Privacy-Preserving Face Recognition for Access Control Systems

Face recognition (FR) technology, a highly secure biometric authentication method, has been widely applied in physical access control systems (ACSs). However, the facial information uploaded by users in the system is vulnerable to third parties, which can be utilized to train unauthorized FR models, which could be used for illegal identification of users in unknown contexts, posing significant threats to user privacy. Although many privacy-preserving FR methods have been proposed, they are rarely directly applicable to existing physical ACSs because they require invasive modifications to the core FR algorithms in ACSs. In this paper, we present a non-intrusive facial privacy protection method based on adversarial example technique, which does not require to modify any software or hardware of the target ACS. It first adds subtle perturbations that do not affect visual perception to the facial photos submitted by users, ensuring that the photos could pass the identity verification of human administrators of ACSs but the FR models built on them would mis-recognize real users. Then, it trains physical adversarial example stickers using masks as carriers, which, when worn on the mouth, users could be correctly recognized, thus passing through the ACS successfully. According to experiments conducted on multiple test subjects and various face recognition models, the validity of the proposed method has been demonstrated.