Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Prof. Anil K. Jain: Keynote Lecture at ICB 2013


Published on

Slides of Prof. Anil K. Jain's Keynote Lecture at ICB 2013 - excellent review of the past, existing technology and the future of biometrics.

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Prof. Anil K. Jain: Keynote Lecture at ICB 2013

  1. 1. 50 Years of Biometric Research: Almost The Solved, The Unsolved, and The UnexploredAnil JainAnil JainMichigan State University June 5, 20131
  2. 2. 50 Years of Biometrics Research• Fingerprint– M. Trauring, “On the Automatic Comparison of Finger Ridge Patterns”,    Nature, vol. 197, pp. 938–940, 1963V i• Voice– S. Pruzansky, “Pattern‐Matching Procedure for Automatic Talker Recognition”,  J. Acoustic Society of America, vol. 35, pp. 354–358, 1963• Face• Face– W. W. Bledsoe, “Man‐Machine Facial Recognition”, Tech. Report                  PRI 22,  Panoramic Res. Inc., 1966– T. Kanade, “Picture Processing System by Computer Complex andT. Kanade,  Picture Processing System by Computer Complex and Recognition of Human Faces”, Doctoral Dissertation, Kyoto University, 1973 • Hand geometry – R.H. Ernst, “Hand ID System”, US Patent No. 3576537, 1971y• Iris – L. Flom and A. Safir, “Iris Recognition System”, US Patent 4641349 A, 1987– J. G. Daugman, “High Confidence Visual Recognition of Persons by a Test of Statistical Independence”, IEEE Trans. PAMI, vol. 15, pp. 1148–1160, 19931960s: Beginnings of research in AI, pattern recognition & image processing2
  3. 3. Over 1 billion people have been covered by biometricIdentification for Development: The Biometrics RevolutionOver 1 billion people have been covered by biometric identification programs in the Low Middle Income Countries*Identification for Development: The Biometrics Revolution, A. Gelb and J. Clark, Center for Global Development, NW, Washington DC, Working Paper 315, Jan. 2013,
  4. 4. Objectives & Outline• What is biometrics?H did bi t i t t t d?• How did biometrics get started?• Where are we now?• Where do we go from here?4
  5. 5. The Biometrics Conundrum• Demographics• Credentials• Sensor data• Transaction• Authentication• PrivilegeUserBiometric RecognitionApplication• Match score• ConfidenceBiometric RecognitionApplicationRecognize a person by body traits & link the body to an externally assigned identity 5
  6. 6. Biometric ChallengeFind a representation & similarity measure such that• Intra‐subject similarity is very high• Inter‐subject similarity is very lowInter subject similarity is very lowProbe GalleryMATCHMATCH6Similarity score > T implies a “match”
  7. 7. Why Biometrics?• Security: Does the person have a prior criminal record?• Convenience: No need to carry credentials (pw, ID)• Audit trail: Who accessed the bank vault?• Audit trail: Who accessed the bank vault?• Fraud: Is the credit card holder the rightful owner?• De‐duplication: One person, one document!Palm vein scanners used for patient registration in Houston hospital system; 2,488 patients are named Maria Garcia and 231 of them have the same birth date7‐scanners‐technology‐schools/1726175/
  8. 8. India’s Aadhaar ProjectBasic demographic data and biometricsNameParentsGender1568 3647 4958Basic demographic data and biometricsstored centrallyUID = 1568 3647 495810 fi i t 2 i i & f iDoBPoBAddress10 fingerprints, 2 irises & face imageCentral UID database“To give the poor an identity”~350 million unique ID numbers have already been issuedg p y 8
  9. 9. PersonalizationApplicationsV ifi iTransactionsHealthcareVerificationConvenienceTravelSafety9
  10. 10. Historical PerspectiveHistorical Perspective10
