be seen that the size effect is dominant below 10nm.
θ D (T )
(9)
T
where, N- is the Avogadro Number and kB is the Boltzmann constant and θ D (T) is the Debye temperature at a particular temperature. The right hand side of eqn (8) is determined by Simpson’s rule for different values of x (from 1 to 100). For improving accuracy the intervals are taken as smaller. A permanent tabulation is formed connecting the x values and right hand side of eqn.(8).
290
The values of specific heat capacities are collected from [12] for tungsten nanoparticles with sizes 8nm, 10nm and 12nm at different temperatures. The x value corresponding to the collected specific heat at a particular temperature is obtained from the permanent tabulation. Knowing the values of x and T, the corresponding values of θ D (T) are calculated from eqn (9). Now a graph is drawn connecting T and θ D (T). When the curve is extrapolated to meet the Y axis (i.e. T = 0K), the Y intercept gives the Debye temperature of the solid. In order to improve the accuracy, the Y intercept is determined by second order least square fitting method. The fitting curves are shown in fig.4, fig.5 and fig.6 for tungsten with particle size 8nm, 10nm and 12nm.
270
20000
260 Liquid drop model From Sp.Ht.
250 240 230 220 210 200 0
0.1
0.2
0.3
1/Size
0.4
0.5
0.6
Debye Tem perature (K)
Debye Tem perature (K)
280
16000 12000 8000 4000
(nm)
Fig.2 Size dependent Debye temperature for nanotungsten. It can be seen that the Debye temperatures determined from the idea proposed by Sadaiyandi [10] and the Debye temperatures calculated from specific heat data [13] are agreeing well.
0 0
1000
2000
3000
4000
Temperature (K)
Fig.4 Debye temperature for nanotungsten with size 8nm. Points are the calculated values. The solid line is the fitting curve. The fitted value of
θD
is
271K
Velammal College of Engineering and Technology, Madurai
Page 429
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
VII.
RESULTS AND DISCUSSION Determination of size dependent
θ D from specific heat measurements by Herr et al [33] showed that θ D decreases significantly for nanocrystalline materials. The depressed Debye temperature in nanocrystalline sample implies a decrease in the cohesion of atoms in the nanocrystallites, which agrees well with the measured grain size dependence of lattice parameter[34]. 20000
D e b y e T e m p e r a tu r e (K )
18000 16000 14000 12000 10000 8000 6000 4000 2000 0 0
1000
2000
3000
4000
fig.2, it can be seen that the values of melting temperatures and Debye temperatures obtained by the two methods are agreeing well, which shows that the size dependent melting temperatures can be determined by liquid drop model proposed by Nandha et al [8] by considering the idea of size dependent lattice contraction proposed by Qi and Wang [25]. Also fig.1 shows that the melting temperature and Debye temperature of Tungsten decreases with the reduction in size which is in agreement with the earlier predictions [33]. Fig.3 shows the size dependent MSDs at 300K for nanotungsten. It shows that the MSD values are increasing with reduction in size and also it can be seen that the size effect on MSDs is predominant in the sizes ranging from 2nm to 10nm. When the size is above 20nm the changes are too small to be neglected. CONCLUSION VIII. The size effect on θ D for tungsten nanoparticles is determined (i) by the liquid drop model considering the size dependent lattice contraction [10] and (ii) from the available specific heat data [12, 13]. The calculated values by the two methods are agreeing well, which supports the determination of size dependent melting temperature and Debye temperature by the liquid drop model by accounting lattice contraction. The procedure is extended to determine the MSDs of Tungsten nanoparticles for sizes ranging from 2nm to 30nm at 300K.
Temperature (K)
Fig.5 Debye temperature for nanotungsten with size 10nm. Points are the calculated values. The solid line is the fitting curve. The fitted value of
θD
is
273.5K.
2) ACKNOWLEDGEMENT The author wishes to acknowledge the management of Velammal College of Engineering and Technology, Madurai and Udaya School of Engineering, Nagercoil for constant encouragement.
18000
3) REFERENCES
Debye Temperature (K)
16000 14000 12000 10000 8000 6000 4000 2000 0 0
1000
2000 Temperature (K)
3000
4000
Fig.6 Debye temperature for nanotungsten with size 12nm. Points are the calculated values. The solid line is the fitting curve. The fitted value of
θD
is
278K.
In the present work the size dependent melting temperature and Debye temperature are determined by liquid drop model and specific heat calculations. From fig.1 and
Velammal College of Engineering and Technology, Madurai
[1] A.S.Edelstein and R.C.Cammarata, Nanomaterials: Synthesis, Properties and Applications ( Inst. Phys., Bristol, 1996) [2] H.Gleiter, Prog. Mater Sci. 33, 223 (1989) [3] M.Takagi, J.Phys.Soc.Jpn. 9, 359 (1954) [4] S.W.H.Yih and C.T.Wang , Tunsten : Sources, Metallurgy, Properties and Applications , (Plenum, New York1979). [5] E.Lassner and W.D.Schubert, Tungsten : Properties, Chemistry, Technology of the Element , (Plenum, New York 1999) [6] Ph.Buffat and J.P.Borel, Phys.Rev.A 13(6), 2287 (1976) [7] Q.Jiang, N.Aya, F.G.Shi, Appl. Phys. A 64, 627 (1997) [8] K.K.Nanda, S.N.Sahu and S.N.Behera, Phys Rev. A 66, 013208 (2002) [9] C.Brechignac, et al., J.Chem. Phys. 101, 6992 (1994) [10] K.Sadaiyandi, Mat. Chem. Phys 115, 703 (2009) [11] F.A.Lindemann Phys.Z. 11, 609 (1910) [12] A.Moitra, S.Kim, J.Houze, B.Jelinek, S.G.Kim, S.J.Park, R.M.German and M.F.Horstemeyer, J.Phys.D: Appl. Phys, 41, 185406 (2008) [13] J.A. Reisland, The Physics of Phonons, Wiley, London1973. [14] P.R.Couchman and W.A.Jesser, Nature(London) 269, 481 (1977) [15] M.Wautelet, J.Phys.D. 24, 343 (1991) [16] Q.Jiang, H.Y.Tong, D.T.Hsu, OK.Okuyama and F.G.Shi Thin Solid Films 312, 357 (1998) [17] R.Lamber, S.Warjen and I.Jaeger Phys. Rev. B. 51, 10968 (1995) [18] X.F.Yu, X.Liu, K.Zhang and Z.Q.Hu J.Phys.C 11, 937 (1999) [19] C.Solliard and M.Flueli Surf. Sci. 156, 487 (1985)
Page 430
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
[20] V.I.Zubov, I.V.Mamoatov and J.N.T. Rabelo Nanostructure.Mater. 8, 595 (1997) [21] M.Fukuhara Phys. Lett. A 313, 427 (2003) [22] P.A.Montana, W,Schulze, B.Tesche, G.K.Shenoy and T.I.Motrison Phys. Rev.b 30, 672 (1984) [23] Q.Jiang, L.H.Liang and D.S.Zhao J.Pjys.Chem B 105, 6275 (2001) [24] L.H.Liang, J.C.Li and Q.Jiang Physica B 334, 49 (2003) [25] W.H.Qi and M.P.Wang J.Nanopart. Res. 7, 51 (2005) [26] C.W.Mays, Vermank and D.KuhlmannWilsdorf, Surf. Sci. 12, 134 (1968) [27] H.J. Wassermanand J.S.Vermaak Surf. Sci. 22 164 (1970) [28] J. Woltersdorf, A.S.Kepijiko and E.Pippel. Surf.Sci. 106, 64 (1981) [29]W.H.Qi, B.Y.Huang, M.P.Wang, Z.M.Yin and J.Li, J.Nanopart.Res.11(3) 579 (2009) [30] G.Grimvall. Thermophysical Properties of Materials, (North Holland, Amsterdam, 1956) [31] R.C.G.Killean and E.J.Lisher, J.Phy.C: Solid State Phys. 8, 3510 (1975) [32]K. Lu and Y. H. Zhao, Nanostruct. Mater. 12, 559 (1999) [33] U.Herr, M.Geigl and K.Samwer, Philos. Mag. A77, 641 (1998) [34] B.S.Murty, M.K.Datta and S.K.Babi. Sadhana 28, 23 (2003)
Velammal College of Engineering and Technology, Madurai
Page 431
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Dynamic Key Management to minimize communication latency for efficient group communication Dr.P.Alli 1 , G.Vinoth Chakkaravarthy 2 , R.Deepalakshmi 3 1
Department of Computer Science and Engineering, Velammal College of Engg.&Tech,Tamilnadu , India Department of Computer Science and Engineering, Velammal College of Engg.&Tech,Tamilnadu , India 3 Department of Computer Science and Engineering, Velammal College of Engg.&Tech,Tamilnadu , India 1 [email protected], 2 [email protected], 3 [email protected] 2
Abstract:
Introduction:
Group key management is one of the basic building blocks in securing Group Communication (GC), the majority of research in this area mainly concerned with increasing the security while minimizing cryptographic computation cost. Due to the reliable GC platform, network delay is amplified by necessary acknowledgments between the group members. The bottleneck is shifted from computation to communication latency. In this paper, a new Virtual Binary Tree Technique (VBT) to specific group key agreement is proposed which supports dynamic group membership and handles network failures. VBT communication-efficient and secure against hostile eavesdroppers and various attacks specific to group settings. In VBT, common group key is generated based on the share taken from each participant and transmitted to all participants in the team. The member, who is having the group key, is allowed to communicate with other members in the group. Communication latency occurs while authentication is carried out every time between the team of users is reduced. To make this scheme as a dynamic one, when a new user arrives, a new shared secret key is formed , when a person leaves from the group , their share is taken from the existing shared secret key and new shared secret key is formed based on the contents of existing participants and it is broadcasted. So the person who left from the team is prevented to participate in the communication .Thus VBT assures the confidentiality of the team with minimum communication latency.
Security is an important requirement in reliable group communication over open networks in order to prevent intruder attack. The majority of research in group key management was mainly concerned with increasing the security while minimizing cryptographic computation cost. But heavyweight computation such as large number arithmetic that forms the basis of many modern cryptographic algorithms is the greatest burden imposed by security protocols. However, the continuing increase in computation power of modern workstations speed up the heavyweight cryptographic operations. In contrast, communication latency has not improved appreciably. Due to the reliable group communication platform, network delay is amplified by necessary acknowledgments between the group members [2]. However, the continued advances in computing power have not been matched by a decrease in network communication delay. Thus, communication latency, is increasingly dominating the key setup latency, replacing computation delay as the main latency contributor. Hence, there is a need to minimize the size of messages and especially the number of rounds in cryptographic protocols.
Keywords: communication latency, group key, dynamic membership, member join, member leave, Diffie-Hellman key exchange , confidentiality
Velammal College of Engineering and Technology, Madurai
The bottleneck shift from computation to communication latency prompts us to look at cryptographic protocols in a different light: allowing more liberal use of cryptographic operations while attempting to reduce the communication overhead. The proposed scheme in this paper is efficient to avoid such latency with out minimizing cryptographic operations. We consider a protocol, is formed by a virtual tree based approach that extends the 2-party Diffie-Hellman key exchange and supposes the formation of a secure static group[2].This protocol involves heavy computation and communication requirements: O(n) communication rounds and O(n) cryptographic operations are necessary to establish a shared key in a group of ‘n’ members. We extend it to deal
Page 432
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
with dynamic groups and network communication-efficient manner
failures
in
a
Cryptographic Properties: The following cryptographic properties is assured in the group key formation[2].
Group Key Secrecy: It guarantees that it is computationally infeasible for a passive adversary to discover any group key. Forward Secrecy: It guarantees that a passive adversary who knows a contiguous subset of old group keys cannot discover subsequent group keys. Backward Secrecy: It guarantees that a passive adversary who knows a contiguous subset of group keys cannot discover preceding group keys. Key Independence: It guarantees that a passive adversary who knows any proper subset of group keys cannot discover any other group key not included in the subset. Backward and Forward Secrecy properties assume that the adversary is a current or a former group member. From our definition of shared secret key , it would be more desirable to guarantee that any bit of the group key is unpredictable. Thus group key secrecy guarantees that it is computationally infeasible for a passive adversary to distinguish any group key from random number.
