Mobile Banking in Finger Print Recognition Combine

Published on 2 weeks ago | Categories: Documents | Downloads: 1 | Comments: 0 | Views: 55
of 6
Download PDF   Embed   Report

Comments

Content

 

Fusion of Gait and Fingerprint for User Authentication on Mobile Devices Mohammad Derawi∗† , Davrondzhon Gafurov∗ , Rasmus Larsen† , Christoph Busch∗ and Patrick Bours∗ ∗

Norwegian Information Security Lab., Gjøvik Univeristy College, Norway ∗ Email: firstname . lastname @hig.no  { }{ } †

Department of Informatics, Technical University of Denmark, Denmark  † Email: [email protected] 

multi-modal -modal biometric biometric authe authentic ntication ation ap Abstract—A new multi proach using proach using gait signals and finger fingerprin printt image imagess as biometric biometric traits is proposed. The individual comparison scores derived from the gait gait and fingers fingers are normal normalize ized d using using fo four ur met method hodss (min(minmax, z-score, median absolute deviation, tangent hyperbolic) and four fusion fusion approache approachess (simple (simple sum, user-wei user-weightin ghting, g, maximu maximum m scoree and minimum core). The proposed scor proposed method is eval evaluated uated using 7200 fingerprint images and gait samples. Fingerprints are collected by a capacitive line sensor, an optical sensor with total internal reflection and a touch-less optical sensor. Gait samples are obtained by using a dedicated accelerometer sensor attached to the hip. And by applying the described commercial fingerprint

recognizing a unlawful person from a security camera video until recently, when accelerometer-based gait recognition has been suggested [9][10][11]. An individuals gait is known to differ from person to person and to be fairly stable [12], whereas intentional imitation of  another anot her person’ person’ss gait is comp complica licated ted [13] [13][14]. [14]. How Howeve ever, r, the biometri biom etricc recognit recognition ion perf performa ormance nce of gait reco recogniti gnition on is not as good as fingerprint recognition since gait recognition is still in its infancy infancy [15] and research researcher er are today stil stilll improvin improving g results resu lts when using accel accelerom erometer eterss [11][10][ [11][10][? ?]. Even though

scanners and dedicated gait sensors, the fusion results of these two biometrics shows an improved performance and a large step closer for user authentication on mobile devices.

Mobile devices Mobile devices – particul particularly arly mobile phones – are being found fou nd in almost almost ev every eryone one’’s hip poc pocket ket these these days days all over over the world. The security issues related to ever-present mobile devices dev ices are beco becoming ming critical, critical, sinc sincee the stored stored informat information ion in the them m (name (names, s, addres addresses ses,, me messa ssages ges,, pic pictur tures es and future future pla plans ns sto stored red in a user user cal calend endar) ar) has a sig signifi nifican cantt person personal al va value lue.. Moreo Moreove ver, r, the servic services es whi which ch can be acc access essed ed via mobile mobi le devices devices (e.g., m-ba m-bankin nking g and m-co m-comme mmerce, rce, e-ma e-mails ils etc.) represent a major value. Therefore, the danger of a mobile device ending up in the wrong hands presents a serious threat to information security and user privacy. Statistics in the UK show that ”a mobile phone is stolen approximately every third minute” [1]. Unlike Unli ke passwords passwords,, PINs, tokens etc. biometr biometrics ics cannot cannot be sto stolen len or for forgot gotten ten.. The mai main n adv advant antage age of bio biomet metric ric authentication is that it establishes explicit link to the identity because becau se biom biometri etrics cs use huma human n phys physiolo iological gical and behavior behavioral al characteristics. Fingerpri Fing erprint nt recogniti recognition on is a broa broadly dly researched researched area with many commercial applications available [2]. Recent publications show that the performance of a baseline system deteriorates from Equal Error Rate (EER) around 0.02 % with very high quality images to EER = 25.785 % due to low qualities images [3] [4]. Thus active research is still going on to improve

