82 P. VARCHOL, D. LEVICKÝ, USING OF HAND GEOMERY IN BIOMERIC SECURIY SYSEMS Using of Hand Geomery in Biomeric Securiy Sysems Peer VARCHOL, Dušan LEVICKÝ Dep. of Elecronics and Mulimedia Communicaions, echnical Universiy of Košice, Park Komenského 3, 04 20 Košice, Slovak Republic Peer.Varchol@uke.sk, Dusan.Levicky@uke.sk Absrac. In his paper, biomeric securiy sysem for access conrol based on hand geomery is presened. Biomeric echnologies are becoming he foundaion of an exensive array of highly secure idenificaion and personal verificaion soluions. Experimens show ha he physical dimensions of a human hand conain informaion ha is capable o verify he ideniy of an individual. he daabase creaed for our sysem consiss of 408 hand images from 24 people of young ages and differen sex. Differen paern recogniion echniques have been esed o be used for verificaion. Achieved experimenal resuls FAR=0,82% and FRR=4,583% show he possibiliies of using his sysem in environmen wih medium securiy level wih full accepance from all users. Keywords Biomeric securiy, hand geomery recogniion, Gaussian mixure model, expecaion-maximizaion algorihm.. Inroducion Associaing an ideniy wih an individual is called personal auhenicaion. he person can be recognized by wha he knows (e.g. password, PIN, or piece of personal informaion, by wha he owns (e.g. card key, smar card, or oken like a SecurID card or by his human characerisics (biomerics. Biomeric mehods of person auhenicaion belong in modern approaches in field of access securiy. he main advanage of biomeric is ha human characerisics canno be misplaced or forgoen []. One of he mos dangerous securiy hreas is he impersonaion, in which somebody claims o be somebody else. he securiy services ha couner his hrea are idenificaion and verificaion. Idenificaion is he service where an ideniy is assigned o a specific individual, and verificaion (auhenicaion he service designed o verify a user's ideniy. Biomeric mehods can be generally divided ino wo caegories: behavioral-based mehods physiological-based mehods. Behavioral-based mehods perform he auhenicaion ask by recognizing people s behavioral paerns, such as signaures keyboard yping or voice prin. he main problem wih behavioral mehods is ha hey all have high variaions, which are difficul o cope wih. On he oher hand, while behavioral characerisics can be difficul o measure because of influences such as sress, faigue, or illness, hey are usually more accepable o users and generally cos less o implemen. Physiological-based mehods verify a person s ideniy by means of his or her physiological characerisics such as fingerprin, iris paern, palm geomery, DNA, or facial feaures. In general, rais used in he physiological caegory are more sable han mehods in he behavioral caegory because mos physiological feaures are virually nonalerable wihou severe damage o he individual [3]. 2. Hand Geomery All biomeric echniques differ according o securiy level, user accepance, cos, performance, ec. One of he physiological characerisics for recogniion is hand geomery, which is based on he fac ha each human hand is unique. Finger lengh, widh, hickness, curvaures and relaive locaion of hese feaures disinguish every human being from any oher person. Hand geomery is considered o achieve medium securiy, bu wih several advanages compared o oher echniques: medium cos as i only needs a plaform and medium resoluion reader or camera, i uses low-compuaional cos algorihm, which leads o fas resuls, low emplae size (from 352 o 209 byes, which reduces he sorage needs, very easy and aracive o users leading o grea user accepance, subconscious connecion wih police, jusice, and criminal records. he availabiliy of low cos, high speed processors and solid sae elecronics made i possible o produce hand scanners a a cos ha made hem affordable in he commercial access conrol marke. Environmenal facors such as dry weaher or individual anomalies such as dry skin do no appear o have any negaive effecs on he verificaion
RADIOENGINEERING, VOL. 6, NO. 4, DECEMBER 2007 83 accuracy of hand geomery-based sysems. he performance of hese sysems migh be influenced if people wear big rings, have swollen fingers or no fingers. Alhough hand analysis is mos accepable, i was found ha in some counries people do no like o place heir palm where oher people do. Sophisicaed bone srucure models of he auhorized users may deceive he hand sysems. Paralyzed people or people wih Parkinson's disease will no be able o use his biomeric mehod. Since here is no much open lieraure addressing he research issues underlying hand geomery auhenicaion, i is difficul o describe he sae-of-he-ar in using i in biomerics. Insead much of he available informaion is in he form of applicaion-oriened descripion. 