Proposed tree based group key generation: The proposed tree structure for group communication [6] is shown in figure: 1. Here all the nodes will be allotted with a value. The leaf nodes are members Ml,M2.....Mn. A member's value is obtained from root node to its member leaf node. By evaluating the values in the nodes, group key is formed for communication. Then the key is broadcasted across the tree structure to all its nodes. In a proposed scheme called Group Key Management Protocol (GKMP) [1] , the entire group shares the same key called as session-encrypting key (SEK).
Figure 1. Tree for Group key formation The tree has two types of nodes: leaf and internal. Each leaf node is associated with a specific group member. An internal node IN(i) always has two children: another (lower) internal node IN(i-1) and a leaf node LN(i). The exception is IN(1) which is also a leaf node corresponding to M1. Each leaf node LN(i) has a session random ri chosen and kept secret by Mi. The blinded version thereof is bri = α ri mod p. Every internal node IN(j) has an associated secret key kj and a public blinded key (bkey) bkj = α kj mod p. The secret key ki (i > 1) is the result of a Diffie-Hellman key agreement between the node’s two children (k1 is an exception and is equal to ri.), which is computed recursively as follows: ki = (bki-1)ri mod p = (bri)ki-1 mod p = α ri ki-1 mod p if i > 1. The group key with the root node is : K4 = The root (group) key is never used directly for the purposes of encryption. Instead, such special-purpose subkeys are derived from the root key, e.g., by applying a cryptographically secure hash function to the root key. All bkeys (bki) are assumed to be public. Group Key Agreement: A comprehensive group key agreement solution must handle adjustments to group secrets subsequent to all membership change operations in the underlying group communication system. The following membership changes are considered: Single member changes include member join or leave, and multiple member changes include group merge and group partition. Join occurs when a prospective member wants to join a group [6]. Leave occurs when a member wants to leave (or is forced to leave) a group[6].
Member joins: Consider, initially the group has n users {M1,M2…. Mn}. When the group communication system announces the arrival of a new member, both the new member and the prior group members receive this notification simultaneously. The new
Velammal College of Engineering and Technology, Madurai
Page 433
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
member Mn+1 broadcasts a join request message that contains its own bkey bkn+1 Upon receiving this message, the current group’s sponsor Mn refreshes its session random, computes brn, kn, bkn and sends the current tree BT(n) to Mn+1 with all bkeys. Next, each member Mi increments n = n + 1 and creates a new root key node IN(n) with two children: the root node IN(n-1) of the prior tree Ti on the left and the new leaf node LN(n) corresponding to the new member on the right. Note that every member can compute the group key since: • All existing members only need the new member’s blinded session random. The new member needs the blinded group key of the • prior group. Step 1: The new member broadcasts request for join Mn+1
brn+1 = α r n+1
event from the group communication system, each remaining member updates its key tree by deleting the nodes LN(d) corresponding to M d and its parent node IN(d) The nodes above the leaving node are also renumbered. The former sibling IN (d-1) of M d is promoted to replace (former) Md’s parent. The sponsor Ms selects a new secret session random, computes all keys (and bkeys) just below the root node, and broadcasts BT(s) to the group. This information allows all members (including the sponsor) to recompute the new group key. Step 1: Every member • updates key tree as described above, • removes all keys and bkeys from the sponsor node to the root node
C = { M1, …. Mn }
Step 2: Every member updates key tree by adding new member node and • new root node, removes bkn, • The sponsor Mn additionally • generates new share rn and computes brn, kn, bkn broadcasts updated tree BT(n) •
The sponsor Ms additionally generates new share and computes all (keys, bkeys) • • and broadcasts updated tree BT(s) C – { Md }
BT(s)
Ms
Step 2: Every member computes the group key using BT(s)
C U { M n+1 } = {M1 , ….., M n+1 } BT(n) Mn Step 3: Every member computes the group key using BT(n)
Figure 3. Leave operation The leave protocol provides forward secrecy since a former member cannot compute the new key owing to the sponsor’s changing the session random. The protocol also provides key independence since knowledge of the new key cannot be used to derive the previous keys; this is, again, due to the sponsor refreshing its session random. Figure 2.Tree update in join
Member Leaves: From a a group of n members when a member M d (d ≤ n) leaves the group. If d > 1, the sponsor Ms is the leaf node directly below the leaving member, i.e., M d-1. Otherwise, the sponsor is M2. Upon hearing about the leave
Velammal College of Engineering and Technology, Madurai
Performance analysis and communication efficiency: The above VBT key management system is analyzed with an existing key management schemes TGDH .The overhead of the TGDH protocol depends on the tree height, the balancedness of the key tree, the location of the joining tree, and the leaving nodes. The number of modular
Page 434
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
exponentiations for a leave event in the VBT depends on the location of the deepest leaving node.
Map",Information and Communications University (ICU), 584, Hwaam-Dong, Yuseong-gu, Daej eon, 305-732, Korea.
Conclusion:
[2] Yongdae,Perrig and Tsudik, "Group Key Agreement efficient in communication", IEEE transactions on computers, Oct 2003.
The proposed key management scheme VBT instantly generates a key. The key is computed as a function of all current group members participates in the communication. In this scheme, the currently operating valid users only can take part in communication and it ensures the presence of users without much difficulty. This scheme supports all dynamic peer group operations: join, leave. Furthermore, it easily handles cascaded (nested) membership events and network failures. Hence the proposed VBT ensures that the efficient communication between the groups with less delay with out minimizing cryptographic computations.
[3] Yair Amir, Yongdae Kim, Cristina Nita-Rotaru, and Gene Tsudik. “On the performance of group key agreement protocols” Technical Report CNDS-2001-5, [4] David A. McGrew and Alan T. Sherman, " Key Establishment in Large Dynamic Groups Using One- Way Function Trees" Cryptographic Technologies Group, TIS Labs at Network Associates, May 20, 1998. [5] C.Becker and U.Wille. "Communication complexity of group key distribution”. In 5th ACM Conference on Computer andCommunications Security, November 1998.
References: [1] Sang won Leel, Yongdae Kim2, Kwangjo Kiml "An Efficient Tree-based Group Key Agreement using Bilinear
[6] Dr.K.Duraiswamy, S.P.Shantharajah “Key Management and Distribution for Authenticating Group Communication” IEEE transactions on computers,2006
Velammal College of Engineering and Technology, Madurai
Page 435
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Integrated Biometric Authentication Using Finger Print And Iris Matching A.Muthukumar1 and S.Kannan2
1. Faculty, Dept of ECE, Kalasalingam University,[email protected] 2. Faculty, Dept of EEE, Kalasalingam University,[email protected] 1
[email protected]
Abstract A biometric measures an individual's unique physical or behavioral characteristics in order to recognize or authenticate his identity. Biometric identification is made up of two stages: enrolment and verification. Automatic and reliable extraction of minutiae from fingerprint images is a critical step in fingerprint matching. The quality of input fingerprint images plays an important role in the performance of automatic identification and verification algorithms. Error-correcting codes are proposed as a means of correcting iris readings for authentication purposes. This paper presents a technique that combined with a bimodal biometric verification system that makes use of finger print images and iris images. Each individual verification has been optimized to operate in automatic mode and designed for security and authentication access application. 1. INTRODUCTION Fingerprint based identification has been one of the most successful biometric techniques used for personal identification. Each individual has unique fingerprints. A fingerprint is the pattern of ridges and valleys on the finger tip. A fingerprint is thus defined by the uniqueness of the local ridge characteristics and their relationships. Minutiae points are these local ridge characteristics that occur either at a ridge ending or a ridge bifurcation[1]. A ridge ending is defined as the point where the ridge ends abruptly and the ridge bifurcation is the point where the ridge splits into two or more branches. Automatic minutiae detection becomes a difficult task in low quality fingerprint images where noise and contrast deficiency result in pixel configurations similar to that of minutiae. This is an important aspect that has been taken into consideration in this project for extraction of the minutiae with a minimum error in a particular location.
Velammal College of Engineering & Technology, Madurai
Fig 1: Finger print and Iris In this paper, error-correcting codes are proposed as a means of correcting iris readings for authentication purposes. We store a hashed version of an individual's biometric template (along with check digits) to verify their identity. A major advantage of this approach is that it is unnecessary to store the individual's biometric template explicitly, a situation which puts the biometric at risk from dishonest systems operators. Given that the potential theft of biometrics is one of the main objections to their usage, any such increase in security is highly desirable. If a valid message is defined to be a codeword, then a code is the set of all code words. The code distance, d, is the minimum number of bits which differ between any two code words. If d=2t+ 1, (1) then up to 2t errors can be detected and up to t of these corrected; if there are t errors or less, the corrupted codeword is
Page 436
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2.1.1.Pre-Processing IRIS
Fin ger prin t
Edge detect ion
Norm alisati on
Norm alisat ion
Frequ ency and Orient ation
Iris Matc hing using Corre i Mask Gener ation and Match
Comb ine Iris and
The input image is segmented from the background which ensures the removal of noise. For this, the whole image is divided into blocks of size 16×16 and the variance of each block is computed. The variance is then compared with the threshold value. If the variance of a block is less than the threshold value, then it is deleted from the original figure. This process is carried out for the whole image. The image obtained from the above step is then normalized to get the desired variance of the given image. The normalized image is given by
Fig 2: Overview of the proposed work automatically closer to one valid codeword than any other, but if there are more than t errors, there is no way of knowing to which codeword one should correct[3]. Therefore, a good code has as large a code-distance as possible, as this enables more errors to be corrected. Each codeword in the code is of the same length, usually represented by n. It is divided into k information bits and (n - k) check digits. Such a code is known as a systematic (n, k) code. As suggested by their name, check digits are used to correct the corrupted codeword to its original value. Finger print with false minutiae removal was studied for enhancement of image[1],[2].Then iris with error correcting codes used for improving the image enhancement and edge detection of the image[3],[4]. 2.PROPOSED WORK
⎧M0 + v(VAR0(I(i, j) - M)2 /(VAR)) ⎪ if I(i, j) > M ⎪ G(i, j) = ⎨ ⎪ M0 - v(VAR0(I(i, j) - M)2 /(VAR)) ⎪⎩ otherwise
where I(i,j) denotes the gray-level value at pixel (i, j), M and VAR denote the estimated mean and variance of I respectively and G(I,j) denotes the normalized gray-level value at pixel (i, j). and VAR0 are the desired mean and variance values respectively. The estimation of the orientation of the image is then carried out as the next step. The whole image is divided into blocks of size 16×16 and the local orientation in the figure is computed by
The whole problem is divided into the following three steps: 1. Pre-Processing 2. Normalization 3. Post-Processing(matching) This fig.2 shows the details of the first step named preprocessing which gives an insight into the process that has been followed by the enhancement of the inputs fingerprint and iris images.
2.1FINGER PRINT MATCHING In this section , image is segmented and it is normalized using gabor filtering are explained. Then in PostProcessing ,how the false minutiae removed was explained.
Velammal College of Engineering & Technology, Madurai
Vx(i,j) = Vy(i,j) =
i +ω/2
j+ω/2
u =i -ω/2
v = j -ω/2
i +ω/2
j+ω/2
∑
∑ 2∂x(u, v)∂y(u, v)
∑ω
∑ω (∂ x(u, v) − ∂ y(u, v))
u =i - /2
2
2
v = j - /2
θ (i, j) = 0.5 tan (Vx(i,j)/Vy(i,j)) −1
where θ(i, j) is the least square estimate of the local ridge orientation at the block centered at pixel (i, j). The angles between the blocks are then smoothened by passing the image through a low pass filter as follows.
