Theory, Characterization and Testing of General Biometric Technologies (Advanced Course) Jim Wayman Office of Research and Graduate Studies and College of Engineering San Jose State University 1 For Further Details See “National Biometric Test Center Collected Works 1997-2000” www.engr.sjsu.edu/biometrics/nbtccw.pdf “Biometric Standards Development Final Report” (1997) www.engr.sjsu.edu/biometrics/fhwa.htm 2 A SHORT HISTORY • Bertillion - 1880 • Fingerprinting -1880 – Faulds – Herschel – Galton – Twain 3 A DIGRESSION: GALTON ON BERTILLION “There was...a want of fulness in the published accounts of it, while the principle upon which extraordindary large statistical claims to its quasi-certainty had been founded were manifestly incorrect, so further information was desirable. The incorrectness lay in treating the measures of different dimensions of the same person as if they were independent variables, which they are not.... The chances against mistake have been overrated enormously owing to this error; still, the system was most ingenious and very interesting”. Galton, Memories of My Life (1908), p. 251.4 DÉJÀ VU IN THE 21st CENTURY DNA -- NRC I and NRC II 1 error in 103 vs. 1 error in 1010 US DOJ at Daubert hearing on fingerprinting 1 error in 1097 Iris Recognition 1 error in 1078 5 A SHORT HISTORY (cont’) • • • • • • Speaker Verification - 1950 Fingerprinting - 1960 Hand & Retina - 1970 Face - 1980 Poock’s access control lab - 1985 Miller’s Personal ID News - 1987 6 MILLER’S DEFINITION “The automatic identification or identity verification of living, human individuals based on physiological or behavioral characteristics” 7 A SHORT HISTORY (cont) • • • • • • • Sandia National Lab Report - 1991 Biometric Consortium - 1992 Iris recognition - 1994 FERET - 1994 SJSU’s Biometric Lab - 1995 Real Biometricians get mad -- 1996 BC goes to Dept. of Commerce - 1999 8 THE PROBLEM • What can be generalized? • What can be stolen? 9 GENERAL APPROACH REQUIREMENTS • Applicable to all biometric technologies • Minimally disruptive to current approaches/vocabularies • Leading to mathematically elegant descriptions 10 OUR FINDINGS (THE REALLY GOOD STUFF) • • • • • • System description New words/concepts Distributional issues “Best Practices” Test results Statistical analysis – Bickel’s equation – Cotton ball squashing 11 SYSTEM DESCRIPTION DATA COLLECTION SIGNAL PROCESSING DECISION BIOMETRIC PATTERN MATCHING DECISION PRESENTATION QUALITY CONTROL SENSOR FEATURE EXTRACTION TRANSMISSION COMPRESSION STORAGE DATABASE EXPANSION TRANSMISSION IMAGE STORAGE 12 NEW WORDS/CONCEPTS • False rejection/acceptance of a hypothesis • Two possible hypotheses: – Sample in the database – Sample not in the database 13 NEW WORDS/CONCEPTS • Basic measures: – – – – – – False match False non-match Failure to acquire/enroll Throughput rates Bin error Penetration rate • Equations to derive false rejection/acceptance from basic measures 14 DISTRIBUTIONS • Good progress on analytic representations 15 “BEST PRACTICES FOR TESTING AND REPORTING BIOMETRIC DEVICE PERFORMANCE” www.cesg.gov.uk/biometircs • Stolen directly from NIST SV protocols and Philips, Martin, Przybocki (2000) • Technical, scenario, operational tests • Good faith genuine attempts • Zero-effort, unknown impostors • Uniformity or randomization of environmental (i.e.channel) variation 16 TEST RESULTS 17 FVC2000 -- UNIV. OF BOLOGNA FRR 1 FNMR 1 Sag1 Sag2 Cspn Cetp Cwai Krdl Uinh10^-1 Utwe Diti Fpin Ncmi 10-1 10^-2 10-2 10-3 10-5 10-4 10-3 10-2 10-1 FMR 10^-3 10^-5 10^-4 10^-3 10^-2 FAR 10^-1 18 DoD FRT2000 (www.dodcounterdrug.com/facial recognition/FRT2000/documents. htm) • One-Year Aging under FERET Lighting 19 CESG/NPL TEST PROGRAM 20 “BEST OF THREE” DET 21 KNOWN AND UNKNOWN IMPOSTORS standard normalised (best practices followed) normalised (best practices not followed) False Rejection Rate 100% 10% 1% 0.1% 0.0001% 0.001% 0.01% 0.1% 1% 10% 100% False Acceptance Rate 22 BICKEL’S EQUATIONS • Cross-comparisons are not independent • Find confidence interval for false match rate when cross comparisons are used 23 BICKEL’S EQUATIONS 24 COTTON BALL SQUISHING • For biometric measures in vector space, can we relate – size of space – number and distribution of templates – measurement error radius (no “zoo”) • to the probability of successful impostors? 25 COTTON BALL SQUISHING by James2 and Wayman P{no squishing} = P{Y>2r} • where Y is minimum distance between any of the M templates and r is the uniform template error radius in Rd by Onoyama, et al (1983) • P{Y>2r} = exp (-c M2 2d rd) • c = volume of ball in Rd * ||f(x)||2 /2 • and f(x) is the density function of the M templates26 THE ROAD AHEAD 1. Statistical analysis -- “Doddington’s Zoo” 2. Security Assessment National Information Assurance Partnership (US) Common Criteria (ISO 15408) 2nd Annual International CC Conference Brighton July 18-19 27 THE ROAD AHEAD 3. Vulnerability X9.84 Annex E T. van der Putte and J. Keuning, “Biometrical Fingerprint Recognition: Don’t Get Your Fingers Burned”, Proc IFIP TC8/WG8.8 (Kluwer Academic Press, 2000) WVU Center for Identification Technology 28 Research THE ROAD AHEAD 4. Privacy National Research Council, “Authentication Technologies and their Impact on Privacy” 5. Business case Fingerprinting in social services INSPASS 29 Banking industry