An Introduction to Biometric Recognition Anil K. Jain, Fellow, IEEE, Arun Ross, Member, IEEE, and Salil Prabhakar, Member, IEEE, IEEE Transactions on Circuits and Systems for Video Technologies, vol. 14, no. 1, Jan. 2004 Multimedia Security
Outline (1/2) Part Ⅰ. Introduction Part Ⅱ. Biometric System Part Ⅲ. Biometric System Errors Part Ⅳ. Comparison of Various Biometrics Part Ⅴ. Application of Biometric Systems 2
Outline (2/2) Part Ⅵ. Advantage and Disadvantage of Biometrics Part Ⅶ. Limitation of (Unimodal) Biometric Systems Part Ⅷ. Multimodal Biometric Systems Part Ⅸ. Social Acceptance and Privacy Issues 3
Ⅰ. Introduction (1/5) The term biometric comes from the Greek words bios (life) and metrikos (measure). Biometrics individuals physiological and/or behavioral characteristics. 4
Ⅰ. Introduction (2/5) Biometric Recognition who she is vs. what she possesses 5
Ⅰ. Introduction (3/5) What biological measurements qualify to be a biometric? a) Universality b) Distinctiveness c) Permanence d) Collectability 6
Ⅰ. Introduction (4/5) In a practical biometric system, there are a number of other issues that should be considered a) Performance b) Acceptability c) Circumvention 7
Ⅰ. Introduction (5/5) In conclusion, the system should meet a) Accuracy b) Speed c) Resource requirement d) Be harmless to the users e) Be accepted by the intended population f) Be sufficient robust to various attack 8
Ⅱ. Biometric System (1/10) A biometric system is essentially a pattern recognition system. 9
Ⅱ. Biometric System (2/10) A biometric system is designed using the following four main modules. 1) Sensor module (encapsulating a quality checking module) 2) Feature module 3) Matcher module (encapsulating a decision making module) 4) System database module 10
Ⅱ. Biometric System (3/10) A sample flow chart: template Sensor Matcher Qualify checker Feature Extractor Decision Maker System Database The templates in the system database may be updated over time. True / False 11
Ⅱ. Biometric System (4/10) A biometric system may operate either in verification mode or identification mode. a) Verification mode: Does this biometric data belong to Bob? b) Identification mode: Whose biometric data is this? 12
Ⅱ. Biometric System (5/10) Template Login Interface Quality Checker Feature Extractor System Database Get Name & Snapshot Check Quality Enrollment 13
Ⅱ. Biometric System (6/10) Claimed identity Login Interface Get Name & Snapshot Feature Extractor Extract Features Matcher One match One template System Database Verification True / False 14
Ⅱ. Biometric System (7/10) Login Interface Get Name & Snapshot Feature Extractor Extract Features Matcher N match N templates System Database Identification User s identity or user unidentified 15
Ⅱ. Biometric System (8/10) Recognition is the generic term of verification and identification. We do not make a distinction between verification and identification. 16
Ⅱ. Biometric System (9/10) Describing the verification problem: a) An input feature vector: X Q b) A claimed identity: I c) The biometric template corresponding to I : X I d) The similarity between X Q and X I : S(X Q, X I ) e) The predefined threshold of similarity: t f) True (a genuine user): ω 1 ; False (an imposter): ω 2 ( I, X Q ω1, ) ω2, if S( X Q, X I otherwise ) t 17
Ⅱ. Biometric System (10/10) The identification problem a) The identity enrolled in the system: I k, k=1, 2,, N b) The reject case: I N+1 c) The biometric template corresponding to I k : X Ik X Q I I k, N 1, if max{s( X Q, X Ik )} t, otherwise k 1, 2,, N 18
Ⅲ. Biometric System Errors (1/9) A biometric verification system makes two types of errors: 1) mistaking biometric measurements from two different persons to be from the same person (called false match) 2) mistaking two biometric measurements from the same person to be from two different persons (called false non-match) 19
Ⅲ. Biometric System Errors (2/9) Hypothesis testing: 1) H 0 : input X Q does not come from the same person as the template X I 2) H 1 : input X Q comes from the same person as the template X I 20
Ⅲ. Biometric System Errors (3/9) Decision: D 0 : person is not who she claims to be D 1 : person is who she claims to be. If S (X Q, X I ) t, then decide D 1, else decide D 0. 21
Ⅲ. Biometric System Errors (4/9) Such a hypothesis testing formulation contains two type of error: Type Ⅰ(α): false match (D 1, when H 0 ) Type Ⅱ(β): false non-match (D 0, when H 1 ) FMR is the probability of Type I error FNMR is the probability of Type II error 22