3. Sysem Archiecure ypical archiecure of all biomeric sysems consiss of wo phases: enrollmen, recogniion. In he phase of enrollmen, several images of hand are aken from he users. he images, called emplaes, are preprocessed o ener feaure exracion, where a se of measuremen is performed. Final model depends on he mehod used for recogniion. Models for each of he users are hen sored in he daabase. In he phase of recogniion, a single picure is aken, preprocessed, and feaures are obained. In he proposed sysem, he process of verificaion is used, where he inpu emplae is compared only wih he model of claimed person. he feaure vecor is compared wih feaures from he model previously sored in he daabase. he resul is he person is eiher auhorized or no auhorized. o evaluae a biomeric sysem s accuracy he mos commonly adoped merics are he False Rejecion Rae (FRR and False Accepance Rae (FAR. FRR is he percenage of auhorized individuals rejeced by he sysem and FAR is he percenage ha unauhorized persons are acceped by he sysem []. he poin where FAR and FRR have he same value is called Equal Error Rae (ERR. he proposed sysem is dedicaed for verificaion and herefore requires he user o claim ideniy hrough an arificial ID (e.g., magneic card or PIN before he sysem can sar process of enrollmen or auhenicaion. Due o assisance of arificial IDs, verificaion sysems require considerable less compuaional resources bu he FRR may increase slighly. his is because he combined FRR for a sysem ha uses boh arificial IDs and biomeric is: FRR = FRR of ID + FRR of biomeric. ( On he oher hand, he combined FAR can be grealy reduced wih arificial ideniies: FAR = FAR of ID FAR of biomeric. (2 Requiring an arificial ID can minimize casual aacks o he biomeric verificaion sysem because random claims can ofen be rejeced as unknown o he daabase. 4. Enrollmen 4. Image Capure Enrollmen involves a process of adding users o he daabase. he image acquisiion sysem which we have designed (inspired from [4], [5] comprises of a scanner and a fla surface. A user places his righ hand on he surface of he device. he palm is facing downwards and he pegs are used as conrol poins for fixing he appropriae posiion of he hand. o obain an image, scanner is used in he nex sep (Fig.. Before obaining a new hand picure, he user was insruced o remove he whole hand from he surface. his muliple placemens allow he sysem o capure images of he hand in slighly differen posiions. ha s also he oher advanage compared o behavioralbased mehods, because enrollmen can be done in shor ime. For example, in case of voice recogniion sysem, he process of enrollmen mus be realized in a long ime period o include all possible aspecs influencing he voice. Fig.. emplae capured by scanner. he final daabase conains 26 people, where for every user 20 emplaes were capured. Because of possible incorrec placemen of hand during enrolmen, he bes picures have been chosen and 7 emplaes for each of he users lef. he 5 of hem are used for process of raining and 2 of hem for esing he sysem. Daabase consiss of people of differen sex and young ages. 4.2 Preprocessing Afer he image is capured, i is preprocessed o obain only he area informaion of he hand. he firs sep in preprocessing is is ransforming o binary image. Since here is clear disincion in inensiy beween he hand and he background, a binary image is obained hrough MA- LAB funcion im2bw. he oupu binary image has values of 0 (black for all pixels in he inpu image wih lumi-
84 P. VARCHOL, D. LEVICKÝ, USING OF HAND GEOMERY IN BIOMERIC SECURIY SYSEMS nance less han a level and (whie for all oher pixels. he level is a normalized inensiy value obained by Osu's mehod, which chooses he hreshold o minimize he inraclass variance of he black and whie pixels. Background lighning effecs and he noise make fake pixels in he image. MALAB funcion imfiler is used o remove hese pixels and o jusify edges of he hand in he nex sep. he funcion provides filering of mulidimensional images. he imfiler funcion compues he value of each oupu pixel using double-precision, floaing-poin arihmeic. Inpu image pixel values ouside he bounds of he image are assumed o equal o he neares array border value. Hand boundary is easily locaed aferwards. 