AFTI: High-generalized facial template inversion under attribute guidance.

Abstract

Face recognition has brought tremendous convenience to daily life, enabling applications like access control, mobile payment, and criminal identification. However, facial templates stored on servers pose significant privacy risks due to potential security breaches. To underscore this concern, facial template inversion-reconstructing face images from templates-has been widely studied. While recent approaches adopt pixel-wise/perceptual losses and elaborate networks to minimize reconstruction errors, they suffer from limited performance in Type-II attack scenarios and unsatisfied generalization across unseen models. In this work, we propose Attribute-guided Facial Template Inversion (AFTI), a strategy that leverages facial attributes (e.g., eye shape, nose structure, mouth characteristics) to impose explicit semantic constraints, mitigating the ambiguity of template-to-image mapping. We introduce an attribute loss and integrate it into two frameworks: UCNet (a simple deep network) and the state-of-the-art M-StyleGAN3. Experimental results show AFTI enhances existing SOTA methods, achieving higher True Accept Rates under Type-II attacks and improved cross-model generalization.