X [k ] =
1 ω −1
∑ G(u, v), k = 0,1,...l − 1
ω d =0
Page 437
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 u=i+(d-(ω/2))cosO(i,j)+(k-(l/2))sinO(i,j), v=i+(d-(ω/2))sinO(i,j)+((l/2)-k)cosO(i,j), The following method is adopted for the calculation of the frequency of the local blocks. X-signatures of each block are computed along the direction perpendicular to the orientation angle in each block. The window used for this purpose is of size 16×32. The frequency is then computed by the distance between the peaks obtained in the X-signatures. The window for this is given by the formula
X [k ] =
1 ω −1
∑ G(u, v), k = 0,1,...l − 1
ω d =0
u=i+(d-(ω/2))cosO(i,j)+(k-(l/2))sinO(i,j),
2.1.3Post-Processing The minutiae points obtained in the above step may contain many spurious minutiae. This may occur due to the presence of ridge breaks in the given figure itself which could not be improved even after enhancement. This results in false minutiae points which need to be removed. These unwanted minutiae points are removed in the post-processing stage. False minutiae points will be obtained at the borders as the image ends abruptly. These are deleted using the segmented mask. As a first step, a segmented mask is created. This is created during segmentation carried out in the stage of preprocessing and contains ones in the blocks which have higher variance than the threshold and zeros for the blocks having lower variance. Finally after enhancing and false removal of the minutiae, then finger print is matched with the template to give the person is authenticated or not. This is first matching output of this paper.
v=i+(d-(ω/2))sinO(i,j)+((l/2)-k)cosO(i,j),
2.2.IRIS MATCHING
In general, the frequency of image constitutes has a certain frequency for the hole image and hence the above step can be omitted if the global frequency of the given figure is known.
In this section , image is segmented and it is normalized using gabor filtering are explained. Then in PostProcessing ,how the error correcting are added and how the matching was occurs was explained.
As the next step, each block is filtered along the direction of the orientation angle using the value of the frequency obtained for each block. A Gabor filter is used for this process and a suitable value of local variances is taken for carrying out the process of filtering. A Gabor filter takes care of both the frequency components as well as the spatial coordinates. The inputs required to create a Gabor mask are frequency, orientation angle and variances along x and y directions. Filtering is done for each block using the local orientation angle and frequency. Pre-processing of the image is completed by the steps as mentioned and the enhanced image is obtained. 2.1.2.Normalization The next step after enhancement of the image is the extraction of minutiae. The enhanced image is binarised first in this step. The skeleton of the image is then formed. The minutiae points are then extracted by the following method. The binary image is thinned as a result of which a ridge is only one pixel wide. The minutiae points are thus those which have a pixel value of one (ridge ending) as their neighbor or more than two ones (ridge bifurcations) in their neighborhood. This ends the process of extraction of minutiae points.
Velammal College of Engineering & Technology, Madurai
2.2.1.Pre-Processing & Normalization In this paper, an improved means of matching iris readings was proposed; however, prior to this stage, the segmentation and image processing stages must take place and in this paper, use of Daugman's concept was strongly recommended algorithm for this purpose. Thus, an Iris Code and corresponding mask are produced, but we do not subsequently employ his technique of Hamming Distance measurements for iris matching. The 256-byte mask contains the locations of bits to be omitted from the reading because of occlusions by eyelids, eyelashes and so on. The enrolment reading is strictly controlled by an operator; occlusions which occur in the enrolment mask are extremely likely to re-occur in subsequent readings. Therefore, bit positions included in the enrolment mask are automatically ignored in all further readings for this user, and we "reduce" the iris reading to its valid bits only. This reduced iris data acquired at enrolment becomes the k information bits of a codeword. Check digits are calculated in the usual manner, and the codeword is then hashed and stored in the user's file. Here also first the input image is segmented from the background which ensures the removal of noise with various filters of Prewitt, Laplace and canny operator A Gabor filter is used for this process and a suitable
Page 438
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 value of local variances is taken for carrying out the process of filtering. A Gabor filter takes care of both the frequency components as well as the spatial coordinates. The inputs required to create a Gabor mask are frequency, orientation angle and variances along x and y directions. Filtering is done for each block using the local orientation angle and frequency. Pre-processing of the image is completed by the steps as mentioned and the enhanced image is obtained. Normalization is done by same process which done in finger print process.
Fig.3 Original Fingerprint Image
2.2.2Post-Processing If we choose to work with a Reed-Solomon code over GF(28), this has a maximum length of 255 octets. The ground field of 28 lends a natural ease to dealing with octetbased calculations, an obvious advantage when working with bytes of data. A further advantage is that it is of characteristic two, which means that the operations of addition and subtraction are identical. This greatly simplifies arithmetic in the field, allowing us to make use of the XOR (exclusiveOR) function. Let the code length, n, be equal to 255, the maximum value it can attain in this code. The number of information symbols, k, cannot be greater than 200, as we know that this is the maximum number of valid bytes in a good Caucasian image. In addition, the check digits are effectively unhidden and thus could be accessed by an unscrupulous systems operator. As such, we must ensure that their number does not exceed n/2, as otherwise this would allow the codeword to be deciphered. Finally after finding the results of finger print and iris matching which will given to xor operation to prove that both finger print and iris will be come from same person in order to authenticate the user.
Fig.4 Normalized Image
Fig.5 Enhanced Image-Gabor Filtering 3. RESULTS AND DISCUSSIONS The original image of the finger print is given as input which was shown in Fig.3. Then the image was normalized using Gabor filtering and it was enhanced which was shown in fig.4,5.After enhancing the image, the false minutiae was removed and the given image is matched with stored image for matching which is shown in Fig.6,7. Fig.6 Minutiae obtained after deleting spurious minutiae at the borders
Velammal College of Engineering & Technology, Madurai
Page 439
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Fig.7Final extracted Minutiae after applying the windowing technique The original image of the Iris is given as input which was shown in Fig.8. Then the image was converted into gray scale image which is normalized using Gabor filtering with various operators like laplacian , prewitt and it was enhanced which was shown in fig.9-13.After enhancing the image, the error detecting codes are combined with the binary image and that image is matched with stored image for matching..
Fig.10 Prewitt filter
Fig.11 Laplacian filter
Fig.12 Normalized image Fig.8 Original Image
Fig.12 Enhanced image using edge detection Fig.9Gray scale image
Velammal College of Engineering & Technology, Madurai
Page 440
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 4.CONCLUSION The main benefit of this method is its fast running speed. The method identifies the unrecoverable corrupted areas in the fingerprint and removes them from further processing. By treating the enrolment iris reading as a codeword, and the verification reading as a corrupted codeword, we have shown that it is possible to reliably match iris readings using ReedSolomon codes. Such an approach provides greater security than matching algorithms currently in use, as it only requires that a hashed version of the iris reading be stored, rather than an explicit template.
5.REFERENCES [1] D.H. Ballard, “Generalizing the Hough Transform to Detect Arbitrary Shapes,” Pattern Recognition, vol. 13, no. 3, pp. 111-112,1997 [2] A. Ross, J. Shah, and A. Jain, “Towards Reconstructing Fingerprints from Minutiae Points,” Proc. SPIE Conf. Biometric Technology for Human Identification II, pp. 68-80, 2005 [3] J. Daugman. “High confidence visual recognition of persons by a test of statistical independence” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol.15,no.11, 1148-1161, Nov 1993. [4] J. Daugman. “How iris recognition works”, IEEE Transactions on Circuits and Systems for Video Technology, vol.14 no.1, 21-30,Jan 2004. [5] A. K. Jain, L. Hong, S. Pantanki and R. Bolle, “An Identity Authentication System Using Fingerprints”, Proc of the IEEE, vol.85, no.9,1365-1388, 1997
Velammal College of Engineering & Technology, Madurai
Page 441
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A New Frame Work For Analyzing Document Clustering Algorithms Mrs. J. Jayabharathy* and Dr. S. Kanmani** *Senior Lecturer, Department of Computer Science & Engineering **Professor & Head, Department of Information Technology Pondicherry Engineering College, Puducherry, India [email protected] [email protected]
ABSTRACT Nowadays all paper documents are in electronic format, because of quick access and smaller storage. So, it is a major issue to retrieve relevant documents from the larger database. Clustering documents to relevant groups is an active field of research finding various applications in the fields of text mining, topic tracking systems, intelligent web search engines and question answering systems. This paper proposes a frame work for comparing the existing document clustering algorithm. From the frame work, the details regarding the algorithms, their capabilities, evaluation metrics, data set and performance of the various methods are analyzed. From the analyses it is revealed that majority of the researches consider vector space model for document representation, F-measure, Isim & Esim , Precision and Recall are the frequently used metrics. Document clustering still has future scope in incremental document clustering, semi-supervised clustering, ontology based clustering, topic detection and document summarization. KEY WORDS— Text mining – Document clustering – unsupervised learning – Semi-supervised learning – ontology. INTRODUCTION Information extraction plays a vital role in today’s life. How efficiently and effectively the relevant documents are extracted from World Wide Web is a challenging issue[16]. As today’s search engine does just string matching, documents retrieved may not be so relevant according to user’s query. A good document clustering approach can assist computers in organizing the document corpus automatically into a meaningful cluster hierarchy for efficient browsing and navigation, which is very valuable for overcoming the deficiencies of traditional information
Velammal College of Engineering & Technology, Madurai
retrieval methods. If documents are well clustered, searching within the group with relevant documents improves efficiency and reduces the time for search. A good document clustering algorithm should have high intra-cluster similarity and less inter- cluster similarity. i.e the documents with the clusters should be more relevant compared to the documents of other clusters[12]. Document Clustering is an unsupervised approach, in which documents are automatically grouped into predefined number of groups based on their content. Document clustering methods are categorized as partitioning, hierarchical, densitybased, grid-based, model-based and etc. In this paper we mainly concentrate on partitional and hierarchical based clustering methods. This paper gives detailed view about the various document clustering algorithms with respect to the similarity measures, data set and performance metrics. The following section discusses about the procedure of general document clustering algorithm, section 3 gives the detailed survey of document clustering algorithms. Section 4 gives the overview of the proposed framework for comparing the existing document clustering algorithms. Section 5 concludes the paper and discusses about the challenges in document clustering. A GENERAL CLUSTERING
PROCEDURE
FOR
DOCUMENT
Document clustering problem could be expressed as follows: Let N be the total number of documents to be clustered. The set D={D1, D2, ….DN} be the set of N documents [25]. The documents are grouped into non-overlapping clusters C= { C1, C2, ….Ck} where k denotes the number of groups formed. C1U C2U….U Ck = D, Assume Ci ≠ Φ and Ci ∩ Cj = Φ where i ≠ j. Widely used document representation methods are VectorSpace Model (VSM), Multiword term, Character N-gram
Page 442
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
representation are some the existing representation models. VSM is identified as the most widely representation for documents. Using VSM, the document d is represented by the term-frequency (tf ) vector dtf = (tf1; tf2; : : : ; tfm), where tfi is the frequency of the ith term in d. Currently, the TF-IDF (Term Frequency times Inverse Document Frequency) model is the most popular model. In the TF-IDF model, a term ti is weighted by its frequency in the document tfi times the logarithm of its inverse document frequency, i.e., tfi * log2(N/n), where N is the number of documents in the collection and n is the number of documents where the term ti occurs at least once. Using the TF-IDF model, terms that appear too rarely or too frequently are ranked lower than the other terms. Widely e-documents suffer from high dimensionality; document representation [15] is itself a challenging issue. Usually document sources are of unstructured format, transforming these unstructured documents to structured format requires some preprocessing steps. Figure 1. Represents the sequence of steps involved in document clustering. Commonly used pre-processing steps are Tokenization: Tokenization is the process of mapping sentences from character strings into strings of words. For example, the sentence “This is a Survey Paper” could be tokenized into \This, \is, \a, \Survey, \Paper. Data cleaning: Non-textual information, such as HTML tags, punctuation marks and digits, are removed from the document. Stopwords removal: Stop words are typical frequently occurring words that have little or no discriminating power, such as \a", \about", \all", etc., or other domain-dependent words. Stopwords are often removed. Stemming: Removes the affixes in the words and produces the root word known as the stem [17]. Typically, the stemming process will be performed so that the words are transformed into their root form. For example connected, connecting and connection will be transformed into connect. Most widely used stemming algorithms are Porter [21], Paice stemmer [ 20], Lovins [19], S-removal[18]