it is appare apparent nt that that differ differenc ences es in wal walkin king g sty styles les of one individua dividuall are caused by shoe wear and othe otherr env environ ironment mental al factors, the impact on gait recognition can be controlled [16]. Accelero Acce leromete meter-b r-based ased gait recognition recognition can toda today y be used for detect detecting ing whe whethe therr a mobile mobile devi device ce is bei being ng car carrie ried d by one and the same subject [17], however this has not been applied for embedded accelerometer-based gait recognition in mobile devices. Instead, we see a variety of other biometric modalities that have been planned and used for this idea, such as signature [18], voice [19][20] and fingerprints, which have been employed in a commercial PDA device [21] and newer mobile phones [22]. All of these approaches except gait recognition (and voice) voice) need expl explicit icit procedure proceduress for user authe authentic ntication ation,, e.g. writing on a touch screen. And in view of the fact that more more and mor moree mob mobile ile devi devices ces at the prese present nt tim timee emb embed ed accelerometers (and few fingerprint sensors), people can walk  directly to their school, job, friends, family without perceiving gait recognition as a major threat to their privacy. On the other hand, mobile devices are often used under difficult conditions that make the users walk unstable in walking situations when  jumping, walking downhill, uphill, etc. In this paper we present a fusion of fingerprint recognition and accelerometer-based gait recognition as means of verifying the identity of the user of a mobile device. The main purpose of this paper is to study how it is possi possible ble to lo lower wer dow down n the user effort while keeping the error rates in an acceptable and practical range. However, a fusion between three single modalities in the same time (fingerprint, voice and gait) have alr alread eady y bee been n pro propos posed ed [23 [23], ], bu butt our propos proposal al is dif differ ferent ent

these numbers. Videoideo-based based gait recogniti recognition on has been studied for a long timee [5] tim [5][6] [6][7] [7][8] [8] for the use in surve surveill illanc ancee sys system tems, s, e.g. e.g.

since we are as only on con gait-recognition recogn recogniti ition on a focusing who whole. le. In contra trast st to [23 [23], ],and wefingerprintalso also have have a differ different ent set setti ting ng for both both mod modali alitie ties. s. We are tes testin ting g out

I. I NTRODUCTION

 

multiple fingerprint scanners with with multiple extractors and comparators for the fingerprint recognition where two of the scanners which are not optical, are more suitable for mobile devices. dev ices. And finally finally we are also analyzing analyzing gait gait-rec -recogni ognition tion differently. Therefore, this proposal is a realistic approach to be implemented in mobile devices for user authentication. I I . M ULTIMODAL  B IOMETRICS Multi-modal and Multi-biometric fusion is a way of combining two biometric into one single decision. wrapped biometric system to makemodalities a unified authentication During the past years of increased use of biometrics to authenticate or identify people, there has also been a similar increase in use of  multimodal fusion to overcome the limitations of single-modal biometric system. There are several benefits when combining multiple mult iple biometri biometricc systems. systems. The cohe cohesiv sivee deci decision sion leads to a sign significan ificantt impr improve ovement ment in precisio precision n and simu simultan ltaneous eously ly reduces the false acceptance rate and false rejection rate. The second benefit is that the more biometric attributes we apply the harder it is to spoof them, such that the impostor makes the verification harder to grant. The Third benefit is the reduction of noisy input data, such as a humid finger or a dipping eye-lid, since if one the input is highly noisy, then the other biometric sample sam ple migh mightt have have a very very hig high h qua qualit lity y to make make an overa overall ll reliable decision. This can also be seen as the fault-tolerance, that is, to continue operating properly in the event of the failure if one system system breaks breaks do down wn or com compro promis mised ed the then n the other other might be sufficient to keep the authentication process running. [24][25] Several of applications in the real world require a higher level of biometric performance than just one single biometric measure to improve security. These kinds of applications will replace and prevent national identity cards and security checks with fusion for example air travel, hospitals and et cetera. And for the individual who are not able to present a stable biometric characteristic to an application, then provision is needed. III. DATA C OLLECTION  A. Finge Fingerprint rprint Image Data

The fingerpri fingerprint nt dat dataa used used in this this pap paper er are captu captured red by thr three ee comme commerci rcial al sensor sensorss as sho shown wn in Figure Figure 1. Furthe Furtherr detailed detai led informa information tion of the sensors is desc describe ribed d in Table I. The experiment had 40 participated volunteers for providing fingerprints for DB1, DB2, and DB3, where 10 were female and 30 males.

Fig. 1. Left: tou touchles chlesss optical senso sensorr (TST BiRD3), BiRD3), Midd Middle: le: optic optical al sensor  R (DP U.4000), Right: capacative line sensor (IDEX SmartFinger IX 10-4 ) and a fingerprint image from each database, at the same scale factor.