Probability (p ) Ⅲ. Biometric System Errors (5/9) Imposter Distribution p (s H 0 ) Decision Threshold (t ) Genuine Distribution p (s H 1 ) FNMR = P (D 0 H 1 ) FMR = P (D 1 H 0 ) - Matching Score (s ) 23
Ⅲ. Biometric System Errors (6/9) The errors in identification mode: FMR N : the identification false match rate FNMR N : the identification false non-match rate FMR N = 1 (1 FMR) N ~ N FMR FNMR N ~ FNMR 24
Ⅲ. Biometric System Errors (7/9) Some situation may lead to following formulation of FMR N and FNMR N. a) FNMR N = RER + (1 - RER) FNMR RER: retrieval error rate b) FMR N = 1 (1 FMR) N P P: the average percentage of database searched during the identification 25
False Match Rate (FMR) Ⅲ. Biometric System Errors (8/9) Forensic Applications Civilian Applications High-security Applications False Non-match Rate (FNMR) 26
Ⅲ. Biometric System Errors (9/9) Important specifications in a biometric system: 1) FMR: false match rate 2) FNMR: false non-match rate 3) FTC: failure to capture (e.g., a faint fingerprint) 4) FTE: failure to enroll 27
Ⅳ. Comparison of Various Biometrics (1/10) Each biometric has its strengths and weaknesses. No biometric is optimal. A brief introduction of the commonly used biometrics is given below 28
DNA Ears Ⅳ. Comparison of Various Biometrics (2/10) 1-D ultimate unique code identical twins have identical DNA patterns contamination and sensitivity automatic real-time recognition issues privacy issues The shape of the ear the structure of the cartilaginous tissue of the pinna 29
Ⅳ. Comparison of Various Biometrics (3/10) Face - Also used by humans 1) the location and shape of facial attributes 2) the overall analysis of the face image Requiring a simple background and illumination In practice, Detect the face Locate the face Recognize the face 30
Ⅳ. Comparison of Various Biometrics (4/10) Facial, hand, and hand vein infrared thermogram A thermogram-based system does not require contact and is non-invasive Infrared sensors are prohibitively expensive 手掌靜脈辨識系統資料來源 :FUJITSU, Taiwan 31
Fingerprint Ⅳ. Comparison of Various Biometrics (5/10) A fingerprint scanner costs about US $20 Single vs. Multiple 32
Ⅳ. Comparison of Various Biometrics (6/10) Gait Hand and finger Geometry 33
Iris Ⅳ. Comparison of Various Biometrics (7/10) stabilize during the first two years of life the irises of identical twins are different extremely difficult to surgically tamper the texture of the iris 34
Keystroke Odor Palmprint Ⅳ. Comparison of Various Biometrics (8/10) 35
Ⅳ. Comparison of Various Biometrics (9/10) Retinal scan the most secure biometric reveal some medical conditions Signature professional forgers may be able to reproduce signatures that fool the system Voice a combination of physiological and behavioral biometrics 36
Ⅳ. Comparison of Various Biometrics (10/10) 37
Ⅴ. Application of Biometric Systems (1/3) The application of biometric can be divided into three main groups: 1) Commercial ATM, credit card, cellular phone, distance learning, etc. 2) Government ID card, driver s license, social security, passport control, etc. 3) Forensic terrorist identification, missing children, etc. 38
Ⅴ. Application of Biometric Systems (2/3) 39
REVENUE (US$m) Ⅴ. Application of Biometric Systems (3/3) 2000 1905.4 1800 1600 1400 1440.6 1200 1000 1049.6 800 729.1 600 400 250.9 399.5 523.9 200 0 1999 2000 2001 2002 2003 2004 2005 SOURCE: The `123' of Biometric Technology 40
Advantage Ⅵ. Advantage and Disadvantage of Biometrics (1/2) All the users of the system have equal security level. Between 20% and 50% of all help desk calls are for password resets. 41
Ⅵ. Advantage and Disadvantage of Biometrics (2/2) Disadvantage Speed is perceived as the biggest problem. FMR will increase when scaling up an identification application. 42
Ⅶ. Limitation of (Unimodal) Biometric Systems (1/2) 1) Noise in sensed data 2) Intra-class variations 43
Ⅶ. Limitation of (Unimodal) Biometric Systems (2/2) 3) Distinctiveness e.g. Hand geometry, face, etc. 4) Non-universality 5) Spoof attacks 44
Ⅷ. Multimodal Biometric Systems (1/19) Data Fusion Level of Fusion a) Fusion at Sensor level b) Fusion at Feature level c) Fusion at Opinion level d) Fusion at Decision level 45
Ⅷ. Multimodal Biometric Systems (2/19) Biometric snapshot Feature Extraction features System Database Fusion features Matching Decision Making Biometric snapshot Feature Extraction decision 46
Ⅷ. Multimodal Biometric Systems (3/19) This combination strategy is usually done by a concatenation of the feature vectors extracted by each feature extractors. This yields an extended size vector set. 47