Velammal College of Engineering & Technology, Madurai
Figure 1 General Procedure for Document Clustering Document Clustering methods [3] are categorized as partitioning, hierarchical, graph-based, model-based and etc. Partitional Clustering – It divides the data objects into nonoverlapping subsets (clusters) such that each data object is in exactly one subset. Hierarchical clustering is a set of nested clusters organized as a hierarchical tree. In this approach, a hierarchy is built bottom-up by iteratively computing the similarity between all pairs of clusters and then merging the most similar pair. Graph based model – This algorithm transforms the problem so that graph theory can be used [4]. In this approach, the objects (documents or words) to be clustered can be viewed as a set of vertices. Two vertices are connected with an undirected edge of positive weight based on certain measurement. In the clustering process, a set of edges, called edge separator, are removed so that the graph is partitioned into k pair-wise disjoint sub-graphs. Objective of the partitioning is to find such separator with a minimum sum of edge weights. Clustering algorithms are based on certain similarity measures [2]. There are many different ways to measure how similar are the two documents. Similarity measures are highly dependent on the choice of terms to represent text documents. Accurate clustering requires a precise definition of the closeness between a pair of objects, in terms of either the pair-wise similarity or distance. Cosine similarity and Jaccard correlation coefficient are widely used similarity or
Page 443
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
distance measures. Euclidean distance and entropy are being applied to calculate pair-wise distance between objects in clustering. CLUSTERING ALGORITHMS This section gives a detailed survey of recent work in document clustering. Benjamin C.M Fung et al [1] elaborated Frequent itemset based hierarchical clustering [22] technique for addressing the problems in hierarchical document clustering. This technique uses frequent item sets to construct clusters and to organize clusters into a topic hierarchy. They have experimentally evaluated FIHC and compared with Hierarchical Frequent Term based clustering (HFTC) proposed by Beil, Ester & Xu, [23] and Unweighted Pair Group Method with Arithmetic Mean (UPMA) proposed by Kaufman & Rousseuw, [24]. This paper concludes that FIHC outperforms well with respect to time, accuracy and scalability. A quantitative and qualitative method was introduced by Yi Peng et al [5] which extract phrases/noun from the documents for clustering. Two hypotheses are designed as follows 1. Instead of single word, this approach uses noun/phrases and words representation and used an optimization method from CLUTO tool kit to group the documents. 2. More meaningful subject headings of document collection could be obtained by merging the similar, smaller clusters instead of generating larger clusters directly. The data set considered for clustering is Data Mining and Knowledge Discovery Publications. This data set was clustered as 8-way and 100-way category using CLUTO tool kit and Intra-cluster similarity and Inter-Cluster similarity are considered as performance metrics. The author concludes that the clustering documents using phrase/word representation generates more meaningful appropriate clusters compared to single word representation and merging smaller and similar documents yields more relevant documents. A framework named SS-NMF (Semi Supervised Non-negative Factorization) which incorporates user defined constraints for document clustering was proposed by Yanhua Chen et al [6], users are able to provide supervision for clustering in terms of must-link and cannot link-link pairwise constraints. An efficient iterative algorithm had been considered for computation of symmetric tri-factorization of the document-document similarity matrix to infer the document clusters. MEDLINE Database, Reuters-21578 Text Collection and Foreign Broadcast Information Service Data are the data sets utilized. This SS-NMF frame work has been compared with supervised and semi-supervised
Velammal College of Engineering & Technology, Madurai
clustering algorithms k-means, kernel k-means, Spectral Normalized Cuts, Non-negative Factorization (NMF), Semisupervised kernel k-means (SS-KK), semi – supervised spectral clustering with normalized cuts (SS-SNC). Confusion Matrix and Accuracy as evaluation metrics and concludes that SS-NMF works better than other supervised and unsupervised algorithms. Jiangtao Qiu [7] et al proposed a Topic Oriented Semantic Annotation algorithm. The author considers semisupervised document clustering approach. User’s need should be represented as multi-attribute topic structure. Using the topic structure, the next step is to compute topic semantic annotation for each document and topic semantic similarity between documents. Ontology based Document topic semantic annotation (ODSA) algorithm was proposed for computing topic semantic annotation for each document. They have also proposed a dissimilarity function using which the dissimilarity matrix have to be constructed. The dissimilarity of two documents is their topic-semantic dissimilarity in topic oriented document clustering. Then the documents are clustered using proposed Optimizing Hierarchical clustering algorithm which is based on agglomerative hierarchical document clustering algorithm. They have also proposed a new cluster evaluation function DistanceSum. In this paper, a comparison between ODSA and Term Frequency and Inverse Document Frequency (TFIDF) representation have been experimented and proved that ODSA takes less time for constructing dissimilarity matrix when compared to TFIDF. They have also assessed the quality of the cluster using the F-measure as the performance metric, i.e. the clustering algorithm based on ODSA and TFIDF have been experimented and judged. The discovery of Topic by Clustering Keywords without any prior knowledge was proposed by Christian Wartena and Rogier Brussee [8]. Identification of most informative keyword is done by using natural probability distributions and the documents are clustered according to various similarity functions like Cosine similarity, JensenShannon divergence distance between document distribution and term distribution. A collection of 8 Wikipedia topics are considered as data set. Considering F-measure as the performance metrics, the experiments show that distribution of terms associated with keyword gives better results. An Active Learning Framework for Semisupervised document clustering with language modeling [9] proposed by Ruizhang Huang and Wai Lam incorporates user defined constraints for clustering. This paper mainly deals with semi-supervised document clustering. This framework concentrates on pair-wise constraints for user provided information i.e. user can provide hierarchical constraints for
Page 444
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
indicating the parent to child relationship between documents. Using these constraints the clustering processes learns actively. Also they have devised a gain function for selecting document pairs. Instead of document represented as bag of words they include term-term dependence relationship using term co-occurrences. Term- Term dependency is used for estimating the terms related to the same cluster. A collection of documents from 20-newsgroup corpus, Reuters RCV1 corpus TDT3 (Topic detection and Tracking) are taken as dataset for experiments. NMI (Normalized Mutual Information) and Pair-wise F-measure are considered as performance evaluation metrics and compares the proposed framework with Active semi-supervision for pair-wise constrained clustering algorithm and concludes that the proposed approach outperforms. Document Clustering Description Extraction and its Applications [10] is a problem of labeling the clustered documents. This paper mainly focuses on automatic labeling concept based on machine learning. Clustering problem transformed into classification for labeling the clusters. Considering two Benchmark models Baeline1 and Baseline1 and SVM (support vector machine), Multiple Linear Regression (MLR) and Logistical regression models (Logit) machine learning approaches are considered for labeling. More than 2,000 academic documents from Information centre for Social Sciences of RUC were taken as the dataset for experiments. Considering Precision, recall and F1 value as performance metrics, the author justifies that SVM outperforms better compared to the other models. Baseline1 model behaves worst compared to five models. Yanjun Li et al [13] propose two clustering algorithms, Clustering based on Frequency of Words (CFWS) and Clustering based on Frequency of Word Meaning Sequence (CFWMS) instead of considering bag of words representation. Word is the one which is simply represented in the document, whereas word meaning is the concept which expressed by synonyms of word forms. Word meaning sequence is the sequence of frequency of the word occurring in the documents. Authors have addressed many issues document representation, dimensionality reduction, labeling the clusters, overlapping of clusters. They preferred Generalized suffix tree for representing the frequency words in each documents. Dimensionality is reduced because of word meaning representation instead of words. CFWS is evaluated to be better algorithm compared to bi-secting kmeans and FIHC with respect of accuracy, overlapping
Velammal College of Engineering & Technology, Madurai
cluster quality and self labeling features. CFWMS achieves better accuracy when compared to CFWS, and modified Bisecting k-means using background knowledge (BBK)[11] Another significant topic in document clustering is document summarization, gives only the main points of the original document. This helps the user to grasp the contents of the entire documents in a moment. Ontology Enchanced clustering based summarization of medical documents[16] was proposed by Kokilavani and Balasubramanie. They have chosen MESH descriptor set for matching string and used kmeans algorithm for clustering. The summary generated was evaluated using Precision and Recall as metrics. Khaled B. Shaban [11] proposed a new frame work namely a Semantic Approach for Document clustering, which adopts meaning based representation of text, and their use in measuring similarities between documents using the knowledge representation structure. Author has used Semantic Graph Model (SGM) for document representation, which is similar to graph based representation. SGM is based on syntactic and symantic information extracted from the documents. For comparative analysis the author takes Vector Space Model and evaluated the performance with respect to F-measure, Entropy, overall similarity and concludes that keywords alone cannot give better cluster quality. He has also proposed a similarity measure precisely for semantic based mining. FRAME WORK FOR ANALYZING THE EXISTING DOCUMENT CLUSTERING TECHNIQUES This study mainly highlights the recent research work in the field of document clustering. This paper mainly focuses about the proposed frame work for comparing various document clustering techniques. The comparative study is based on the survey, which is made by analyzing the existing algorithms, considering the topics like Document clustering technique, Data set used for experiments, Performance metrics and summarization about the performance of the proposed technique. Column 1 in this frame work states the title of the survived papers. Frame work, algorithm and techniques which are discussed in the existing papers are stated in column 2. The 3rd column gives the details of the data set which is considered for conducting the experiments. Metrics considered by the authors for performance evaluation are given in column 4. The concise details about the performance of the algorithms are listed in column 5.
Page 445
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PROPOSED FRAME WORK PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
PERFORMANCE
Recent Trends in Data Mining (DM): Document clustering of DM Publications – 2005
K-means Implemented in CLUTO - Toolkit
1436 papers from different data mining publications
Human Inspection
Forms 47 clusters according to the category
Hierarchical Document Clustering - 2006
Frequency Itemsetbased Hierarchical clustering (FIHC)
100k documents
F-measure
Improved Accuracy
A Hybrid Strategy for Clustering Data Mining Documents- 2006
A qualitative method which extract noun/phrases instead of words from the documents.
Books, thesis & Publications of Data mining and knowledge discovery 1626 papers.
Isim – Intracluster s Similarity Esim – intercluster similarity
Clustered into 8 and 100 categories and concluded that 100way clustering gives quality clustering- Human inspection
Incorporating User Provided constraints into Document clustering's - 2007
SS-NMF : semi supervised nonnegative matrix factorization framework
MEDLINE Database, Reuters-21578 Text Collection and Foreign Broadcast Information Service Data
Confusion matrix Accuracy metric
Concludes that SSNMF outperforms well when compared with semi supervised and un supervised clustering methods
Topic Oriented Semi- Supervised Document Clustering - 200 7
Ontology based Document topic semantic annotation (HOWNET ontology knowledge base)
Chinese web pages about “Li Ming” – same name different personality football player who belongs different clubs
Time analysis for computing Dissimilarity matrix and Dimensionality analysis
Compares with TFIDF w.r.t time (construction of dissimilarity matrix ) And dimensionality analysis (less)
A document Clustering and Ranking system for Exploring MEDLINE
CLUTO tool kit
Society of surgical oncology Bibliography– 10 categories of
Document clustering is a part of Information retrieval in this
This framework retrieves reduced set of articles because of clustering, topic extraction and ranking methods
Bisecting K – means clustering algorithm
Velammal College of Engineering & Technology, Madurai
Scalability
and
Page 446
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
citations - 2007
Spectroscopy
cancer
paper
Text Document Clustering based on Frequent word meaning sequences
Clustering based on Frequent Word Sequences(CFWS) and Clustering based on Frequent Word Meaning Sequences (CFWMS) – CLUTO kit
9 categories from Reuters data set, Classic data set and 200 documents from search engine results.
F-measure Purity
CFWS & CFWMS – has better accuracy compared to Bisecting K-means and Frequent Itemset based Hierarchical clustering(FIHC)
Similarity Measures for Text Document Clustering
Similarity measure – Euclidian, Jaccard coefficient,Cosine simiarity, Pearson correlation coefficient, Average Kullback – Leibler Divergence
20 newsgroups, Classic – academic papers, wap – web pages, web knowldege base - webkb
Purity Entropy
Except Euclidian other S.M are quite similar. Jaccard and Pearson are slightly better with higher purity value
Research Field Discovery Based on Text Clustering
Newman Fast clustering algorithm for network model
Research proposals of National Natural Science and Foundation – 3423 relevant proposals
Modularity
Gives insight of the year wise research field .