Database

DB1

Sensor Name

TST

Dig. Persona

IDEX

BiRD3

U.4000

500 DPI   8-bit 19 19x1 x16 6 [m [mm] m]   5-50 [C] 160x115x95

512 DPI 8-bit 14.6 14.6x1 x18. 8.1[ 1[mm mm]] 5-35 [C] 79x49x19

 R SmartFinger IX 10-4 500 DPI 8-bit 10 10x4 x4[m [mm] m] -40-85 [C] 10x4x0.8

Model

 

Resolution   Gray Scale Acquisition   Temperature Dimension  

 

DB2

 

DB3

TABLE I S ENSOR  I NFORMATION . [C] =  CELSIUS AND  [ MM ] =  MILIMETER .

as shown in Figure 2. It is also equipped with a storage unit capable of storing 64 megabyte of acceleration data and has both a USB and a bluetooth-interface, which makes it possible to transfer the data to either a computer, a cellular phone or a PDA. The sampling frequency of the MR100 sensor was about 100 samples per second and its dynamic range was between -6g and +6g. During walking trials the MR100 was attached to the hip of the persons. Thus, we analyze hip movements for recognition purposes.

 B. Gait Data

In this experim experiment, ent, 40 subj subjects ects partici participated pated and walk walking ing weree record wer recorded. ed. The gender gender distri distribu butio tion n was the sam samee as with the fingerprin fingerprintt experime experiment. nt. Subj Subjects ects were told to walk  normally for a distance of about 20 meters in a hall on flat gr grou ound nd.. At th thee end end of the the ha hall ll the the subj subjec ects ts had to wait wait 2 seconds, turn around, wait, and then walk back. A so called Motion 100 (MR100) sensor wasinused toorthogonal record the motion. Recording The MR100 measures acceleration three directions, namely up-down, forward-backward and sideways

Fig. 2. Accel Accelerati eration on of motion reco recordin rding g in three dimensiona dimensionall axis. To Top: p: x-acceleration, Middle: y-acceleration and Botton: z-acceleration

 

V. S CORE  L EVEL  F USION

C. Mult Multi-bi i-biometr ometric ic Data

Representations ations - Assigning Gait To Fing Finger  er  The sub subjec jects ts in the fingerp fingerprin rintt and gai gaitt ex exper perim iment entss are  A. Represent different. Howev different. However, er, assuming non-correlation of persons finAs mentioned in section III, each participant acquired all of  gerprint patterns and gait (walking style) we randomly pick  his or her 10 fingers in 6 sessions, resulting in 60 templates up a gait sample and assign it to fingerprint sample. per scanner. In the gait experiment, we retrieved 12 templates for eac each h per person son.. Whe When n com combin bining ing two bio biome metri tricc agains againstt IV. IV. F EATURE  E XTRACTION AND  C OMPARISON each other, we must ensure that the template ratios from all  A. Finge Fingerprint rprint Analysis biometrics are in the same domain. By having 10 fingerprints In order to measure the sensor performance we have applied of 6 sessions are not comparable with 12 gait templates of one threee dif thre differen ferentt commerci commercial al minu minutia tia ext extract ractor or for the feature feature sessio session. n. Som Someho ehow w we mus mustt ens ensure ure that the domain domain of two extraction: were within the same range domain. We have two possi possible ble opportunities: 1) Neurotechnology, Neurotechnology, V Verifinger erifinger 6.0 Extended SDK 2) TST Biometri Biometrics, cs, SDK 2.1 1) Distr Distribu ibute/c te/copy opy the 12 templat templates es into 60 templat templates. es. 3) NIST, NIST, NIST2 SDK (min (mindtct, dtct, bozorth3) bozorth3) 2) To reduce the number of finger fingerprin prints ts (60 templates) templates) to 12 templates. All of the above mention mentioned ed SDKs includes includes funct functional ionality ity to extract a set of minutiae data from an individual fingerprint Second approach would not be a reasonable approach since a imagee and comp imag compute ute a comp comparis arison-sc on-score ore by comp comparin aring g one lot of data information is lost and performance would change set of mi minut nutiae iae dat dataa with with ano anothe therr. The im image age pro proces cessin sing g slightly. Therefore we chose the first mentioned approach. And of obtain obtaining ing the templa templates tes can be fou found nd in the each SDKs the solution that was used had the important fact and awareness documention report. What can be seen from the description to ensure that duplicate templates in the different sessions for is that NIST and TST are only designed to compare images each finger are not assigned. Thus, the solution for assigning originati origi nating ng from the same template template ext extracto ractorr only only.. Such ex- was done in the following way: tractors or comparators are identified as non-standardized (e.g.  From the gait templates, we chose 6 randomly templates •

proprietary). However, Neurotechnolgy supplier provides ISO and ANSI interoper interoperabil ability ity due to the standardi standardized zed template template formats they offer. These are therefore known as standards.

