Privacy-First Attribute Preserving Face Anonymization via ID-Augmented Contrastive Learning

Face anonymization suppresses identity in visual data while retaining useful expression, gaze, and pose. Imperfect face swappers can leave identity cues and distort these attributes. We introduce ID-augmented contrastive learning: synthetic identity variants supply attribute-consistent positives, while different real images of the same person supply negatives that discourage residual identity encoding. The learned attribute code conditions a latent inpainting diffusion model with inference-time identity guidance. On FFHQ, negative guidance reduces target-identity similarity to 0.04 while retaining low pose and expression errors; ablations support the benefit of same-ID negatives. This framework applies AI to visual data sanitization, with a controllable privacy--utility trade-off when images are shared or processed downstream.