Face image anonymization aims to preclude re-identification while preserving images’ analytical utility and visual quality. Although existing image anonymization approaches can conceal identity cues of faces, they either lack privacy guarantees and/or result in an unacceptable trade-off between privacy and utility. To address this challenge, we propose Generative BlendPose-Latent Diffusion Model (LDM), a framework for face image anonymization that leverages k -anonymity privacy guarantees with the capabilities of latent diffusion models, to generate high-quality anonymous images. Our approach uses uniformly blended pose-guided and class-consistent k -anonymous images together with text prompts aimed at preserving non-identifying demographic features and guide the diffusion-based denoising. Comprehensive comparisons against related works –including plain k -anonymous and differentially private mechanisms and generative methods–, show that our method results in stronger anonymization, better utility preservation, and improved visual fidelity and realism.
Ali et al. (Wed,) studied this question.