• •

 B. Gait Analysis

The fea featur turee ex extra tracti ction on for the gait-s gait-sign ignals als was don donee by applying apply ing dif differe ferent nt signal signal proc processin essing g meth methods, ods, in cont contrast rast to fingerprin finger print. t. The ext extract raction ion of feat features ures is described described in more details in [11], but roughly was the extraction performed in the following order 1)   T Time ime interpolation: Lin Linear ear tim timee int interp erpola olatio tion n on the three axis data (x,y,z) since the time intervals between two observation points are not always equal. 2)   Noise reduction: The weighted moving average filter has been applied since it is fast and implementation is easy. 3)   G-force conver conversion sion: The raw data does not contain gforce values. Therefore it must be converted by using the properties of the sensor in order to achieve values of g. 4)   Resultant Vector : The result resultant ant vecto vectorr wil willl be create created d from the converted values from all three directions. 5)   Cycle Detection: From From the result resultant ant vecto vector, r, ste steps ps are being detected meaning that cycles can be extracted. 6)   Feature Vector Creation : All cycles are being normalized to ha have ve equa equall le leng ngth th and and th thee medi median an cycl cyclee will will be the represent representati ative ve feat feature ure vector vector.. This step is slightly slightly different than performed in [11] where all cycles were maintained as a matrix for the feature vector. For the comparison part, the feature vector was compared to a reference the dynamic time between warping (DTW) since itfeature is ablevector to findusing the optimal alignment two time series.



out of 12  These templates were assigned to the first finger   To avoid duplication for when assigning all 10 fingers for one session, we just choose the next gait template in the list.   Tabl Tablee II shows shows how how the poi points nts mentio mentioned ned above above are distributed into a gait matrix. S ID   ⇓  /  F ID   ⇒

1 2 3 4 5 6

  1 (Rnd)

2

3

4

   

G3 G5

   

G4 G6

   

G5 G7

   

G6 G8

   

G11 G7

   

G12   G8  

G1 G9

  G2   G10

   

G1 G9

   

G2   G3   G4 G10   G11   G12

...            

... ... ... ... ... ...

10   G12   G2    

G8 G4

  G10   G6

TABLE II A N EXAMPLE OF RANDOMLY ASSIGNING  1 2   GAIT TEMPLATES  ( FROM ONE SUBJECT)  T O  1 0   FINGERS  . R ND  = [ RANDOM PICKED ],  S ID  = [ SESSION-ID],  F ID  = [ FINGER -ID]  A ND  G 1−12  = [ GAIT TEMPLATE FROM 1-12].

 B. Score Normalization

The com compar pariso ison n scores scores at the out output put of the ind indiv ividu idual al comparators may not be homogeneous like in our case. For example, the dynamic time warping comparator used for gait outputs a distance (dissimilarity) measure while each of the fingerprint software comparators output a proximity (similarity) measure. Thus, we simple simple calc calculate ulate the mult multipli iplicati cative ve in inve verse rse or rec recipr iproca ocall for the distan distance ce score score lik likee shown shown in Equation 1. Scoresimilarity   =

1

Scoredistance

· factor

 

(1)

 

Furthermore, the outputs of the individual comparators need not to be on the same numerical scale (range). And finally, the comparison scores at the output of the comparators may follow different statistical distributions [26]. Score Sco re normal normaliza izati tion on is the theref refore ore use used d to ma map p the scores scores of each simple-biometric into one common domain. Some of  the methods are based on the Neyman-Pearson lemma, with simplifying assumptions. Mapping scores to likelihood ratios, for ex examp ample, le, allow allowss them them to be com combin bined ed by mul multip tiplyi lying ng under an ind under indepe epende ndence nce assump assumpti tion. on. The oth other er app approa roache chess may be based on modifying other statistical measures of the comparison score distribution. Whatt is rel Wha relev evant ant to know know is tha thatt score score normal normaliza izati tion on is related rela ted very close close to score-le score-level vel fusion since it aff affects ects how scores sco res are combin combined ed and interp interpret reted ed in ter terms ms of bio biomet metric ric performance. Table able IV sho shows ws the nor normal maliza izatio tion n functi functions ons,, which which are applie app lied d in this this pap paper er.. The rel relev evant ant abbre abbrevia viatio tions ns for the statistically measures are given Table III. St Stat atis isti tica call meas measur ures es

Ge Genu nuin inee distribution

Impostor distribution

Minimum score

 

S Min

G

 

S Min

Maximum score

 

S Max

G

 

S Max

G

 

Mean

S Mean G

 

Median score

 

S Med

G Score standard standard devia deviation tion   S SD

   

 

S Min



 

S Max







S SD

B

B

 

