Abstract

No abstract

Abstract. This paper investigates a new approach to formulate performance indices of biometric system using information theoretic models. The performance indices proposed here (unlike conventionally used FAR, GAR, DET etc.) are scalable in estimating performance of large scale biometric system. This work proposes a framework for identification capacity of a biometric system, along with insights on number of cohort users, capacity enhancements from user specific statistics etc. While incorporating feature level information in a rate-distortion framework, we derive condition for optimal feature representation. Furthermore, employing entropy measures to distance (hamming) distribution of the encoded templates, this paper proposes an upper bound for false random correspondence probability. Our analysis concludes that capacity can be the performance index of a biometric system while individuality expressed in false random correspondence can be the performance index of the biometric trait and representation. This paper also derives these indices and quantifies them from system parameters.

Performance of biometrics in human authentication can be limited when multiple samples of a user's biometric information differ due to intra-class variability in acquisition, storage or transmission of biometrics. Thus, random correspondence results between users. We formulate Probability of Random Correspondence (PRC) by developing an information model of biometrics features as a noisy source. The information in features represented by N s bits inherently has t error bits attributed to the intra-class variabilities. The values of t and bit error probability are shown to be determined from second order statistics of the features. These are used respectively, to formulate information rate of the noisy biometric and to characterize a binary symmetric channel that models the occurrence of errors in a biometric template. Finally, information rate and error probability are combined in the framework of error exponents to formulate PRC of biometrics. We illustrate our approach with simulations, using freely available data, to obtain numerical values of PRC of fingerprint biometrics.

No abstract