Topic Detection by Clustering Keywords
Identifies most informative keyword by Probability , Clustering using Bisecting K-means algorithm
Wikipedia articles of various categories like architecture, popmusic etc
F-measure
Implemented using 3 similarity measures like Cosine, Jensen shanon divergence for term and document distribution. Jensen shanon for term gives better result
Velammal College of Engineering & Technology, Madurai
PERFORMANCE
Page 447
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
PERFORMANCE
An Active Learning Framework for Semi-supervised document Clustering with language modeling
DevisedGain Function for selecting documents pairs, Language Model, Term-Term dependence instead of bag of words Designed a Framework – Active Semi
20 newsgroups, Reuters – newswire stories And TDTTopic detection and detection project
F-measure
Proposed Active Semi gives better clustering performance
Document Clustering Description Extraction and its Applications
Uses Machine Learning approach like SVM, Multiple Linear Regression Model and Two Benchmarks – Baseline1 and 2 for Cluster Description
Documents from Information Center for Social Science of RUC – 2000 documents
Precision Recall
SVM performs better compared to other 4 Machine Learning methods
A Semantic Approach for Document Clustering
A Framework designed – Parses document syntactically, Semantically converts to Semantic Graph Model Algorithm – KNearest Neighbor, HAC, Single Pass Clustering
Reuters Set
F-measure Entropy
Quality of Clustering outperforms well compared to VSM
Ontology Enhanced Clustering Based Summarization of Medical Documents
K- means and MesH ontology descriptor for matching sentences for summarization
PubMed documents
Precision Recall
MesH descriptor based summarization gives better summarization
Velammal College of Engineering & Technology, Madurai
Data
NMI (Normalized mutual information)
Page 448
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
PAPER TITLE
ALGORITHM/ TECHNIQUES
DATA SET
METRICS
PERFORMANCE
Text Document Clustering Based on Neighbors
3 methods : 1. new method for selecting initial cluster centroid 2. Similarity – links and cosine 3. Heuristic function for spiting the cluster
Reuters and MEDLINE
F-measure Purity
Compared with K-Means, Bisecting K-Means and with proposed techniques. Gives better results.
A Comparison of Document Clustering Technique - 2000
Similarity Measure – Centroid, UPGMA and Inter cluster similarity
Reuters, web AceA Web Agent for Document Categorization and Exploration
F-measure
Compares Bi-secting kmeans is better than UPGMA and K-means.
From the study it is inferred that more than 70% of the document clustering algorithms are based on Partitional and hierarchical based algorithms. It is understood that k-means for document clustering dominates in late 1990’s. Later Bisecting k-means algorithm produces good cluster quality compared to k-means. Recent researches mainly concentrates on Ontology based document clustering
Velammal College of Engineering & Technology, Madurai
techniques. It is also inferred that supervised document clustering techniques drives more importance than normal clustering algorithms. Nowadays we have broad scope in supervised document clustering, topic detection and in ontology based document clustering. From the survey it is understood that most of the document clustering algorithms are implemented in CLUTO tool kit.
Page 449
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
P recision & R ecall
35
C onfusion matrix 30 25
Inter&Intra clus similarity
20
E ntrophy & Modularity NMI
15 Misclassification index
10
P urity 5 F ‐measure 0 ME T R IC S
Figure 2 Percentage of Metrics considered for documents clustering algorithms The graph in Figure 2 states the various metrics considered in document clustering in X-axis and percentage of references in Y-axis. More than 30% of the document clustering algorithms concentrates on Intra and Inter cluster similarity measure. 30% of the authors prefer Precision and Recall which gives the quality about the document relevancy. 20% of the document clustering algorithm considers Fmeasure and Purity as performance metrics. Around 10%, the qualities of some clustering techniques are evaluated by Human intervention. 5% to 10% researchers consider (Normalized mutual information) NMI, Modularity, Confusion matrices as performance metrics. Figure 3 shows about the choice of data sets for document clustering with the percentage of usage. From the frame work it is analyzed that majority of the researcher’s attention has been drawn by 20 newsgroups and MEDLINE data set for experiments. Data documents from Reuters which has categories like ORGs, PEOPLE, PLACES etc has received next importance. Wikipedia document sets receive next priority. PubMed and other data sets receive next
Velammal College of Engineering & Technology, Madurai
precedence. From the analyzes it is understood that recent researchers mainly concentrate on semi-supervised.
Page 450
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
35 30 25
R euters
20
20News group
15
Medline
10
W kipedia
5
C las s ic
0
P ubmed
D ata s ets for experiments Figure 3 Data sets for documents clustering algorithms
CONCLUSION In this paper a frame work to compare existing document clustering algorithms was proposed, which gives the summarization of recent research work on various algorithms in document clustering. This frame work precisely states the details about the algorithms, data sets, metrics and performance. Document clustering still has scope in various issues like incremental document clustering, semi-supervised clustering, topic detection and in document summarization.
REFERENCES 1.
2. 3. 4.
5.
Benjamin C. M. Fung, Ke Wang, Martin Ester, “Hierarchical Document Clustering”, The Encyclopedia Of Data Warehousing and Mining,2005. Anna Huang, “Similarity Measures for Text Document Clustering,”, NZCRSC, 2008. Jiawei Han and Micheline Kamber. “Data Mining: Concepts and Techniques., Mogan Kaufmann”, San Francisco, CA, USA, 2001. Eui-Hong Han, George Karypis, Vipin Kumar, and Bamshad Mobasher. “Clustering based on association rule hypergraphs”, In Research Issues on Data Mining and Knowledge Discovery, 1997. Yi Peng, Gang Kou, Yong Shi and Zhengxin Chen, “ A Hybrid Strategy for Clustering Data Mining
Velammal College of Engineering & Technology, Madurai
6.
7.
8.
9.
10.
11.
Documents”, Sixth IEEE International Conference on Data Mining – Workshop (ICDMW’06). 2006. Yanhua Chen, Manjeet Rege, Ming Dong, Jing Hua, “Incorporating User Provided constraints into Document clustering's”, - Proc. of IEEE International Conference on Data Mining, 2007. (Regular Paper). Jiangtao Qiu, Changjie Tang, “Topic Oriented SemiSupervised Document Clustering”, SIGMOD2007 Ph.D. Workshop on Innovative Database Research 2007(IDAR2007). Christian Wartena, Rogier Brussee, “Topic Detection by Clustering Keywords”, IEEE -19th International Conference and Expert System Application,2008 Ruizhang Huang, Wai Lam, “An Active Learning Framework for Semi-supervised document Clustering with language modeling”, Data & Knowledge Engineering,Elesiver 2008. Chenzhi Zhang, Huilin Wang, Yao Liu, Hongjiao Xu, “Document Clustering Description Extraction and its Applications”, Computer Processing of Oriental Languages. Language Technology for the Knowledge-based Economy, 22nd International Conference,ICCPOL 2009, Springer 2009. Khaled B. Shaban, “A Semantic Approach for Document Clustering”, Journal of Software 2009.
Page 451
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
12.
13.
14.
15. 16.
17. 18.
M. Steinbach, G. Karypis, and V. Kumar. “A comparison of document clustering Techniques”, In Proceedings of Workshop on Text Mining, 6th ACM SIGKDD International Conference on Data Mining (KDD’00), pages 109–110, August 20–23. Yanjun Li, Soon M. Chung, John D. Holt, “Text Document Clustering based on Frequent word meaning sequences”, Data Knowledge Engineering, Elesiver 2007. A.Hotho, S. Staab, and G. Stumme, “Ontology improves tect document clustering”, Proceedings of 3rd IEEE International Conference on Data Mining”,2003. Pp 551-554. Document Representation and Dimension Reduction for Text Clustering A. A. Kogilavani, Dr. P. Balasubramanie, “Ontology Enhanced Clustering Based Summarization of Medical Documents”, International Journal of Recent Trends in Engineering, 2009 WB Frakes, CJ Fox , “Strength and Similarity of Affix Removal Stemming Algorithms”, ACM SIGIR Forum, 2003 Harman, D. "How Effective is Suffixing." Journal of the American Society for Information Science 42 (1), 1991, 7-15.
Velammal College of Engineering & Technology, Madurai
19. 20. 21. 22 .
23.
24. 25.
Lovins, J. B. "Development of a Stemming Algorithm." Mechanical Translation and Computational Linguistics 11, 1968, 22-31. Paice, Chris D. "Another Stemmer." SIGIR Forum 24 (3), 1990, 56-61. Porter, M. F. "An Algorithm for Suffix Stripping."Program 14, 1980, 130-137. Fung B, Wnag K & Ester .M, “Hierarchical Document Clustering using Frequent itemsets”, SIAM International Conference on Data Mining, SDM ’03.2003. Pp 59-70. Beil F, Ester M & Xu. X , “ Frequent Term Based Test Clustering”, International Conference on Knowledge Discovery and Data Minig, KDD’02. 2002.pp 436-442. Kaufman, L & Rousseeuw P.J, “ Finding groups in data: An introduction to cluster analysis”, John Wiley & Sons, New York. 1990. Zhaogu Xuen, Li chen, Yanzhangg Dang, Jiang Yu, “ Research Field Discovery Based on text clustering”, IEEE 2008.
Page 452
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Optimization of Tool Wear in Shaping Process by Machine vision system Using Genetic Algorithm S.Palani¹ G.Senthilkumar² S.Saravanan³ J.Ragunesan³ 1. Professor, Mechanical department, Mount Zion Engineering College, Pudukkottai. [email protected]
2. Lecturer, Mechanical department, Velammal Engineering College, Madurai. [email protected]
3. Student, Mechanical department, Sudharsan Engineering College, Pudukkoattai. [email protected]
3. Student, Mechanical department, Sudharsan Engineering College, Pudukkoattai. [email protected]
Abstract Optimization has significant practical importance particularly for operating in machines. In order to increase the accuracy of finishing product the tool must be in good condition always as much as possible. The cutting tool must be in good condition when go for high precision of components i.e. the tolerance is very closer there the optimized cutting conditions were needed. To achieve good condition of tool, the machining parameters like speed, feed, depth of cut and average gray intensity level should be optimized. This paper aims to find out the safety cutting conditions for minimizing the tool wear by applying the optimized input parameters using genetic algorithm technique. Introduction The growing demand for product quality and economy necessity have forced incorporation of monitoring the process parameters in automated manufacturing systems. The greatest limitation of automation in the machining operation is tool wear. If the tool wear increases then the tool life will be minimum. So the tool wear has been optimized by selecting the optimal cutting parameters such as speed and depth of cut. The main machining parameters which are to be considered as variables of the optimization are speed and depth of cut, the required output is minimum tool wear. The optimization of tool wear is very important in modern machining processes. The cutting tool is the most critical part of the machining system, and recent advances in tool technology mean that many traditionally difficult materials
Velammal College of Engineering & Technology, Madurai
can now be more readily machined. Added to his, the desire for faster, more flexible manufacturing systems to improve productivity places even greater demands on cutting tools. Reliable wear data can be used not only for adaptive process control and optimization. Tool wear monitoring can also be used to determine the optimum conditions for a given machining process, such as finding the most economical cutting speed. The optimum set of these three parameters are determined for a particular job-tool combination of High Speed Steel-Mild Steel during which optimizes the tool wear. Many graphical and analytical techniques were initially used for optimization. Some of them are overcome by numerical optimization methods which may be either simplex or gradual based methods. These and other similar numerical optimization techniques have been found to give reasonably accurate results only for simply problems. But in the case of highly nonlinear optimization problems, the numbers of iterations are required to get optimum point, thus consuming significant computation time. So an alternative method for optimization is probabilistic methods. Genetic Algorithm is searches from a population and gives a set of reasonably good solution to the user. Computer vision technique Computer vision techniques have many particular advantages to offer when used for tool wear optimization. The measurement is non-contacting. And the magnification can also be obtained based on the power of the lens used. Hence clarity of the image obtained is more. Computer vision techniques were used to include a parameter affecting the
Page 453
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
tool wear i.e. the gray levels. With the given speed, feed, and depth of cut, shaping operation is performed on work piece, and the image of this tool is captured through the machine vision system. The image is imported to the computer and image processing techniques, the average gray level is found out. Genetic algorithm A genetic algorithm (GA) is a search technique used in computing to find exact or approximate solutions to optimization and search problems. Genetic algorithm is a global population search technique based on the operations of natural genetics and mimics the natural biological process. Genetic algorithm first encodes all variables into a finite bit binary string called as chromosomes. Chromosomes represent a possible solution of chromosomes are formed. Each chromosome is decoded and is evaluated according to the fitness function.