S Med

Method

Formula 

(i=1 to N)  S i

Simple Sum Minimum Score

 

min  (i=1

to N)   S i

Maximum Score

 

max  (i=1

to N)  S i



User Wei Weighting ghting

   

B

S Mean B

S Med B

S SD

(i=1 to N)  W i∗  ·   S i

TABLE V E XAMPLES OF SCORE FUSION METHODS . [24]

Both



S Mean

score le score leve vel, l, see Figur Figuree 3. The com compar pariso ison n sco score re out output put by a comparator contains the richest information about the input biometric sample in the absence of feature-level or sensor-level informat info rmation. ion. Furtherm Furthermore, ore, it is relativ relatively ely easy to access access and combine the scores generated by several different comparators. Consequen Cons equently tly,, integrat integration ion of informati information on at the comparis comparison on score sco re le leve vell is the mos mostt common common app approa roach ch in multi multi-m -moda odall biometric systems. Table V lists the fusion approaches applied in this paper and outlined from [24].

VI . R ESULTS The results results sho shown wn below below are algorith algorithm m perf performa ormances nces for biometric verification purposes. Experiments were performed in order to compare the following configuration: 1) Perf Performa ormance nce of sing single le modalit modalities, ies, i.e. fingerprin fingerprintt recogrecognition and gait recognition separately 2) Perf Performa ormance nce of mult multi-mo i-modalit dalities, ies, i.e. fingerprin fingerprintt reco recoggnition and gait recognition fused Table VI gives gives an ove overvie rview w of the single-m single-modal odality ity perforperforma mance nces. s. In gen genera erall point point of vie view w we see that Neurot Neurotech ech--

TABLE III S YMBOLS USED FOR

SCORE NORMALIZA NORMALIZATION TION FORMULAS .

Method

Formula

Min-Max





[24]

B B B =   (S  − S Min )  /  ( S Max   -   S Min )

NIST

DB1 : TST DB2 : Digital Persona DB3 : IDEX

29.91 19.80 18.56

Neurotechnology 1.23 1.12 2.56

TST

Gait

11.08 5.82 5.50

9.39 9.61 9. 9.43

TABLE VI EE R S OF

FINGERPR FINGERPRINT INT RECOGNITION  ( COLUMN  2

- 4)  A ND  G AIT

R ECOGNITION ( LAST COLUMN )

(MM) Z-Score

Scanner

 



I  I  =   (S  − S Mean )  /  ( S SD )



B B =   (S  − S Med )   /   median(S  − S Med )



G B = 0.5 (tanh(0.01 (S-S Mean ) /  S SD )+



C. Scor Scoree Fusion

nology’ nolo gy’ss extr extracto actorr and comparato comparatorr is performi performing ng bett better er than NIST’s and TST’s for all three fingerprint databases with an EER of 1.12 %. The performances of gait recognition for all three databases using usin g dynam dynamic ic time warping lies approxima approximately tely around the same with an EER of 9.43 %. Table VII takes all of Neurotechnologys fingerprint scores (si (since nce the are per perfor formi ming ng best) best) and is fus fused ed wit with h gait gait data. data. Gi Give ven n an EER of 1.23 for fing fingerp erprin rintt and an EER of 9.39 9.39 we ga gain in an ov over eral alll fu fuse sed d pe perf rfor orma manc ncee of EER EER = 0. 0.23 23 %. However, in an greater improvement point of view, then Table VIII shows how large an improvement can be done by having

When Whe n matches indiv individu idual al bio biomet metric ric compar parato ators out output put a (comset of  possible along with thecom quality ofrseach match pariso par ison n sco score) re),, int integ egrat ration ion can be don donee at the compar compariso ison n

high Given thatwe fingerprint has an EER of  19.80numbers and gait of hasEERs. an EER of 9.61 gain an improved EER of 1.63.

Media Median n Abso Abso-lute Deviation Deviation



Hyperbolic Tangent



1) TABLE IV

A PPLIED

SCORE NORMALIZATION APPROAC APPROACHES HES .

[24]

 

Fig. 3.

Fin inge gerr 1.23 1.12 2.56

Ga Gait it 9 9..39 9.61 9.43

Fi Fing nger er + Ga Gait it 0.23 0.39 0.57

S MALLEST E ER S AFTER

Overvie Overview w of the proposed met method hod in the scorescore-lev level el fusion

No orm rmal aliz izat atio ion n MinMax MAD MAD

Fusi Fusion on Weighted Simple Sum Simple Sum

TABLE VII T HE TWO LAST COLUMNS SHOWS

FUSION .

