Generating Reversible and Identifiable Faces for Privacy Protection

Abstract

Face anonymization, which aims to conceal a person’s identity in visual contents while preserving visual utility of anonymized faces, plays a critical role in privacy protection of visual data. The challenge of face anonymization is how to achieve a high protection level without compromising the usability of original data, which has yet to be well addressed in previous works. In this paper, we propose RIFT, a Reversible and Identifiable Face Transformer to generate reversible and identifiable faces (RIFs) that meet this challenge. Our RIFs are visually and statistically different from original faces to ensure privacy, while being bound to virtual identities for direct authentication. Moreover, the original faces can be accurately recovered, enabling original identity tracing and better virtual identity management. Concretely, RIFT employs a normalizing flow to project original faces into a latent space, then reversibly transforms them into RIFs using condition embeddings derived from key-dependent latent. To support this, we introduce an attention-based conditional affine coupling layer to integrate identity control, and design a set of loss functions to enforce reversibility, anonymity, diversity, and identifiability. Experimental results show that RIFT significantly outperforms existing methods in anonymity, identifiability, reversibility, and robustness against malicious attacks.