Cancelable binary face templates generation based on density-sensitive hashing and feature hashing

Abstract

The increasing utilization of face recognition technology prompts concerns about the security of stored templates. Nevertheless, existing biometric template protection methods often incur high computational overhead or depend on two-factor input. To address these issues, we propose a cancelable template generation strategy that integrates density-sensitive hashing and feature hashing. The density-sensitive hashing transforms the facial feature vector into binary codes by leveraging the geometric characteristics of the data. Feature hashing then derives a permutation seed from the facial features to shuffle a random key, which is encoded using the binary codes, producing an encoded key retained within the cancelable template. Experimental results on the LFW, FEI and CASIA-FaceV5 databases show that our method achieves an EER below $0.72$%, a GAR exceeding $97.5$% at a FAR=$0.01$% and an average template generation time of 5.3 ms, confirming its efficiency and recognition performance. Furthermore, related experimental and theoretical evaluations prove that the proposed method guarantees the characteristics of irreversibility, revocability unlinkability, and resilience against various attacks.