WHICH NORMALIZATION NORMALIZATION AND FUSION APPROACHES WERE APPLIED

Fin inge gerr 29.91 19.80 18.56

M OST

Ga Gait it 9. 9.39 9. 9.61 9. 9.43

Fi Fing nger er + Ga Gait it 3.45 1.63 3.27

No orm rmal aliz izat atio ion n MAD MAD MAD

Fusi Fusion on Max Score Simple Sum Simple Sum

TABLE VIII THE TWO LAST COLUMNS SHOWS

IMPROVED  E ER S AFTER FUSION .

WHICH NORMALIZATION NORMALIZATION AND FUSION APPROACHES WERE APPLIED

VII. D ISCUSSION

in places where the probability of losing a handhold device are high, a fusion of gait processing with biometrics such as fingerpri finge rprint nt recogniti recognition on is an oppo opportuni rtunity ty to protect protect personal personal devices in noisy and normal environments. A possible application cati on scena scenario rio of a multi-mo multi-modal dal biom biometri etricc user verifica verification tion system in a mobile device could be as follows; When a device as a mobile phone,mode is first into use ittime would enter asuch ”practicing” learning fortaken an appropriate session, say 24 hours. For this period of time the system would not only form the gait and fingerprint templates, but also investigate the solidity of the behavioral biometrics with respect to the user in question. Password-based or PIN code user authentication would be used during the learning session. If the solidity of  the gait and fingerprint biometrics was sufficient enough, the system would go into a biometric authentication ”state”, a state that will need confirmation from the owner. In this state the system would asynchronously verify the owner’s identity every ti time me the ow owner ner wa walke lked d while while car carryi rying ng the phone phone dif differ ferent ent places or eventually talked into it. The system would be in a safe state for a certain period of time after verification. If  new verification failed, the system would use other means to

Since personal handhold devices at present time only offer means mea ns for exp explic licit it use userr authen authentic ticati ation, on, thi thiss authen authentic ticati ation on usually usual ly takes place one time; time; only when the mobi mobile le device has been switched on. After that the device will function for a long time without shielding user privacy. If it is lost or stolen, a lot of pri priva vate te inf inform ormati ation on suc such h as add addres resss boo book, k, pho photos tos,, financi fina ncial al data data and use userr calend calendar ar ma may y bec become ome access accessibl iblee to a stranger. Even the networking capabilities on the handhold de devi vice ce can can be us used ed with withou outt rest restra rain intt unti untill the the ho hold lder er of  the device device disco discove vers rs the loss of it and infor informs ms this this to the network provider. In turn to decrease the risks to the owner’s security secur ity and pri privac vacy y, mobi mobile le devices devices shou should ld verify verify regu regularl larly y and discreetly who in fact is carrying and using them. Gait recognition is well-suited for this purpose but is difficult under unusual and challenging conditions. In view of the fact that the risk of a handhold device being stolen is high in public

Although the use of gait biometrics alone might be insufficient for user authentication, experiments during this project has has show shown n that that it itss us usee as a comp comple leme ment ntar ary y moda modali lity ty to

area (transport, shopping areas etc), the method for unobtrusive user authenticati authentication on shoul should d work at high complica complicated ted lev levels. els. Since people frequently move about on foot (at short distances)

fingerpri finge rprint nt recogniti recognition on impr improve ovess the performa performance. nce. Even for cases where the perf cases performa ormance nce of fingerprint fingerprint was worse worse than the gait, we still saw an improvement.

verify the user, e.g. asking for fingerprint. Gaitt bio Gai biomet metric ricss is a behav behavior ioral al biomet biometric rics, s, and gait gait can be aff affecte ected d by dif differe ferent nt factors. factors. Using wear wearable able sensors in gait gait rec recogn ogniti ition on is a quite quite new new fiel field d and ther therefo efore re a lot of  further research would be needed. By looking at topics that are directly directly connec connected ted to thi thiss paper paper it is natura naturall to inc includ ludee more testing testing condi condition tions, s, like e.g. walk walking ing up- or down downhill hill,, injuries inju ries,, tiredness tiredness,, heavy heavy load carr carrying ying , high high-heel -heeled ed shoes wearing etc. but it would also be interesting to look at several types of environments like the surface, e.g. walking on grass, bad grounds, gravel, sand, etc.