4.3 Feaure Exracion Preprocessing simplifies a measuremen algorihm and enables us o ge feaures of he hand. An algorihm for feaure exracion was creaed in programming environmen MALAB and i is based on couning pixel disances in specific areas of he hand. Since he sysem uses special surface wih pegs o fix he appropriae posiion of he hand, i can obain pixel disance of he given measuremen. he algorihm looks for whie pixels beween wo given poins and compues a disance using geomerical principles. he resul is a vecor of 2 elemens (Fig. 2: Widhs: each of he fingers is measured in 3 differen heighs. hump finger is measured in 2 heighs. Heighs: he heigh of all fingers and humb is obained. Palm: 2 measuremens of palm size. 5. Recogniion he feaure vecor obained by he verificaion should ener a comparison process o deerminae if he person whose hand image was aken is he user who claims o be. his comparison is made agains user model, which will be calculaed depending on he comparison algorihm used. Experimens were made wih differen mehods: Euclidian disance, Hamming disance, and Gaussian mixure model. 5. Euclidian Disance Euclidian disance, considered he mos common echnique of all, performs is measuremens wih following equaion: d = L í = 2 ( x i i (3 where L is he dimension of he feaure vecor, x i is he i-h componen of he emplae feaure vecor, and i is he he i-h componen of he model feaure vecor. Model in his case, is hen represened as he mean of he resuling se of feaure vecors. Fig. 2. Image afer preprocessing and locaion of measuremen poins for feaure exracion. 5.2 Hamming Disance Hamming disance does no measure he difference beween he componens of he feaure vecors, bu he number of componens ha differ in value. As i is ypical ha all he componens differ beween samples of he same user, i is necessary o follow anoher approach for he emplae calculaion differen from one used for he Euclidian disance. Based on he assumpion ha he feaure componens follow normal disribuion, no only he mean of he se of iniial samples is obained, bu also a facor of sandard deviaion of he samples. In he comparison process, he number of componens of he feaure vecor falling ouside he area defined by he model parameers (represened by mean and sandard deviaion is couned, obaining he Hamming disance. 5.3 Gaussian Mixure Model In order o obain beer resuls han in previous approaches, echnique of Gaussian mixure models (GMM has been implemened for recogniion block. GMM is paern recogniion echnique ha uses an approach of he saisical mehods [6]. he vecor of each hand measuremen can be described by normal disribuion, also called Gaussian disribuion. Each hand measuremen may be hen defined by wo parameers (for our case, where measuremen vecor is one dimensional: mean (average and sandard deviaion (variabiliy. Suppose ha he measuremen vecor is he discree random variable x. For he general case, where vecor is mulidimensional, he probabiliy densiy funcion of he normal disribuion is a Gaussian funcion [2]: ( x μ 2 ( x μ p ( x, μ, = exp (4 L (2π where μ is he mean, is he covariance marix and L is he dimension of feaure vecor. Covariance marix is he naural generalizaion o higher dimensions of he concep
RADIOENGINEERING, VOL. 6, NO. 4, DECEMBER 2007 85 of he variance of a random variable. If we suppose he random variable measuremen is no characerized only wih simple Gaussian disribuion, we can hen define i wih muliple Gaussian componens. GMM is a probabiliy disribuion ha is a convex combinaion of oher Gaussian disribuions [2]: J j = ( j ( j ( j p( x, μ, p( x = π (5 where J is he number of Gaussian mixures and π (j is he weigh of each of he mixure. Afer GMM is rained, he model of each user will be he final values of π (j, μ (j, (j and J, which grealy increases he daabase size. ab. shows he differences in he size of he model depending on he compuaional mehods used. k h = n x. x k μ =, (7 ( k h = n x ( k ( μ ( x μ ( k k h = n x x k = h = n ( k k = hn ( x = ( k ( x, (8 π. (9 Afer convergency of he main model parameers, he muliple Gaussian disribuions can be described by one single funcion. In he case in Fig.3, he GMM has seven mixures and wo dimensional feaure vecor. Mehod used Raw emplae Euclidian disance Hamming disance GMM 2 mixures Model size,395 MB 352 B 520 B,209 kb ab.. Comparison of he model sizes for differen echniques. 5.3. Expecaion-Maximizaion Algorihm o esimae he densiy parameers of a GMM saisic model, cluser esimaion mehod called Expecaionmaximizaion algorihm (EM is adoped. he EM is he ideal candidae for solving parameer esimaion problems for he GMM. Each of he EM ieraions consiss of wo seps Esimaion (E and Maximizaion (M. he M-sep maximizes a likelihood funcion ha is refined in each ieraion by he E-sep. he GMM parameers can be divided ino wo groups: one conaining π (j s and anoher conaining μ (j s and (j s. he former indicaes he imporance of individual mixure densiies via he prior probabiliies π (j s, whereas he laer is commonly regarded as he kernel parameer defining he form of mixure densiy. Unlike oher opimizaion echnique in which unknown parameers can be arranged in any order, he EM approach effecively makes use of he srucural relaionship among he unknown parameers o simplify he opimizaion process. Afer iniializaion of parameers, he EM ieraion is as follows:. he E-sep deermines he bes guess of he membership funcion h (j n (x, which is he funcion for each elemen of x and each mixure []: ( j ( j ( j ( j p( x δ =, φn π n hn ( x = (6 J ( k ( k ( k = p x = k δ, φn π n where x /δ j = defines ha x is generaed by he j-h mixure, φ j is densiy funcion associaed wih he j-h mixure. 