Deep Secure Quantization: On secure biometric hashing against similarity-based attacks

Abstract

Abstract The widespread application of biometric recognition has emerged solid protection on the privacy of biometric templates. Non-invertible transformations such as random projections are popular solutions for this purpose, yet their security has been recently challenged due to the advent of similarity-based attacks (SA). To address this issue, we developed Deep Secure Quantization (DSQ), a new biometric hashing scheme for privacy-preserving biometric recognition. DSQ essentially takes into account the information leakage between the original distance and the hashed distance, which is the security blind spot of existing hashing models. This leakage is further incorporated into an optimal hashing objective which well balances between security and utility. Hashing is then modeled as a highly nonlinear problem solved by a novel deep neural network. Experiments on CASIA-v4-interval demonstrate that DSQ not only offers strong resistance to SA but also yields comparable or even superior recognition performance over existing biometric hashing methods, including deep framework-based ones.