  11. 11. Habitual Criminal Act (1869) “What is wanted is a means ofclassifying the records of habitualcriminal, such that as soon as theparticulars of the personality of anyparticulars of the personality of anyprisoner (whether description,m e a s u r e m e n t s , m a r k s , o r, ,photographs) are received, it may bepossible to ascertain readily, andi h i h h hi i iwith certainty, whether his case is inthe register, and if so, who he is.” Habitual Criminals Returns BookG. Pavlich, “The Emergence of Habitual Criminals in 19th Century Britain: Implications for Criminology”, Journal of Theoretical and Philosophical Criminology, 2(1), pp. 1–59, 201011
  12. 12. Bertillon System (1882)H.T. F. Rhodes, Alphonse Bertillon: Father of Scientific Detection, Harrap, 195612First use of soft biometrics and multi‐biometrics
  13. 13. Friction Ridge PatternFriction Ridge Pattern“Perhaps the most beautiful andcharacteristic of all superficial marks(on human body) are the small furrows(on human body) are the small furrowswith the intervening ridges and theirpores that are disposed in a singularlycomplex yet even order on the underp ysurfaces of the hands and feet.”Francis Galton Nature June 28 1888Francis Galton, Nature, June 28, 1888
  14. 14. Three Most Popular Biometric Traits• Legacy database • Legacy database • 1: N search• 1:N search• NIST evaluation• Covert capture• NIST evaluation• High accuracy• NIST evaluation
  15. 15. Fingerprint Recognition Milestones1880Henry FauldsArticle on fingerprints published in Nature1963Mitchell TrauringFirst paper on automatic fingerprint matching2004DHSUS‐VISIT2009UIDAIIndia started issuing 12‐digit UID no. to 1901Scotland YardAdopted Galton/Henry system of classificationAD 600ChinaFingerprint to seal1893ArgentinaFirst use of fingerprint1924FBISet up fingerprintp g p g gits residents2008FBINGI1970sFBIInitiation of AFISFingerprint to seal contracts  and legal documentsFirst use of fingerprint as forensic evidenceSet up fingerprint identification division~1000‐2000 B.C. 1892 1990s 20051997NGIInitiation of AFISAncient BabylonFinger mark on clay sealJuan VucetichRolled ink on paperLive‐scan technique• Optical sensor• Solid‐state sensor• Ultrasound sensorTBSTouchless 3D sensorThomson‐CSFSwipe sensor~19952003 US‐VISIT‐VISIT_(CBP).jpg‐eikon‐touch‐fingerprint‐reader‐eikon‐touch‐700.html‐to‐start‐informing‐aadhar‐numbers‐through‐emails‐smses/article2624078.ece‐sensor/‐3‐d‐fingerprinting.html 15Siemens ID mouse
  16. 16. Face Recognition Milestones1991Turk & Pentland1964Woodrow Bledsoe1973Takeo Kanade2001Viola & Jones2006Ahonen et2009Wright et al.1997Belhumeur et1999Blanz & VetterTurk & PentlandEigenfaceWoodrow BledsoeFace recognitionTakeo KanadeFirst FR thesisViola & JonesFace detectorAhonen et al.LBPWright et al.Sparse rep.Belhumeur et al.FisherfaceBlanz & VetterMorphable face191535mm still camera1991KodakDigital camera1024p2000SharpFirst camera phone320p2010‐2013Wearable camera480P @ 30fpsGoogle GlassApril 2013Samsung Galaxy S41080p @ 30fps1990sSurveillance camera480p @ 30fpsBledsoe, W. W. 1964. The Model Method in Facial Recognition, TR PRI 15, Panoramic Research, Inc., California.‐history‐timeline/M. Turk and A. Pentland, Eigenfaces for recognition. Journal of Cognitive Neuroscience 3 (1): 71–86, 1991.Takeo Kanade, Picture Processing System by Computer Complex and Recognition of Human Faces, Kyoto Univ.,1973. Viola, Jones: Robust Real‐time Object Detection, IJCV 2001.Ahonen,et al. Face Description with Local Binary Patterns: Application to Face Recognition, PAMI, 2006.J. Wright et al. Robust Face Recognition via Sparse Representation, PAMI, 31‐2, 2009.‐video‐camera.jpg‐galaxy‐s4‐infographic/720p @30fpsBelhumeur, P.N. et al., Eigenfaces vs. Fisherfaces: recognition using class specific linear projection, PAMI, 19‐7, 1997.V. Blanz and T. Vetter, A morphable model for the synthesis of 3D faces, SIGGRAPH 1999.