2. he M-sep maximizes funcion o find new parameers π (k*, μ (k*, (k* using (5, (6, (7. Afer ha algorihm incremen n by and repea E-sep unil convergence []. Fig. 3. GMM model - superposiion of seven Gaussian disribuions. Verical axe represens probabiliy densiy, and parameers on he horizonal axes are observaions of 2-dimensional vecor. 6. Experimenal Resuls Sysem has been esed on he daabase described in secion 4. oally wih 408 hand emplaes. Sysem behavior can be managed depending on environmen for is using and securiy policy. his is done by a hreshold, which influences boh values, FAR and FRR. he hreshold for GMM mehod is a value, which is compared o he probabiliy obained from GMM for a given user. If he probabiliy offered by GMM is higher han he hreshold, verificaion of he given user is posiive, and vice versa. Likewise, he hreshold for Euclidian disance or Hamming disance is a value, which is compared o he disance obained from he recogniion process. If he Euclidian or Hamming disance is lower han he hreshold, verificaion is posiive. he bes values of FRR, FAR and ERR achieved for differen compuaional mehods are shown in ab.2. he sysem was esed wih differen hresholds depending on used mehods and he resuls in ab.2 are he bes achieved values.
86 P. VARCHOL, D. LEVICKÝ, USING OF HAND GEOMERY IN BIOMERIC SECURIY SYSEMS FRR (% FAR (% ERR (% GMM 4,583 0,82 4,62 ED 0,47 0,272 6,45 HD 2,5 4,076 9,73 ab. 2. Resuls for differen mehods: GMM Gaussian Mixure Model, ED Euclidian disance, HM Hamming disance. As menioned above, radeoff beween FAR and FRR is adjused by a hreshold, which needs o be adjused carefully so ha he wo raes can boh saisfy he prescribed securiy sandards. If a securiy sysem makes users feel uncomforable, eiher psychologically or physically, hen he sysem is inrusive. For example, in compuer nework securiy or access conrol for areas requiring middle or low securiy levels, an inrusive sysem will annoy users and herefore will discourage hem from using i. In high securiy areas, an inrusive sysem someimes can urn ou o be a benefi, since i may appear o be a highly secure recogniion mehod. his elevaed sense of securiy may in iself discourage inruders. ab.3 and Fig. 4 show FRR and FAR values dependen on he adjusable adoped hreshold for mehod GMM. Due o a small value of he hreshold, i is given by a negaive logarihmic value. GMM (2 mixures hreshold (-log FRR (% FAR (% 2,864 6,667 0,0906 30,863 4,583 0,82 87,0676 8,333 3,0797 6,882 0, 6,432 ab. 3 Values FAR and FRR dependen on adjused securiy hreshold. Error (% 8 6 4 2 0 8 6 4 2 0 FRR[%] FAR[%] 0 20 40 60 80 00 20 reshold (-log Fig. 4. FAR and FRR agains hreshold. In order o reach an effecive comparison of differen sysems, he descripion independen of hreshold scaling is required. Receiver Operaing Characerisic (ROC in Fig.5 plos FRR values direcly agains FAR values and eliminaes hreshold parameers. FRR (% 00 0 0, 0, 0 FAR (% Fig. 5. Receiver Operaing Characerisic (ROC. 7. Conclusion Experimens presened show he possibiliies of using hand geomery as he biomeric characerisic for auomaic verificaion sysems. Hand geomery feaures used for he proposed sysem were shown as enough unique o use hem o verify he person s ideniy. From he comparison mehods, Gaussian mixure modeling has been revealed as he one wih he bes performance and became preferred o Euclidean and Hamming disance. he bes resuls achieved GMMs wih 2 mixures: FAR=0,82%, FRR=4,583% and EER=4,62%. All users showed grea accepance and easy of usage of he sysem during process of enrollmen and creaing he daabase. his sysem as designed currenly is considered a good alernaive for securiy applicaions for areas requiring middle or low securiy levels (e.g., aparmens, hospials, sores, aendance. Furher work should be applied o creae mulimodal biomeric sysem wih a fusion of hand geomery and voice prin echniques o ge securiy sysem wih high accuracy. Acknowledgemens he research described in he paper was suppored by he Minisry of Educaion and he Academy of Science of he Slovak republic VEGA under Gran No. /4054/07. References [] KUNG, S. Y., MAK, M. W., LIN, S. H. Biomeric Auhenicaion. Published as Prenice Hall Professional echnical Reference. New Jersey: Firs Prining, Sepember 2004. [2] VARCHOL, P., LEVICKY, D. Implemenaion of Gaussian mixure models for biomeric securiy sysem. In Proceedings Komunikacne a informacne echnologie, aranske Zruby(Slovak Republic, 2007.