Fig: Structure of Genetic algorithm Input parameters
Table: Data’s for computing the regression equation
The following equation called as the regression equation and it is used as fitness function in Genetic programming. TW = - 0.779 - 0.000733 S + 12.5 F + 1.44 D + 0.00221 AG + 0.00131 sd - 20.8 fd
Velammal College of Engineering & Technology, Madurai
Page 454
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Training parameters
Comparison of measured tool wear with GA predicted tool wear
Experimental results
The result shows the average error of the prediction of tool wear in shaping using GA is 2.761 % i.e. the accuracy is 97.239% Optimized cutting conditions
Velammal College of Engineering & Technology, Madurai
Page 455
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Conclusion In this research an adaptive Genetic algorithm has been used to predict the optimum cutting conditions of a tool and tool wear in machining operation by employing the GA learning algorithm. The following conclusions can be drawn from this research. The reason for introducing the Genetic algorithm network technique in the present study is that the machining process is complex and uncertain in nature; the present machining theories are not adequate for the purpose of practical problem.
References 1. Alegre.E, Alaiz.R, Barreiro.J & Ruiz.J, Assessment and visualization of machine tool wear using computer vision.(2008) 2. C.Flix Prasad, S.Jayabal & U.Nataraajan.Optimization of tool wear in turning using genetic algorithm. (2007) 3. C.Bradly & Y.S.Wong. Surface texture indicators of tool wear-A machine vision approach.(2001) 4. Sukhomay pal, P.Stephan Heyns, Burkhard H. Freyer, Nico J.Theron & Surjya K.Pal. Tool wear monitoring and selection of optimum cutting conditions with progressive tool wear effect and input uncertainties.(2009) 5. H. H. Shahabi & M. M. Ratnam. Assessment of flank wear and nose radius wear from workpiece roughness profile in turning operation using machine vision. (2008) 6. Rick Riolo and Bill Worzel. Genetic Programming Theory& Practice.(2006) 7. Wen-Tung chein & Chung-shay Tsai. The investigation on the prediction of tool wear and determination of optimum cutting conditions in machining of stainless steel.( 2003)
8. D.Venkatesh, KKannan & R.Saravanan. A Genetic Algorithm based model for the optimization of machining process.(2009) 9. R.Saravanan, PAsokan & MSachidhanandham. Multi objective GA approach for optimization of tool wear.(2002)
Velammal College of Engineering & Technology, Madurai
Page 456
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
A Comparative Study Of Various Topologies And Its Performance Analysis Using Wdm Networks P. Poothathan1, S. Devipriya2 , S. John Ethilton3 P.Poothathan1 ,Department of Physics, Velammal College of Engg. Tech., Madurai-9, India. [email protected]. S.Devipriya2 , Department of Physics, Velammal College of Engg. Tech., Madurai-9, India. [email protected] S.John Ethilton3 ,Department of Physics, Velammal College of Engg. Tech., Madurai-9, India. [email protected]
ABSTRACT Optical wavelength division multiplexing (WDM) networking technology has been identified as a suitable candidate for future wide area network (WAN) environments, due to its potential ability to meet rising demands of high bandwidth and low latency communication. Networking protocols and algorithms are being developed to meet the changing operational requirements in future optical WANs. Simulation is used in the study and evaluation of such new protocols, and is considered a critical component of protocol design. In this paper, we construct an optical WDM network simulation tool, which facilitates the study of switching and routing schemes in WDM networks. This tool is designed as an extension to the network simulator ns2. In this work, the effectiveness of the proposed approach has been verified through numerical example and simulated results for various network scenarios, such as ring, mesh and interconnection topologies and we analyse the blocking probability of distributed light path establishment in wavelength-routed WDM networks with no wavelength conversion. We discuss the basic types of connection blocking: i) due to the dimension of the network. ii) due to offered network load. iii) due to network load in Erlang and Packet delay due to load in Erlang. Key words: Optical Networking, WDM network, topology, simulator, blocking probability, delay.
I. INTRODUCTION Optical fiber communications was mainly confined to transmitting a single optical channel until the late 1980s. Because the fibre attenuation was involved, this channel required periodic regeneration, which included detection, electronic processing, and optical retransmission. Such regeneration causes a high-speed optoelectronic bottleneck and can handle only a single wavelength. After the new generation
Velammal College of Engineering and Technology, Madurai
amplifiers were developed, it enabled us to accomplish highspeed repeaterless single channel transmission. A WDM system enables the fiber to carry more throughputs. By using wavelength-selective devices, independent signal routing also can be accomplished. Two common network topologies can use WDM, namely, the mesh and the ring networks. Each node in the mesh has a transmitter and a receiver, with the transmitter connected to one of the central passive mesh’s inputs and the receiver connected to one of the mesh’s outputs. WDM networks can also be of the ring variety. Rings are popular because so many electrical networks use this topology and because rings are easy to implement for any network geographical configuration. In an ideal WDM network, each user would have its own unique signature wavelength. Routing in such a network would be straightforward. This situation may be possible in a small network, but it is unlikely in a large network whose number of users is larger than the number of provided wavelengths. In fact, technologies that can provide and cope with 20 distinct wavelengths are the state of the art. There are some technological limitations in providing a large number of wavelengths, for instance: due to channel broadening effects and non-ideal optical filtering, channels must have minimum wavelength spacing. Wavelength range, accuracy, and stability are extremely difficult to control. Therefore, it is quite possible that a given network may have more users than available wavelengths, which will necessitate the reuse of a given set of wavelengths at different points in the network. The same wavelength could be reused by any of the input ports to access a completely different output port and establish an additional connection. This technique increases the capacity of a WDM network. The recent emerge of high-bit rate IP networking application is creating the need for on demand provisioning of wavelength routing with service - differentiated offering within transport layer. To fulfill these requirements, different WDM
Page 457
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 optical network architecture has been proposed. For servicedifferentiated offering network, an accurate engineering of WDM span design is needed. So the factors such as additive nature of signal degradation, limited cascade ability of optical components and traffic dependent signal quality should be taken into account for accurate WDM span design.
probability with network dimension, network load and delay etc. The following topologies have been studied in our project:
In order to demonstrate that our approach performs better than that reported in the literature and to investigate the performance of algorithms, we must resort to simulation studies based upon the ns-2 Network Simulator. Not able to find a suitable simulator that could support our proposed DWP algorithm, we designed and developed a simulator to implement routing and wavelength assignment in all-optical networks for various topologies such as ring , mesh and interconnected rings.
1.RING TOPOLOGY
• Ring • Mesh and • Interconnected rings.
Ring networks are less connected and consequently need less messaging, which in turn generally assures faster convergence towards destination, and hence decreases computation time. We considered rings for topologies where varying network size would not significantly vary the performance (i.e., for bidirectional rings only two paths exist independently of the ring size).
WDM optical networks have gained prime importance due to the rapid growth of internet and the ever increasing demand for voice and video transmission. By allowing several channels to be routed on the same fiber on different wavelengths, the capacity of each link is increased tremendously. However this also calls for more efficient planning before provisioning lightpaths. The recent advent of high bit rate IP network applications is creating the need for on demand provisioning of wavelength routed channels with service differentiated offerings within the transport layer. To fulfill these requirements different optical transport network architectures have been proposed driven by fundamental advances in WDM technologies. The availability of ultra longreach transport and all optical switching has enabled the deployment of all optical networks. While being attractive for their transparent and cost effective operation all optical networks require accurate engineering of WDM spans to meet the requirements of dynamic wavelength routing. The additive nature of signal degradation, limited cascade ability of optical components and traffic dependent signal quality are some of the reasons that make the provisioning of on demand wavelength channels a challenging task.
1.1 Nodes Ring Topology Here, we study a 9-node bi-directional ring topology as shown in Figure1.1, of which the properties per wavelength and network element are shown. As it can be seen from the results, the method DWP, being a multipath strategy, yields superior results for the given service. This is because DWP method yields several feasible paths, out of which the best can be selected, in contrast to the non-DWP is checked on remaining requirements (transmission degradation, reliability).
To overcome the problems of analog WDM design, electronic regeneration is deployed at optical switching nodes in the so called opaque optical networks. However electronic regeneration can also impose limitations on the wavelength routing, such as delay accumulation, reliability reduction and increase in the operational cost. These issues become particularly critical if service requirements force multidimensional optimization such as maximum reliability and minimum transmission degradation. We analyze the blocking
Velammal College of Engineering and Technology, Madurai
Page 458
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 1.2 Snapshot of the simulation for 9 node ring topology
Fig. 1.2 shows the 9 node WDM network ring topology generated by the topology generator along with a snapshot of the simulation run. The bandwidth and the number of wavelengths on each multi-wavelength link are configured by the specified values. The propagation delay of each link is generated at random by the topology generator. The exponential session traffic pairs are randomly distributed amongst all nodes, which are displayed by the packet flows in nam. In the snapshot the update events of the virtual topology are captured and displayed in chronological order in the annotation panel of the main window by the event monitor. For instance, at 7.009009s, a light path is created for the traffic session 0, from node 0 to node 5 and the lightpath is established on the shortest path (path1) between source and destination without wavelength conversion.
2. MESH TOPOLOGY
the shortest path according to a certain constraint is two hops, we include only paths up to three hops (“plus one hop”). The path length limitation as used here reduces the algorithm convergence time, by avoiding messages handling information about paths with hop lengths larger than the limit. Note that the number of hops is an intrinsic variable of the adapted Bellman-Ford algorithm and is therefore advantageous to use, but additionally also a limitation on the used constraints, possibly combined by some cost-function, might be used to further reduce the number of candidate paths. In this example, we assume only nodes and links as network elements. All nodes have the same properties per wavelength along the network. Transmitters and receivers are modeled as ideal. Due to the full connectivity, compared to the 9-node ring network, this network is analyzed for higher loads. Also we assume that reliability to be constant over time. Here, two versions of DWP are used, one which uses a reduced set of feasible paths, where DWP is used for wavelength allocation only, and whiles the other considers all feasible paths, i.e. where DWP is used for routing and wavelength allocation.
We next consider the example of a nine-node mesh network, where a reduced set of pre-routed paths is introduced. To limit the number of candidate feasible paths we define a hop limit that automatically adapts to different connection requests and network loads.
2.2 Snapshot of the simulation for the 9 node mesh topology
3. INTERCONNECTED RING TOPOLOGY
2.1.Mesh Topology
Therefore we use a limit relative to the number of hops of the shortest path found, according to a single constraint. For example, if a limit is defined in terms of number of hops and if
Velammal College of Engineering and Technology, Madurai
As an example of strong practical importance, we study interconnection of networks, where, as specifically shown in Figure 3.1, six ring networks are interconnected (denoted as “6 x 5”). We believe that interconnected networks will be increasingly important to study, since they will have to deal with a variety of network technologies and architectures, originating from different vendors and service providers. To analyze interconnected networks with differentiated
Page 459
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 transmission impairments and service requirements, we consider an example where networks are designed with four different link types and the three link Parameters that are considered for routing: transmission degradation, reliability and delay.
4. PERFORMANCE ANALYSIS In order to demonstrate that our approach performs better than that reported in the literature and to investigate the performance of algorithms, we must resort to simulation studies based upon the ns-2.0 Network Simulator. Not able to find a suitable simulator that could support our proposed DWP algorithm, we designed and developed a simulator to implement routing and wavelength assignment in all-optical networks for various topologies such as ring , mesh and interconnected rings. The simulator accepts input parameters such as the number of nodes in the network, link information with weight, number of wavelengths per fiber, connection requests.
