Generative face privacy protection enables secure learning and analysis

Abstract

To effectively preserve personal privacy, particularly identity and sensitive attribute information in facial recognition, this study aims to develop a facial privacy-preserving mechanism that not only ensures facial privacy but also maintains image quality and practicality. This research introduces a diffusion model-based facial privacy-preserving framework, which leverages an identity encoder and an attribute encoder to effectively decompose facial representations. Additionally, differential privacy is enhanced by introducing Laplacian noise. Finally, the Denoising Diffusion Implicit Model (DDIM) serves as the decoder to generate privacy-preserving facial images. Experimental results demonstrate that this method not only excels in privacy preservation but also preserves image quality, while being compatible with various computer vision tasks, thereby successfully striking a balance between privacy preservation, image quality, and functionality.