 

VIII. C ONCLUSION The multi-mo multi-modal dal biom biometri etricc meth method od for frequ frequent ent authe authentintication of users of mobile devices proposed in this paper was investigated in a technology test. It contained their fingerprints and gait data with placement of the accelerometer module in the hip. Fingerpri Fing erprint-b nt-based ased recog recogniti nition on resu resulted lted in dif differe ferent nt perforperformances of using three different minutia extractors and comparators. The best functioning extractor and comparator pair was Neurotechnologys template extractor and comparator. The algorithm performance resulted in an EER of 1.12 % for DB 2 , while DB1   and DB2  resulated in EER = 1.23 % and EER = 2.56 %, respectively. Further, our experimental results show that in all cases that fused algorithm performance (finger + gait) was significantly improved compared to performances of individual modalities. Under the use of NIST extractor and comparator , where EER exceed exce ed 18 %, mult multi-mo i-modal dal authentic authenticatio ation n achieve achieved d EER of  1.63 % - 3.45 %. In cases, where fingerprint modality alone performed well enough (EER between 1.23 % - 2.56 %), the performance of the combined finger and gait modalities was further improved to EER of 0.23 % - 0.57 % . The sho shown wn result resultss sug sugges gestt the pos possib sibili ility ty of usi using ng the propos pro posed ed met method hod for mobile protec protecti ting ng person personal de es suc such h as PDAs, smart suitcases, phones etc.alIndevic a vices future of truly pervasi perv asive ve computin computing, g, when small small and inexpens inexpensiv ivee hardware hardware can be embedded in various objects, this method could also be used for prot protecti ecting ng val valuable uable personal items. More Moreove over, r, reliably reli ably authent authenticat icated ed mobi mobile le devices devices may also serve as an automated authentication in relation to other systems such as access control system or automated external system logon. I X. ACKNOWLEDGMENTS We would like to thank IDEX (www.idex.no) for using their prototype sensor for testing. And furthermore thank all of our volunteers participating in the data collection. R EFERENCES [1] Biometrics for secure mobile connections. http://www.21stcentury.co.uk/technology/biometrics-for-mobiles.asp. [Online; accessed 08-January-2010]. [2] Nist image group’s fingerprint research. http://w http://www ww.itl .itl.nist .nist.go .gov/ia v/iad/89 d/894.03 4.03/fing /fing/fing /fing.htm .html. l. [Onl [Online; ine; access accessed ed 25-February-2010]. [3] Fvc2006 Fvc2006 the four fourth th internat internationa ionall finge fingerpri rprint nt verificati verification on competiti competition. on. http://b http://bias.c ias.csr sr.uni .unibo.i bo.it/fv t/fvc200 c2006/re 6/result sults/O s/O res db2 a.asp. [Onl [Online; ine; accessed 25-February-2010]. [4] Fvc200 Fvc2004 4 th thee third third in inter ternat nation ional al fingerp fingerprin rintt ve verifi rificat cation ion com compet petiition. http://bias.csr.uni http://bias.csr.unibo.it/fvc20 bo.it/fvc2004/results.asp. 04/results.asp. [Online; accessed 2 255February-2010]. [5] S.A. Niyogi Niyogi and E.H. Adel Adelson. son. Analyzin Analyzing g and recog recognizi nizing ng walkin walking g figures in xyt.   CVPR, 94:469–474. [6] Chiraz BenAbdelk BenAbdelkader, ader, Ross Cutler, Cutler, Harsh Nanda, Harsh N, and Larry Davis. Eigengait: Motion-based recognition of people using image selfsimilarity.  Audio and Video-based Person Authentication (AVBPA). [7] Mark S. Nixon, Nixon, John N. Carter Carter,, Jamie D. Shutler, Shutler, and Michael G. New advances in automatic gait recognition.  Information Security Technical  Report , 7(4):23–35, 2002. [8] L. Wang, T T.N. .N. Tan, W.M. W.M. Hu, and H.Z. Ning. Automatic gait recognition recognition based on statistical shape analysis. 12(9):1120 12(9):1120–1131, –1131, September 2003.