  17. 17. Iris Recognition Milestones2002USA Use of iris recognition in field operations2009UIDAINational ID2010MexicoNational ID2011IndonesiaNational ID1936Frank BurchConcept of using irispatterns for humanidentification1985Flom and Safir1991John Daugman2001UAE2005Flom and Safir2011John Daugmanidentification First iris recognition patentIris recognitionpatent1995Deployed iris recognition system for border controlPatent expired Patent expired1989 20062004 2013 2013IrisScanner SystemOne of the earliest commercial iris camerasJohn DaugmanFirst iris cameraSarnoffIris on the MoveSecuriMetricsPortable iris recognition deviceAOptixApp & device for smartphoneto capture irisDeltaEye‐solutions/iom‐passport‐portal‐system‐biometric‐iphone‐scanner17
  18. 18. Other Biometric Traits18
  19. 19. Traits with Legacy Database• Trace evidence• Innocence project• Real‐time DNA matching• ~30% of latents are of palmprints• FBI’s NGI will  have latent palmprintmatching capabilityCapability to conduct 1:N search 19g p y
  20. 20. Traits in Commercial SystemsPrimarily for 1:1 match 20
  21. 21. Traits in Laboratory Stage21
  22. 22. Which Biometric Trait?22
  23. 23. Requirements of a Biometric Trait• Uniqueness (Is it distinctive across users?)• Permanence (Does it change over time?)• Permanence (Does it change over time?)• Universality (Does every user have it?)• Collectability (Can it be measured quantitatively?) • Performance (Does it meet error rate, throughput..?)Performance (Does it meet error rate, throughput..?)• User experience (Is it acceptable to the users?) • Vulnerability (Can it be easily spoofed?)• Integration (Can it be embedded in the application?)g ( pp )23No biometric trait is “optimal”, but many are “admissible”
  24. 24. Rejected Biometric Traits24
  25. 25. Where is Biometrics Now?Where is Biometrics Now? 25
  26. 26. Fingerprint Matching• Plain‐to‐plain matching– NIST FpVTE (2003): 99 4% TAR at 0 01% FARNIST FpVTE (2003): 99.4% TAR at 0.01% FAR• Latent‐to‐rolled matching– NIST ELFT EFS II (2012): 63 4% rank 1 accuracy– NIST ELFT‐EFS II (2012): 63.4% rank‐1 accuracy– On NIST SD27, state‐of‐the‐art rank‐1 accuracy: ~72%NIST FpVTE NIST ELFT‐EFSNIST FpVTE NIST ELFT EFSC. Wilson et al., Fingerprint Vendor Technology Evaluation 2003: Summary of Results and Analysis Report, NISTIR 7123, 2004M. Indovina et al., Evaluation of Latent Fingerprint Technologies: Extended Feature Sets [Evaluation #2], NISTIR 7859, 2012Matching speed: ~millions/sec for tenprint matching on a single server26
  27. 27. Face RecognitionFRVT 2012 TAR @ FAR = 0.1%Mugshots 96%Visa photos 99%‐2012.cfmVisa photos 99%Search Time (640k gallery faces): 0.3 secondsMEDS‐II TAR @ FAR = 0.1%Eigenfaces 9%Fisherfaces 35%LBP 34%COTS‐A 58%COTS‐B 88%COTS‐C 97%
  28. 28. Iris Recognition• Iris images collected from 4.3 million eyes (6.1 M images)• Over the 95 algorithms evaluated, single‐eye iris false negative identification rates (i.e. “miss” rates) are at 1.5% or higher. For two eyes, FNR = 0.7%• No. of false positives is 25 out of a total of 1013 comparisons• Pupil dilation & constriction can impact iris recognition• Template size varies between 1KB and 20KB  28
  29. 29. Speaker Recognition• Given a target speaker and a test speech segment, is the target speaking in the test segment?NIST SRE12target speaking in the test segment?– A trial consists of: Target Speaker ID + Test segmentNIST SRE12Systems tested 759095100Test Segment Duration 30‐300sNumber of TrialsTarget Speakers~2million2 897808590TAR(%)Target SpeakersKnown Non‐TargetUnknown Non‐Target2,89746,60161,87170751 2 30.01% 0.1% 1%FARC. Greenberg et al., The 2012 NIST Speaker Recognition Evaluation, NIST SRE12, December 11‐12 2012, Orlando, Florida
  30. 30. From Solved to Unsolvedtrained54% TAR @ FAR=0.1%72% Rank‐1 accuracy66.8% TAR @ FAR=10%UnconstFVC2004MBGC CASIA v4 distance NIST SD27LFW UBIRIS v2ConditionsFVC2004MBGC CASIA.v4‐distance NIST SD27LFW UBIRIS.v2Imaging C100% TAR@ FAR=0.1%99.4% TAR@ FAR=0.01%97.8% TAR@ FAR=0.01%nstrainedFRGC, Exp. 1 IREX IIIFpVTE 2003 FERET IIITD alcoholic irisUser distorted imageUsersCooperativeConUncooperative30FVC2006
