Abstract
Advancements in face recognition (FR) technologies have amplified privacyconcerns, necessitating methods that protect identity while maintainingrecognition utility. Existing face anonymization methods typically focus onobscuring identity but fail to meet the requirements of biometric templateprotection, including revocability, unlinkability, and irreversibility. Wepropose FaceAnonyMixer, a cancelable face generation framework that leveragesthe latent space of a pre-trained generative model to synthesizeprivacy-preserving face images. The core idea of FaceAnonyMixer is toirreversibly mix the latent code of a real face image with a synthetic codederived from a revocable key. The mixed latent code is further refined througha carefully designed multi-objective loss to satisfy all cancelable biometricrequirements. FaceAnonyMixer is capable of generating high-quality cancelablefaces that can be directly matched using existing FR systems without requiringany modifications. Extensive experiments on benchmark datasets demonstrate thatFaceAnonyMixer delivers superior recognition accuracy while providingsignificantly stronger privacy protection, achieving over an 11% gain oncommercial API compared to recent cancelable biometric methods. Code isavailable at: https://github.com/talha-alam/faceanonymixer.