RADIOENGINEERING, VOL. 6, NO. 4, DECEMBER 2007 87 [3] VARCHOL, P., LEVICKY, D. Access securiy based on biomeric. In Proceedings Research in elecommunicaion echnology. Nove Meso na Morave (Slovak Republic, 2006. [4] SANCHEZ-REILLO, R. Biomeric idenificaion hrough hand geomery measuremens. IEEE ransacions on Paern Analysis and Machine Inelligence. ISSN: 062-8828. Washingon, 2000. [5] JAIN, A., ROSS, A. A prooype hand geomery-based verificaion sysem. In Proceedings of 2nd In. Conference on Audio- and Videobased Biomeric Person Auhenicaion. Washingon (USA, 999. [6] YOUNG, S. he HK Book (for HK Version 3.2. Firs published December 995, Revised for HK Version 3.2 December 2002. Abou Auhors... Dušan LEVICKÝ for biography see p. 8 of his issue. Peer VARCHOL was born in Sará Ľubovňa, Slovakia, in 980. He graduaed from he echnical Universiy in Košice, Faculy of Elecrical Engineering and Informaics. Since 2004 he has been PhD. suden a he Deparmen of Elecronics and Mulimedia Communicaions, focusing on biomeric securiy, nework echnologies and digial image processing. RADIOENGINEERING REVIEWERS December 2007, Volume 6, Number 4 BAUDOIN, G., ESIEE Paris, France BARDOŇOVÁ, J., Brno Universiy of echnology BÁRÍK, H., Czech echnical Universiy in Prague BILÍK, V., Slovak Universiy of echnology, Braislava, Slovakia ĎAĎO, S., Czech echnical Universiy in Prague DJIGAN, V., ELVEES R&D Cener of Microelecronics, Moscow, Russia DOLEŽEL, I., Czech echnical Universiy in Prague DRUAROVSKÝ, M., echnical Universiy of Košice, Slovakia DŘÍNOVSKÝ, J., Brno Universiy of echnology FRÝZA,., Brno Universiy of echnology HALÁMEK, J., Academy of Sciences of he Czech Republic, Brno HEMZAL, D., Masaryk Universiy, Brno HOZMAN, J., Czech echnical Universiy, Kladno JIŘÍK, R., Brno Universiy of echnology KLÍMA, M., Czech echnical Universiy in Prague KOLÁŘ, R., Brno Universiy of echnology KOULIAKOVÁ, J., Slovak Universiy of echnology, Braislava, Slovakia KOZUMPLÍK, J., Brno Universiy of echnology KRAOCHVÍL,., Brno Universiy of echnology KRŠEK, P., Brno Universiy of echnology KULLA, P., Slovak Universiy of echnology, Braislava, Slovakia LÁČÍK, J., Brno Universiy of echnology LEVICKÝ, D., echnical Universiy of Košice, Slovakia LUKEŠ, Z., Brno Universiy of echnology MACHÁČ, J., Czech echnical Universiy in Prague MARŠÁLEK, R., Brno Universiy of echnology MIHALÍK, J., echnical Universiy of Košice, Slovakia NOVONÝ, V., Brno Universiy of echnology PÁA, P., Czech echnical Universiy in Prague PECHAČ, P., Czech echnical Universiy in Prague PERŽELA, J., Brno Universiy of echnology POLEC, J., Slovak Universiy of echnology, Braislava, Slovakia POLÍVKA, M., Czech echnical Universiy in Prague POMĚNKA, P., hene, Brno PROKEŠ, A., Brno Universiy of echnology PROVAZNÍK, I., Brno Universiy of echnology RAJMIC, P., Brno Universiy of echnology SCHEJBAL, V., Universiy of Pardubice ŠEBESA, V., Brno Universiy of echnology ŠŤASNÝ, J., Czech echnical Universiy in Prague URBANEC,., Brno Universiy of echnology VARGIC, R., Slovak Universiy of echnology, Braislava, Slovakia WIESER, V., Universiy of Žilina, Slovakia ZAVACKÝ, J., echnical Universiy of Košice, Slovakia ZEMČÍK, P., Brno Universiy of echnology ŽALUD, V., Czech echnical Universiy in Prague