  31. 31. Unsolved Problems• Fundamental problemsp• Uniqueness (individuality)• Permanence (persistence)Permanence (persistence)• Application‐driven problems• Unconstrained sensing environment (surveillance)• Unconstrained sensing environment (surveillance)• System security & user privacy • Template security• Template security• Anti‐spoofing31
  32. 32. Uniqueness• Given a 10‐digit PIN, no. of unique identities that can be resolved = 10 billion• But, what can we say about a biometric trait?• How many traits to identify 7 billion individuals?How many traits to identify 7 billion individuals?• Body trait vs. sensed imageNo. of monozygotic twin birth rate is about three in every 1,000 births worldwide 32
  33. 33. Persistence• Human body (hence biometric traits) will age over time • Can we devise an age invariant template?• Can we devise an age‐invariant template?Score=0.84                   Score=0.76                    Score=0.71                    Score=0.58COTS‐ACOTS‐B33Images from PCSO mugshot database; Courtesy Scott McCallum
  34. 34. Cameras Everywhere1M CCTV cameras in London & 4M in U.K.; average Briton is seen by 300 cameras/day;400K cameras in Beijing provide 100% coverage of public places; 150K cameras in Seoul34
  35. 35. Face Recognition in Video35How to detect “persons of interest in a video” and then identify them?
  36. 36. 36
  37. 37. Top Retrieval Ranks  for Tsarnaev Brothers(with demographic filtering)(with demographic filtering)37
  38. 38. Biometric System VulnerabilitiesInsider AttacksEnrollment FraudCollusionCoercionEnrollmentS dCoercionException AbuseFeatureExtractionMatcherEnrollmentVerificationStoredTemplatesMatcherHill Climbing,Replay &Man-in-the-MiddleAttacksSpoofing &ObfuscationAttacksTrojanHorseAttackTrojanHorseAttackTemplateAttacksExternal Adversary AttacksFakeIris AlteredFingerprints Minutiae ReconstructedFingerprint
  39. 39. Biometric Template ProtectionApplication AFeatureExtractionEnrollmentStoredTemplatesTemplateProtectionMatcherVerificationNon-linkable(Enhances PrivacyNon-invertible(Enhances System SecurityXXApplication B(Enhances Privacyby preventingcross-matching)(Enhances System Securityby preventing intrusion attacks)XStoredTemplatesCan we generate a non-invertible AND non-linkable biometric templatewithout compromising the matching accuracy?
  40. 40. Biometric Key GenerationCourtesy: Prof. Tyfun Agkul, Istanbul Technical University40
  41. 41. Circumvention
  42. 42. Face Liveness42
  43. 43. Where is Biometrics Going? g• Processor, memory & sensors, y• Ubiquitous biometricsC• Context • Privacy • Biometrics for societal goodP li ti• Personalization• Biometrics & forensics43
  44. 44. Processor, Memory and Sensing TechnologyMicroprocessor performance vs. cost RAM capacity vs. cost1892 Ink and paper1990sOptical sensor1990sCapacitive sensor1997 First swipe sensor2007 US‐VISITSlap sensor2010 SAFRANTouchless swipe sensor
  45. 45. Ubiquitous BiometricsqBiometrics will become more holistic,h l b h dwhere location, behavior and recentinteraction history fuse with multimodalbiometric ID (strong and weak traits) toprovide a strong assurance of identity.This degree of integration is bothinevitable and necessary for ubiquitousy qbiometrics.Courtesy Rob Rowe, Lumidigm 45
  46. 46. 46
  47. 47. Google Glasses47
  48. 48. Privacy “Cafes ban Googleglasses to protectcustomers privacy:customers privacy:Fears users of futuristiceyewear can recordwithout permission.”‐2323578/Cafes‐ban‐Google‐glasses‐protect‐customers‐privacy‐Fears‐users‐futuristic‐eyewear‐record‐permission.html#ixzz2UgEfuN8F48
  49. 49. Your Digital Footprint Defines YouBrowsing HistoryBookmarksPreferencesYour FavoriteWeb ServicesYour Family& Friendse-TransactionsOnline Interaction Patterns++hWhatyou do?Home & Office MobileWhereyou are?Whatyou are?WHO YOU ARE
