Beyond Texture: Advanced Facial Privacy Protection via Hierarchical Diffusion Autoencoder

Paper i proceeding, 2027

Facial privacy protection on social media is increasingly critical due to the widespread use of face recognition technology. Existing methods often require visually noticeable alterations that compromise the naturalness of faces. In this paper, we develop a hierarchical diffusion autoencoder to achieve subtle semantic-level modifications beyond texture variations, including makeup, eyebrow density, and complexion that preserve the natural appearance of faces while protecting against unauthorized recognition systems. We validate our method on popular face image datasets, including CelebA-HQ and LFW, and demonstrate superior performance in untargeted attack scenarios compared to state-of-the-art methods. Furthermore, evaluations with commercial face recognition APIs confirm the practical efficacy of the proposed approach, showing significant improvements in the Protection Success Rate (PSR) without compromising user experience on social media platforms.