[9] Jan Janii Manty Mantyjar jarvi vi Ele Elena na Vildjio ildjiouna unaite ite Satu-M Satu-Marj arjaa Mak Makela ela Heikk Heikkii J. Ailisto, Ailisto, Mikko Lindhol Lindholm. m. Iden Identify tifying ing people from gait pattern with accelerometers.   In Proceedings of the SPIE , 5779:7–14, 2005. [10] Einar Snekk Snekkenes enes Davrondzho Davrondzhon n Gafurov Gafurov and Patrick Patrick Bours. Impr Improve oved d gait recognitio recognition n perf performan ormance ce usin using g cycle match matching. ing. In   International Conferen Conf erence ce on Adva Advanced nced Info Informat rmation ion Networkin Networking g and Appl Applicat ications ions, 2010. (to appear). [11] Moha Mohammad mmad O. Derawi, Patri Patrick ck Bours, and Kjetil Holie Holien. n. Auto Automated mated non-obtr nonobtrusi usive ve gait recog recogniti nition on usin using g accele acceleromet rometers. ers.   subm submitte itted d to  IWSCN 2010. [12] L. Bianchi, D. Angelini, and F Lacquaniti Lacquaniti.. Indivi Individual dual characteristics of  human walking mechanics. 436:343–35 436:343–358, 8, 1998. [13] Davrondzho Davrondzhon netric Gafurov Gafurov, Snekkenes, and Erik Buvarp. Robustness of biom biometric gait, Einar auth authenti enticatio cation n agai against nst Tor impe imperson rsonatio ation n attac attack. k. In OTM Workshops (1), pages 479–488, 2006. [14] Bendik B. Mjaaland. Mjaaland. Gait mimi mimickin cking g - attack attack resistan resistance ce testing of  gait authentication systems. Master’s thesis, Norwegian Norwegian University of  Science and Technology, 2009. [15] Ruud Bolle, Jonathan Con Connell, nell, Sharanthchandr Sharanthchandraa Pankanti, Nalini Ratha, and Andrew Senior.   Guide to Biometrics. SpringerVerlag, 2003. [16] Ernst A. Heinz, Heinz, Kai S. Kun Kunze, ze, Stefan Sulisty Sulistyo, o, Holger Junker Junker,, Paul Lukowicz, and Gerhard Tr¨ Troster. o¨ ster. Experimental evaluation of variations in primary features used for accelerometric context recognition. In  EUSAI , pages 252–263, 2003. [17] Jonathan Lester Lester,, Blake Hannaford, and Gaetan Gaetano o Borriello. Are you with me?” – using accelerometers to determine if two devices are carried by the same person. In  Proceedings of Second International Conference on Pervasive Computing (Pervasive 2004 , pages 33–50 33–50,, 2004 2004.. [18] [18] Lua Luann nn Rra Rragam gami, i, Mauri Maurice ce Gif Giffor ford, d, and Nichol Nicholas as Edw Edward ards. s. Dsv questions remain...   Biometric Technology Today, 11(11):7, 2003. [19] Lifen Lifeng g Sang, Zhaohui Wu, Wu, and Yingch Yingchun un Ya Yang. ng. Systems, Systems, man and cybernetics.   IEEE Interna Internation tional al Confer Conference ence on Syste Systems, ms, Man and  Cybernetics, System Security and Assurance, 4:3116 – 3121, 2003. [20] Heik Heikki ki Ailisto, Ailisto, Vill Villee Haata Haataja, ja, Vesa Kyll Kyllonen, o¨ nen, and Mikko Mikko Lindholm Lindholm.. Wearable context aware terminal for maintenance personnel. In  EUSAI , pages 149–159, 2003. [21] Jean-Franc¸¸ois ois Mainguet. Biometrics for large-scale consumer products. In  IC-AI , pages 310–314, 2003. [22] Intellig Intelligent ent touch cont controls rols,, enhan enhanced ced secur security ity and touc touch h contr control ol for exceptional mobile devices. http://w http://www.at ww.atrua.com/product rua.com/products-wireless.cfm. s-wireless.cfm. [Online; accessed 25-February-2010]. [23] [23] Ele Elena na Vildji ildjioun ounait aite, e, Satu-M Satu-Marj arjaa Maakel¨ ¨ kelaa, ¨ , Mikk Mikko o Lin Lindho dholm, lm, Vesa Kyll¨onen, Kyll o¨ nen, and Heikki Aili Ailisto. sto. Incre Increasin asing g secur security ity of mobile mobile devices by decreasing decreasing user effor effortt in verificati verification. on. In   ICSNC , page page 80. IEEE Computer Society, 2007. [24]   ISO/IEC TR 24722:2007, Information Information technology - Biometrics - Multimodal and other multibiometric fusion, 2007. [25] Arun A. Ross, Karthik Nandak Nandakumar umar,, and Anil K. Jain.   Handbook of   Multibiometrics  Multibio metrics (Internatio (International nal Series on Biometrics). SpringerSpringer-V Verlag New York, Inc., Secaucus, NJ, USA, 2006. [26] A. Jain, K. Nand Nandakuma akumar, r, and A. Ross. Ross. Scor Scoree norm normaliza alization tion in multi multi-modal biometric systems. 38(12):2270–2285, 38(12):2270–2285, December 2005.

Sponsor Documents

Or use your account on DocShare.tips

Hide

Forgot your password?

Or register your new account on DocShare.tips

Hide

Lost your password? Please enter your email address. You will receive a link to create a new password.

Back to log-in

Close