Learning Biometric Representations with Mutually Independent Features Using Convolutional Autoencoders

Abstract

Representations of biometric traits to be used in automatic recognition systems are typically learned with the goal of obtaining significant discriminative capabilities, that is, generating features that are notably different when produced by traits of different subjects, while maintaining an appropriate consistency for a given user. Nonetheless, discriminability is not the only desirable property of a biometric representation. For instance, the mutual independence of the coefficients in the employed templates is a valuable property when designing biometric template protection schemes. In fact, managing representations with independent coefficients allows to maximize the achievable security. In this paper we propose different learning strategies to obtain biometric representations with the property of statistical independence among coefficients, while preserving discriminability. In order to achieve this goal, different strategies are employed to train convolutional autoencoders. As a proof of concept, the effectiveness of the proposed approaches is tested by considering biometric recognition systems using both finger-vein and palm-vein patterns.