Ⅷ. Multimodal Biometric Systems (4/19) Two drawbacks: 1) There is little control over the contribution of each vector component on the result. 2) Both feature extractors should provide identical vector rates. 48
Ⅷ. Multimodal Biometric Systems (5/19) Although it is a common belief that the earlier the combination is done, the better result is achieved, state-of-the-art data fusion relies mainly on the opinion and decision level. 49
Ⅷ. Multimodal Biometric Systems (6/19) Biometric snapshot Feature Extraction Matching rank values System Database Fusion Decision Making rank values Biometric snapshot Feature Extraction Matching decision 50
Ⅷ. Multimodal Biometric Systems (7/19) The score must be adjusted first: ( Normalization must be done. ) The similarity measures must be converted into distance measures. The score generated by each classifier must have same range. [ex. 0-100] 51
Ⅷ. Multimodal Biometric Systems (8/19) The combination strategies can be classified into three main groups: Fixed rules / equal weight Trained rules / unequal weight Adaptive rules / adaptive weight 52
Ⅷ. Multimodal Biometric Systems (9/19) The most popular schemes are: Weight sum Weight product Decision trees ( base on if-then-else ) 53
Ⅷ. Multimodal Biometric Systems (10/19) Classifier 1 Classifier 2 Classifier 3 Yes Score 1 > t 1 No Yes Score 2 > t 2 No Yes Score 2 > t 2 No False Yes No Score 3 > t 3 False True False True 54
Ⅷ. Multimodal Biometric Systems (11/19) Biometric snapshot Feature Extraction Matching Decision Making decision System Database Fusion decision decision Biometric snapshot Feature Extraction Matching Decision Making 55
Ⅷ. Multimodal Biometric Systems (12/19) In this last case, the Borda count method can be used for combining the classifiers outputs. This approach overcomes the scores normalization that was mandatory for the opinion fusion level. 56
Ⅷ. Multimodal Biometric Systems (13/19) Classifier 1 class 2 class 1 class 3 class 2=2 class 1=1 class 3=0 Classifier 2 class 1 class 2 class 3 class 1=2 class 2=1 class 3=0 class 2=5 class 1=3 class 3=1 Classifier 3 class 2 class 3 class 1 class 2=2 class 3=1 class 1=0 57
Ⅷ. Multimodal Biometric Systems (14/19) One problem that appears with decision level fusion is the possibility of ties. For verification applications, at least three classifiers are needed. 58
Ⅷ. Multimodal Biometric Systems (15/19) An important combination scheme at the decision level is the serial and parallel combination, also known as AND and OR combination. System 1 System 2 System 1 System 2 59
Ⅷ. Multimodal Biometric Systems (16/19) The AND combination improves the False Acceptance Ratio. The OR combination improves the False Rejection Ratio. 60
Ⅷ. Multimodal Biometric Systems (17/19) Multiple sensors optical & capacitance sensors minutiae & non-minutiae based matchers Multiple matchers Multiple biometrics face & fingerprint Multimodal Biometrics two attempts of right index finger Multiple snapshots Multiple units right index & middle fingers 61
Ⅷ. Multimodal Biometric Systems (18/19) Example of Multimodal Biometric Systems Person Identification Using Multiple Cues Face, Voice Expert Conciliation for Multimodal Person Authentication Systems using Bayesian Statistics Face, Speech Integrating Faces and Fingerprints for Personal Identification Face, fingerprint Personal Verification using Palmprint and Hand Geometry Biometric Palmprint and Hand Geometry Bioid: A Multimodal Biometric Identification System voice, lip motion, face 62
Ⅷ. Multimodal Biometric Systems (19/19) A combination of uncorrelated modalities is expected to result in a better improvement in performance. A combination of uncorrelated modalities can significantly reduce the FTE. However, the cost of the system may increase and the system may cause inconvenience. 63
Ⅸ. Social Acceptance and Privacy Issues (1/3) Social Acceptance The ease and comfort in interaction with a biometric system contribute to its acceptance. Biometric characteristics captured without the knowledge of the user is perceived as a threat to privacy by many individuals. 64
Ⅸ. Social Acceptance and Privacy Issues (2/3) Privacy Issues Biometrics can be used as one of the most effective means for protecting individual privacy. Biometric characteristics may provide additional information about the background of an individual. 65
Ⅸ. Social Acceptance and Privacy Issues (3/3) Legislation is necessary to ensure that such information remains private and that its misuse is appropriately punished. Most of the commercial biometric systems available today store a template in an encrypted format. 66