3.1 6x5-nodes network of interconnected rings Figure 3.1 shows the distribution of different link types (Type A, B, C, and D) along the example networks. . For the 6x5 network, we assumed five-node rings using links of Type A only. They are interconnected by two node disjoint links of Type B. Two link types are used for their long-haul interconnection: Type C for connections to the next closest neighbors, and Type D for connections diagonally
Some of the calls may be blocked because of the unavailability of free wavelength on links along the route from the source to the destination. The ratio of the total number rejected requests (blocked calls) to the total number of light path requests in the network is defined as the blocking probability. The output of the simulator is the blocking probability for the specified parameters along with the detailed information of connections. All these parameters can be initialized before running the simulations to obtain results for a given selection of parameters. Extensive simulations are then carried out for every combination of parameters of interest and the results are obtained .
4.1 Blocking probability Vs Number of nodes 4.1 Blocking probability Vs Number of nodes 3.2 Snapshot of the simulation for the 6x5 nodes interconnected rings topology
Velammal College of Engineering and Technology, Madurai
When we analysis the call blocking probability with size of the network (Number of nodes), the blocking probability becomes almost constant with number of nodes. But the graph
Page 460
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 shows that DWP methods are useful for mesh networks, while they bring less benefit for rings. With increasing number of nodes, the blocking probability decreases for mesh, and increases for ring networks. This is due to that DWP can use one alternative path only in bi-directional ring structures, which is N-n hops long (being number of nodes, being length of the shortest path), with which, if selected, even more network resources get allocated.
The call blocking probabilities are obtained as a function of the traffic load arriving at the network for all topologies. Here, the traffic arrival rate, traffic holding time and other conflicts are kept constant. Then the graph is plotted between the load and the blocking probability. There is no blocking probability for low load arriving at the network for all network topologies. When the arrival load to the system will be increased, leading to a remarkable increase in the blocking occur in the Performance Analysis of Constraint Based WDM Networks using DWP Algorithm network due to the offered load.
4.2 Delay due to Network size The time delay is obtained as a function of the network size for various network topologies such as ring and mesh. Here, the traffic load and other conflicts such as traffic arrival rate and traffic holding time are kept as constant. Then graph is plotted between the number of load and delay time. In the simulation, we observed that, at low network size, the delay time I s low for ring network and higher for mesh network. As the network size increases the delay time increases for ring network than the mesh network.
4.3 Blocking probability Vs Load The mesh network shows lower blocking probability for higher loads compared with other network topologies. 4.4 Blocking due to conflicts (Traffic arrival rate & Traffic holding time)
Figure 4.2 Time delay Vs Number of nodes
This is due to the fact that with larger ring networks, larger amounts of network resources are allocated and cause delay of future calls whenever a service is accommodated and also DWP can use one alternative path only in bi-directional ring structures while it can use more candidate paths in mesh structure. 4.3 Blocking due to Load
Velammal College of Engineering and Technology, Madurai
4.4 Blocking probability Vs network load in Erlang
Page 461
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 The call blocking probabilities are obtained as a function of load in Erlang for the network for all topologies. Here, the traffic load and other conflicts are kept constant. Then the graph is plotted between the load in Erlang traffic and the blocking probability. At low arrival rates, primarily the load on the network causes blocking; however, as the load in Erlang increases by increasing the traffic arrival rate while the offered load remains constant, the blocking due to load doesn’t increase by much. However, the blocking due to conflicting connection requests increases as expected and becomes the dominant source of blocking for networks with low loads. We also observe that, in the simulation, the blocking due to load increases slightly as the arrival rate increases. Although the offered load remains constant, the actual network utilization is increasing, since connections, which are blocked still reserve network resources for a short period of time, leading to a slight increase in blocking probability. Furthermore, during the connection setup process when resources are being reserved, the reserved resources will go unused for a short period of time before the connection can begin transmitting data. This resource reservation overhead will be higher when the connections are established for shorter time duration, and the number of connections being established is higher. Thus, as the arrival rate increases, the overall load in the network will tend to increase.
4.5 Delay due to Conflicts Load in Erlang
Then graph is plotted between the delay time and the load in Erlang. The load in Erlang is increased by increasing either traffic arrival rate or traffic holding time. In the simulation, we observed that, the delay time increases with load in Erlang. This is due to the fact that, larger amounts of network resources are allocated and cause delay of future calls whenever a service is accommodated.
5. CONCLUSION In this paper, we proposed a new approach to constraintbased path selection for dynamic routing and wavelength allocation in optical networks based on WDM. Our approach considered service-specific path quality attributes, such as physical layer impairments, reliability, policy and traffic conditions. Here we considered dynamic routing method. Although this method requires a long set of time, it is more efficient than other methods term of blocking probability. To validate the network modeling, we presented the detail of the revised OWns architecture, which is designed to fulfill the key characteristic of WDM network and we implemented the behavior of DWP algorithm for dynamic routing and wave length assignment problem for various networks such as ring, mesh and interconnected topologies with no conversion and we compared the behavior of blocking probability and packet delay for various network topologies. The study of blocking and delay is very useful to know about the suitable topology for various network sizes, traffic offered load and load due to conflicting factors of a network.
The time delay is obtained as a function of the load in Erlang for various network topologies such as ring and mesh. Here, the traffic load and other conflicts are kept as constant.
4.5 Blocking probability Vs Packet delay
Velammal College of Engineering and Technology, Madurai
BIBLIOGRAPHY 1) A. Jukan, and Gerald Franzl, “Path Selection Methods With Multiple Constraints in Service-Guaranteed WDM Networks” IEEE/ACM TRANSACTIONS ON NETWORKING, vol 12,pp.59-71(2004). 2) A. Jukan and H. R. van As, “Service-specific resource allocation in WDM networks with quality constraints,” IEEE J. Select/AREAS COMMUNICATION, vol 18, pp. 2051– 2061(2000). 3) J. P .Jue and Gaoxi Xiao, “Analysis of Blocking Probability for Connection Management Schemes in Optical Networks”, IEEE/ ACM TRANSACTIONS ON NETWORKING,vol 17, pp.1546-1550(2005). 4) Ashwin Sridharan and Kumar N. Sivarajan, “Blocking in All-Optical Networks”,IEEE/ACM TRANSACTIONS ON NETWORKING, vol. 12,pp.384- 396 (2004).
Page 462
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 5) Paramjeet Singh,Ajay K. Sharma, Shaveta Rani, “Routing and wavelength assignment strategies in optical networks”www.sciencedirect.com/ OPTICAL FIBER TECHNOLOGY,vol.13, pp. 191–197(2007). 6) Vinh Trong Le, Xiaohong Jiang, Yasushi Inoguchi, Son Hong Ngoa, Susumu Horiguchib, “A novel dynamic survivable routing in WDM optical networks with/without sparse wavelength conversion” www.sciencedirect.com/ OPTICAL SWITCHING AND NETWORKING, vol.3,pp. 173–190(2006). 7) C. Siva Rain Murthy and Mohan Guruswamy Chapter 4 Wavelength Rerouting Algolithms," WDM OPTICAL NETWORKS Concepts De-sign and Algorithms", PHI 2002. 8) R. Ramaswami, et al., Optical Networks: A Practical Perspective, Morgan Kaufmann Publishers Inc., 2002. Mr. P. Poothathan, Senior Lecturer, Department of Physics, Velammal College of Engg. & Tech., Madurai, completed his M.Sc. Physics from Madurai Kamaraj University. He received his M.Tech. degree in Optoelectronics & Optical communication from Rajiv Gandhi Techincal Univerisity, Bhopal. He is having a rich 14 years of teaching experience. He published many papers in various national and international conferences. His areas of interest are optical communication and Nano-computing. Ms. S. Devipriya, she received her M.Sc., and M.Phil degrees in Physics from the Pondicherry University, India, She is currently working as a Lecturer in the Department of Physics at Velammal College of Engg. & Tech., Madurai. She is having a teaching experience of 5 years and her research areas of interest are nonlinear Physics and Nanotechnology. She is an University rank holder in II position in her M.Sc. Physics. Dr. S. John Ethilton, Assistant Professor & Head, Department of Physics, Velammal College of Engg. & Tech., Madurai. He completed his M.Sc. Physics from Madurai Kamaraj University. He obtained his doctoral degree from Manonmaniam Sundaranar University, Tirunelveli. He is having totally 10 years of teaching experiences. He is having 3 international publications and 1 national publication. His research interests are Nanotechnology and fuel cells.
Velammal College of Engineering and Technology, Madurai
Page 463
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
Modeling of Cutting Parameters for Surface Roughness in Machining M. Aruna#1, P. Ramesh Kumar#2 #
Department of Mechanical Engineering, Velammal College of Engineering and Technology, Madurai, India. 1
2
[email protected] [email protected]
Abstract— In the present-day research, Response Surface Methodology is used to investigate the relationships and parametric interactions among the three controllable variables, namely cutting speed, feed rate and depth of cut, on the Surface roughness. Experiments are conducted on EN 24 steel with Ceramic inserts. To study the proposed second-order polynomial model for Surface roughness, the central composite experimental design is used to estimate the coefficients of the three factors, which are believed to influence the Surface roughness in Machining process. The response was modeled using a Response Surface Model based on experimental results. The significant coefficients are obtained by performing Analysis of Variance (ANOVA) at 5% level of significance. It was found that Cutting speed and feed rate have significant effect on the Surface roughness. This methodology is very effectual, needs only 20 experiments to levy the conditions, and model sufficiency was very satisfactory. Keywords— Response Surface Methodology (RSM), ANOVA, Central Composite Design (CCD)
I. INTRODUCTION There is a heavy demand for the advanced materials with high strength, high hardness, high temperature resistance and high strength to weight ratio, in the present-day’s technologically advanced industries like, Automobile, Aeronautics, Nuclear, Gas Turbine Industries etc. This necessity leads to evolution of advance materials like High strength alloys, Ceramics, Fibre-reinforced composites etc. In machining of these materials, conventional manufacturing processes are increasingly being replaced by more advanced techniques, and it is difficult to attain good surface finish and close tolerance. The appropriate range of feeds and cutting speeds, which provide a satisfactory tool life, is very limited [1,2]. Machinability of material provides an indication of its adaptability to be manufactured by a machining process. In general, machinability can be defined as an optimal combination of factors such as low cutting force, high material removal rate, good surface integrity, accurate and consistent workpiece geometrical characteristics; low tool feed rate and good curling of chip and breakdown of chips [3].
Velammal College of Engineering & Technology, Madurai
Therefore, it is troublesome to establish a model that can accurately predict the performance by correlating the process parameters. The optimum processing parameters are very much essential to boost up the production rate by a large extent and to shrink the machining time, since the advance materials are costlier, when compared to ordinary materials. Improving the surface quality in the advance materials is still a challenging problem. Little research has been reported about machining on EN 24 steel using Surface Response Methodology. In this work, surface response approach is used for development of a model and analysis of surface roughness, with cutting speed, feed rate and depth of cut as input parameters. A Central Composite Design (CCD) for combination of variables and response surface method (RSM) has been used to analyze the effect of the three parameters [4,5]. The ceramic cutting tools can be operated at higher cutting speeds than carbides and cermets for increased productivity and material removal rates. Experiments are conducted on hardened EN 24 steel with the hardness 40 HRC using toughened alumina cutting tool and Ti[C,N] mixed alumina ceramic tool. Flank wear models are developed and tool life analysis is done using the models [6]. Also, Experiments are conducted on S.G.Iron workpiece in a high speed lathe. The machinability of the S.G.Iron with ZTA cutting tool was evaluated in terms of the surface finish of the turned workpiece, flank wear of the cutting tool insert and cutting force during machining. The surface finish increases with cutting speed and the flank wear increases with cutting time and cutting speed [7]. The Taguchi and Shainin experimental design processes are also available for obtaining optimum cutting parameters. But still, it seems that the RSM approach would yield much better results. [8].