  50. 50. GPS Fingerprint:                         Identification Without Biometric DataIdentification Without Biometric DataWith just the anonymous location data, one can figure out “who you are” by tracking your smartphone. Patterns of our movements, when traced on a map, De Montjoye, Hidalgo, Verleysen & Blondel, “Unique in the Crowd: The Privacy Bounds of Human Mobility”, Scientific Reports, vol. 3, 2013create something akin to a fingerprint, unique to every person.50
  51. 51. Affective BiometricsCourtesy Prof. Tayfun Akgul, ITU
  52. 52. Cell Phone Science(By Bill Gates, Nov 09, 2010)“Most of us think of cell phones primarily asa convenient tool to stay in touch withpeople and store information. Butincreasingly, scientists are exploring ways touse cell phones to deliver critical health careto people in developing countries.”Over 1.7 billion mobile phones were sold worldwide in 2012 alone‐Phone‐Science 52
  53. 53. Mobile Phone‐based Vaccination Registry Use fingerprint scans to track children who have received immunizations The goal is to reduce16 months old boy, right ring fingerUse fingerprint scans to track children who have received immunizations. The goal is to reduce redundant doses and increase coverage levels in developing countries (Mark Thomas, VaxTrak)53
  54. 54. Locard’s Exchange PrincipleEdmond Locard (1877–1966), a pioneer in forensic science,stated that the perpetrator of a crime will bring something intothe crime scene and leave with something from it, and that bothcan be used as forensic evidence.“….. Not only his fingerprints or his footprints, but his hair, thefibers from his clothes, the glass he breaks, the tool mark hel th i t h t h th bl d h d itleaves, the paint he scratches, the blood or semen he depositsor collects. All of these and more, bear mute witness againsthim. This is evidence that does not forget. …..Only humanf il t fi d it t d d d t d it di i i h itfailure to find it, study and understand it, can diminish itsvalue.“Paul Kirk, Crime investigation: physical evidence and the police laboratory. Interscience Publishers, NY 1953 54
  55. 55. A Brief History of DNA TestingBefore DNA: Blood TypingPCR, STRs and the Onset of Modern DNA TestingMitochondrial Based TestsRapidDNATests results in    < 2 hoursMatches actual samples rather than templates Blood Typing19161986198819921996of Modern DNA Testing Based Tests DNA2010s19531986 1992RFLP: The first DNA FingerprintingY Chromosome Based TestsNon‐Human DNA ForensicsTouchDNADouble helix structure of DNA first discoveredBlood Typing Y Chromosome Tests Touch DNADouble helix structure of DNA‐dna‐crime‐scene‐crime‐laboratory‐center/resources/dna/55
  56. 56. Summary & Conclusions• Biometrics Recognition is here to stayh h l• Research vs. Technology • Drivers for academic researchers: –Error rates (but what’s the baseline?), new biometric traits, fusion,…biometric traits, fusion,…• Drivers for technology providers:i ( l i–Requirements (error rates, template size, processor, throughput), seamless integration usability return on investmentintegration, usability, return on investment 56
  57. 57. My Observations1. Biometric System: Almost always embedded in an application 2 Bi t i T it N ti l b t b tt th th2. Biometric Trait: No optimal one, but some are better than others3. Matcher Accuracy: Zero error is neither required nor guaranteed4 System Evaluation: Error rates in lab test are lower than field test4. System Evaluation: Error rates in lab test are lower than field test5. Baseline: Improper baseline provides false sense of progress6 Security: Biometrics is an effective tool only if implemented well6. Security: Biometrics is an effective tool only if implemented well7. Biometric Template: Feature extraction is not a one‐way function8 Fusion: Does not guarantee better performance and security8. Fusion: Does not guarantee better performance and security9. Match Score: Gaussian density is not advisable. Tails are critical10 Impact: Not without a perspective on application & technology10. Impact: Not without a perspective on application & technology57
  58. 58. Acknowledgement58