Secure biometric hashing against relation-based attacks via maximizing min-entropy

Abstract

With its irreversibility, low computational cost, and high storage efficiency, biometric hashing is widely used in privacy-preserving biometric recognition systems, yet its security has been recently challenged due to the emergence of relation-based attacks (RA). To address this issue, we model and analyse the RA, and discover that maximizing the conditional min-entropy of the signs of inter-class distance differences in the original space and hash space can minimize the leakage of the distance relation on the original biometric. Consequently, we develop a Secure Biometric Hashing scheme against Relation-Based Attacks (SBH-RA). SBH-RA maximizes the conditional min-entropy by using the cosine function as a distance mapping function. Meanwhile, it learns hash codes by a classification loss and quantization loss to ensure the accuracy of recognition. Our study demonstrates that SBH-RA not only offers higher security but also yields comparable or even superior recognition performance over existing biometric hashing methods experimentally and theoretically. Given 1024 bits hash codes from SBH-RA, it brings a decrease of 21% in Equal Error Rate on face dataset LFW compared with widely used Biohashing. Besides, even under white-box attacks, the probability of a successful attack is smaller than 1.69×2−346.