II. EXPERIMENTATION A number of experiments were conducted to study the effects of various machining parameters on Machining process. These studies have been undertaken to investigate the effects of cutting speed, feed rate and depth of cut on
Page 464
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 surface roughness. In this work, EN 24 steel is considered as work piece material and its composition is given inTable1. TABLE. 1 COMPOSITION OF EN 24 STEEL
C
Si
Mn
Cr
Mo
Ni
0.35-0.45 0.01-0.35 0.45-0.7 0.9-1.4 0.2-0.35 1.25-1.75
Designation: 40 Ni 2 Cr 1 Mo 28
In these methods, there is a possibility that the experiments will stop with fairly few runs and decide that the prediction model is satisfactory. Experiments have been carried out on the machining set up and the data were collected with respect to the influence of the predominant process parameters on surface roughness. 20 experiments are conducted and the value of surface roughness with design matrix is tabulated in the Table.3. TABLE. 3 PLANNING MATRIX OF THE EXPERIMENTS WITH THE OPTIMAL SURFACE ROUGHNESS
III. RESPONSE SURFACE METHODOLOGY Response surface methodology (RSM) is a collection of mathematical and statistical techniques that are useful for modelling and analysis of problems in which output or response influenced by several variables and the goal is to find the correlation between the response and the variables. It can be used for optimising the response [9,10,11]. It is an empirical modelization technique devoted to the evaluation of relations existing between a group of controlled experimental factors and the observed results of one or more selected criteria. A prior knowledge of the studied process is thus necessary to achieve a realistic model. The first step of RSM is to define the limits of the experimental domain to be explored. These limits are made as wide as possible to obtain a clear response from the model. The cutting speed, feed rate and depth of cut are the machining variables, selected in this investigation. The different levels retained for this study are depicted in Table 2. In the next step, the planning to accomplish the experiments by means of Response Surface Methodology (RSM) using a Central Composite Design (CCD) with three variables, eight cube points, four central points, six axial points and two centre points in axial, in total 20 runs are carried out. Total number of experiments conducted with the combination of machining parameters is presented in Table 2. The levels of significant are depicted in this Table. The Central Composite Design is used, since it gives a comparatively accurate prediction of all response variables related to quantities measured during experimentation. CCD offers the advantage that certain level adjustments are allowed and can be used in two-step chronological response surface methods [12].
Run
A
B
C
Ra (µm)
1
1.00000
-1.00000
1.00000
0.678
2
-1.00000
1.00000
-1.00000
0.589
3
0.00000
0.00000
0.00000
1.366
4
1.68179
0.00000
0.00000
0.510
5
0.00000
0.00000
0.00000
0.866
6
0.00000
0.00000
0.00000
1.051
7
1.00000
-1.00000
-1.00000
0.483
8
0.00000
0.00000
-1.68179
0.716
9
1.00000
1.00000
-1.00000
0.983
10
0.00000
0.00000
0.00000
0.516
11
-1.68179
0.00000
0.00000
0.561
12
0.00000
0.00000
0.00000
0.983
13
-1.00000
-1.00000
-1.00000
0.489
14
0.00000
0.00000
1.68179
0.733
15
0.00000
1.68179
0.00000
0.689
16
0.00000
-1.68179
0.00000
0.438
17
-1.00000
1.00000
1.00000
0.923
TABLE. 2 DIFFERENT VARIABLES USED IN THE EXPERIMENTATION AND THEIR LEVELS
18
1.00000
1.00000
1.00000
0.899
19
0.00000
0.00000
0.00000
0.466
20
-1.00000
-1.00000
1.00000
0.456
Variable
Cutting Speed (m/min) Feed rate (mm/rev) Depth of cut (mm)
Coding
A B C
Level 1 2 3 (-1) (0) (1) 80 110 140 0.04 0.06 0.08 0.4 0.45 0.5
Velammal College of Engineering & Technology, Madurai
The mathematical model is then developed that illustrate the relationship between the process variable and response. The behaviour of the system is explained by the following empirical second-order polynomial model.
Page 465
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010
2
ŷ = b x +b x + b x + b x + b x + b x +b x +b x .… 2
0 0
1 1
2 2
3 3
12 12
13 13
23 23
11 11
(1)
Analysis of variance (ANOVA) for the adequacy of the model is then performed in the subsequent step.
TABLE. 4 ANOVA TABLE FOR SURFACE ROUGHNESS, ESTIMATED REGRESSION
Source Regression
DF
Seq SS
Adj SS
Adj MS
F
P
9
0.49487
0.494868
0.054985
0.76
0.652
Linear
3
0.24668
0.246681
0.082227
1.14
0.378
Square
3
0.23974
0.239742
0.079914
1.11
0.390
Interaction
3
0.00844
0.008445
0.002815
0.04
0.989
Residual Error
10
0.71925
0.719246
0.071925
Lack-of-Fit
5
0.13928
0.139282
0.027856
0.24
0.928
Pure Error
5
0.57996
0.579963
0.115993
Total
19
1.21411
F value indicates that the variation of the process parameters makes a big change on the surface roughness and p denotes its TABLE. 5 COEFFICIENTS FOR SURFACE ROUGHNESS
Term
Coef
SE Coef
T
P
Constant
0.87157
0.10938
7.968
0.000
A
0.03663
0.07257
0.505
0.625
B
0.12522
0.07257
1.726
0.115
C
0.03226
0.07257
0.445
0.666
A*A
- 0.09969
0.07065
-1.411
0.189
B*B
-0.08979
0.07065
-1.271
0.233
C*C
-0.03286
0.07065
0.465
0.652
A*B
0.01925
0.09482
0.203
0.843
A*C
-0.02375
0.09482
-0.250
0.807
B*C
0.01100
0.09482
0.116
0.910
R-Sq = 40.76% R-Sq(adj) = 0.00%
The ANOVA Table for the curtailed quadratic model for surface roughness is shown in Table 4. The F ratio is calculated for 95% level of confidence. The value which are less than 0.05 are considered significant and the values greater than 0.05 are not significant. Table 4 indicates that the model is adequate since P values of lack-of-fit are not significant and F-statistics is 0.24. This implies that the model could fit and that it is adequate. From Table 5, the reduced model results indicate that the model is significant (R2 and adjusted R2 are 40.76% and 0.0%, respectively). The R2 value indicates that the machining parameters explain 40.76% of variance in surface roughness. This value indicates that the presented model fits the data well. The p-value shows that the model, linear terms and squared terms are non significant at α-level of 0.005 influence on surface roughness. The results predicted by regression model are compared with experimental measurements. Larger
Velammal College of Engineering & Technology, Madurai
percent contribution on surface roughness. Also, owing to the p-value of interaction is 0.928(>0.05), one can easily deduce that the interactions of distinct design variables are not significant. In other words, the most dominant design variables have minimum interaction with others in the current text. The final response equation for surface roughness is given as follows. Surface roughness=0.720+0.0366A+0.125B+0.0323
(C2)
The final model tested for variance analysis (F-test) indicates that the adequacy of the test is established. For analysing the data, the checking of goodness of fit of the model is very much required. The model adequacy checking includes the test for significance of the regression model, test for significance on model coefficients, and test for lack of fit. For this purpose, analysis of variance (ANOVA) is performed. The fit summary recommended that the quadratic model is statistically significant for analysis of surface roughness. This implies that the proposed model is adequate to illustrate the pattern of surface roughness. IV. RESULT AND DISCUSSION The effect of the machining parameters (cutting speed, feed rate and depth of cut) on the response variable surface roughness has been evaluated by conducting experiments. The results are put into the Minitab software for further analysis. Figure.1 shows the estimated response surface for surface roughness in relation to the process parameters of cutting speed and feed rate. It can be seen from the figure, the surface roughness tends to increase, significantly with increase in feed rate and maximum surface roughness is obtained at medium speeds.
Page 466
Proceedings of International Conference on Computers, Communication & Intelligence, July 22nd & 23rd 2010 Also in Fig. 1, in the 3D graphics, it can be seen that when the cutting speed is low and the feed rate is high the surface roughness is in its maximums. The 3D surface graphs for the surface roughness which have curvilinear profile in accordance to the quadratic model are obtained and the adequacy of the model is thus verified.
Surface Plot of SURFACE ROUGHNESS vs B, A Hold Values C 0
In the Fig. 2, between the cutting speed and feed rate, with their elliptical shape, there is a strong, positive, second degree relationship. V. CONCLUSIONS The present study develops surface roughness models for three different parameters namely cutting speed, feed rate and depth of cut for machining process of EN 24 steel using response surface method. The second-order response models have been validated with analysis of variance. It is found that all the three machining parameters and some of their interactions have significant effect on surface roughness considered in the present study. With the model equations obtained, a designer can subsequently select the best combination of design variables for achieving optimum surface roughness. This eventually will reduce the machining time and save the cutting tools. References
0.75
[1]
FA CE ROUGHNESS 0.50 0.25
1 0
0.00 -2
-1
-1
0 A
[2]
B
-2
1
[3]
[4] Fig. 1 Surface plot of surface roughness vs B, A
Fig. 2 Example of of anSURFACE unacceptable low-resolution Contour Plot ROUGHNESS vs B, A image
1.5 Fig. 3 Example of an image with
1.0
B
0.5
SURFACE ROUGHNESS < 0.2 0.2 – 0.4 acceptable resolution 0.4 – 0.6 0.6 – 0.8 > 0.8
[5]
[6] [7]
[8]
Hold Values C 0
0.0
[9]
-0.5 [10]
-1.0 [11]
-1.5 -1.5 -1.0 -0.5 0.0 A
0.5
1.0
1.5
Fig. 2 Contour plot of surface roughness vs B, A
Velammal College of Engineering & Technology, Madurai
[12] [13]
L.N. Lopez del acalle, J. Perez, J.I. Llorente and J.A. Sanchez, “Advanced cutting conditions for the milling of aeronautical alloys”, Journal of Materials Processing Technology, vol 100, No. 1-3, pp. 111, 2000. E. Brinksmeier, U. Berger and R. Jannsen, “High speed milling of Ti6Al-4V for aircraft application”, First French and German Conference on High Speed Machining, Conf. Proceeding, Metz, pp. 295-306, 1997. M.Y. Noordin, V.C. Venkatesh, S. Sharif, S. Elting and A. Abdullah, “Application of response surface methodology in describing the performance of coated carbide tools when turning AISI 1045 steel”, “Journal of Materials Processing Technology”, vol. 145, No. 1, pp. 4658, 2004. R. Snoyes, F. Van Dijck, “Investigations of EDM operations by means of thermo mathematical models”, Annals of CIRP 20 (1), pp. 35, 1971. A. Senthil Kumar, A. Raja Durai and T. Sornakumar, “Machinability of hardened steel using alumina based Ceramic cutting tools”, International Journal of refractory Metals and Hard Materials, vol. 21, pp 109-117, 2003. T. Sornakumar, R. Krishnamurthy and C.V. Gogularathnam, “Machining performance of ZTA cutting tool”, 12th ISME conference, vol. 2, 2001. A.J.Thomas, and J.Antony, “A Comparative analysis of the Taguchi and DOE techniques in an aerospace environment”, International Journal of Productivity and Performance Management, vol. 54, pp-56, 2001. D. Mandal, S.K. Pal, and P. Saha, “Modeling of electrical discharge machining process using back propagation neural network and multiobjective optimizations using non-dominating sorting genetic algorithm-II”, Journal of Materials Processing Technology, vol. 186, 154-162, 2007. K. Wang, Hirpa L. Gelgele, Yi Wang , Qingfeng Yuan and Minglung Fang.,” A hybrid intelligent method for modeling the EDM process”, International Journal of Machine Tools & Manufacture, vol. 43, pp.995–999, 2003. P. J. Wang and K.M. Tsai, “Semi-empirical model on work removal and tool wear in electrical discharge machining”, Journal of Materials Processing Technology, vol. 114, Issue 1, pp. 1-17, 2001. K. Palanikumar, “Modeling and analysis for surface roughness in machining glass fiber reinforced plastics using response surface methodology”, Materials and Design, vol. 28, pp. 2611–2618, 2007. D. C. Montgomery, “Design and Analysis of Experiments (second ed.)”, Wiley, New John Wiley and Sons, New York, 1984. L. Robert Mason, F. Richard Gunst, Dallas, Texas, James L. Hess. “Statistical Design and Analysis of Experiments with applications to Engineering and Science (Second Edition)”, A John Wiley & sons publication, 2003.
Page 467
