Performance Testing Guide to Biometrics chapter 7 An

Performance Testing “Guide to Biometrics” - chapter 7 “An Introduction to Evaluating Biometric Systems” by Phillips et al. , IEEE Computer, February 2000, pp 56 -63 Presented By: Xavier Palathingal September 21 st, 2005

Background • Two biometric capabilities( matching and ranking) and biometric system errors • Chapter 5 – 1: 1 Biometric Matching • Chapter 6 – 1: m Biometric Searching • Relate “error quotes” to error definitions • Look at accuracy numbers and reconstruct and interpret them

Overview • Measuring performance • Technology Evaluations • Scenario Evaluations • Operational Evaluations • Comparison of the methods • Limits to Evaluations

Overview (cont…) • Implications of error rates • Biometric Authentication - “Why does it reject me? ” • Biometric Screening – “Why does it point to me? ” • • Face, Finger and Voice Iris and Hand geometry Signature Summary of verification accuracies

Overview (cont…) • Identification System Testing • Biometric data and ROC and CMC • Biometric search engines • 1: m search engine testing • Face Recognition and Verification Test 2000 [ FVRT 2000 ] • FVRT 2002

Measuring Performance Evaluation protocols • Why measure performance ? • Determines how you test the system, select the data and measure performance • Evaluation shouldn’t be too hard or too easy • Is just right when it spreads performance over a range that lets to distinguish

Measuring Performance 1 -Technology Evaluations • On laboratory or prototype algorithms • “testing on databases” • Move from general to specific • “training” data A • Sequestered “test” data Q • Two phases • Training phase • Competitive testing phase

Phase 1 of Technology Evaluation. Training phase • The algorithm is trained using “training” data A = (A 1 U A 2 U …) • Then tested on newly made available sequestered “test” data Q

Phase 2 of Technology Evaluation Competitive testing phase • Using database D for each matcher Z, a set of match (genuine) scores X={X 1, X 2, …, XN} and a set of non-match scores Y={Y 1, Y 2, …. , YN} are generated. • FMR and FNMR [FAR and FRR] are calculated as a function of threshold T

Measuring Performance 2 - Scenario Evaluations • Tests complete biometric systems under conditions that model real world applications • Combination of sensors and algorithms • “office environment”, “user tests”

Measuring Performance 3 - Operational Evaluations • Similar to scenario evaluations • Scenario test – class of applications • Operational test – specific algorithm for a specific application • Performed at the actual site • Using actual subjects/areas • Usually not very repeatable

Comparison of methods • Academia tend to use databases i. e. ; technology evaluations • acquisition procedures • user population is closed in scenario evaluations • Not “double blind” – technology and scenario

Limits to evaluation • Biometric authentication should be mandatory to the whole user population • User population should be fairly represented • Subjects should be unaware of the matching decision • Only realistic form of testing is operational evaluation • One cannot measure the true FAR or true FNR – nobody except the actual subject knows • Attempt to measure these “hidden” system parameters will be by trying to defeat the biometric system

Implication of error rates Biometric Authentication “Why does it reject me? ” • Verification protocol – frequent flyer smartcard with biometric - fingerprint template on a smartcard - unique frequent flyer no. and smartcard - FRR = 3% (typical for finger) - 5000 people per hr [Newark airport] in a 14 hr day. 03 x 5000 x 14 = 2100 - will have to handled through exception handling procedures

Implication of error rates Biometric Screening “Why does it point to me? ” • Screening protocol – passenger face images with government face image database - a system that checks a face against a negative database N of n=25 alleged terrorists - FPR = 0. 1% - 300 people request access to a flight 25 x 300 = 7500 matches 7500 x. 001 = 7 false positives

Implication of error rates Biometric Screening “Why does it point to me? ” • The no. of false positives , FPR(n) ≈ n x FPR(1) • Matching a positive data set M of m subjects requires m matches against a database N of n terrorists • m = 300 • n = 25 • # false positives for plane = m x FPR(n) = m x n x FPR(1)

Face , Finger and Voice • Technology evaluations • FARs are operating around 10%

Iris • “normal office environment”, with 200 volunteers over a period of 3 months • In identification mode, not in verification mode • High FRR may be due to environmental error, reflection from glasses, user difficulty

Hand Geometry • Group of 50 users. • 200 volunteers over a 3 month period

Signature • Does not have the characteristic of permanence • Accept = genuine, reject = forgery • Zero-effort forgery, Home-improved forgery, Over-the-shoulder forgery, Professional forgery

Signature (cont…. ) • Improvement of two-try over one-try indicates poor habituation of the biometric on that particular device

Summary of verification accuracies • Best error rates found in literature • One main thing is the volume

Identification system testing: Biometric data and ROC, CMC • Biometric capabilities like ranking and matching need to be developed by modeling biometric data and training using biometric data • Two different biometric statistics – ROC and CMC • ROC – measures the capabilities of a match engine s(B’, B) with some fixed t 0 or as a function of some operating threshold T • CMC – measures the capabilities of a rank engine R((B 1, B 2), B’l) with ordered entries (B 1, B 2) € M and some unknown sample B’l

Biometric search engines • A hybrid approach - ranking followed scoring • Input to the 1: m search engine - B’l , the biometric sample • Output - vector CK(B’l) =(ID(1), …ID(K))T • The 1: m search engine with an enrollment database of M is defined as : CK = (B(1), B(2), …. , B(K))T = (ID(1), ID(2), …, ID(K))T

Biometric search engines (cont…) • A possible architecture: - A biometric rank engine which determine some reordering Cm of vector M by repeatedly applying ranking - A biometric match engine determine using a scoring function s(B’l, B(k)) and decision threshold t 0(B’l), a short candidate vector CK of the K top candidates

1: m search engine testing • The big distinction of a 1: m search engine compared to a 1: 1 matcher - prerequisite of an enrollment database M = (B 1, B 2, …. BM)T • We select the first m samples as database samples [9] • For other samples, denoted as {B’l, l=m+1}= DM, a rank ř(B’l) is estimated as follows: 1. Computes the sets of scores Xl = {s(B’l, Bi); i = 1, …. , m} for l = m + 1

1: m search engine testing (cont. . ) 2. Sort these scores: X~l = (s(B’l, B(1)), s(B’l, B(2)), …. s(B’l, B(m)))T such that s(B’l, B(k)) > s(B’l, B(k+1)), 1 ≤ k < m 3. If (B’l, B(k)) is the mated pair, i. e. , if Bi = B(k) matches B’l, ř(B’l) = k

Face Recognition and Verification Test 2000 • • • 1. 2. 3. 4. 5. First attempt to characterize performance measures 5 participating vendors had to compute an all-against-all match of a database of 13, 872 face images Some results: Compression does not affect performance adversely Pose changes up to 25 degrees was handled by algorithms, beyond 40 the performance degrades sharply Images taken 12 or more months apart are difficult to recognize Distance between camera and person matters a lot Identification is more sensitive to expression changes than verification is

FRVT 2002 • An increase in database size • Difference in results in plain verification tasks • K =sorted list size, m =gallery size –

Thank you ! Especially to: Dr. Bebis for suggesting the additional paper Reza and Chang for help with the scanner