IDDiffuse: Dual-Conditional Diffusion Model for Enhanced Facial Image Anonymization

Muhammad Shaheryar, Jong Taek Lee, Soon Ki Jung; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 4017-4033

The increasing prevalence of computer vision applications in public spaces has raised substantial privacy concerns regarding facial image data. Traditional anonymization methods, despite their potential, often suffer from drawbacks such as limited output variety, inadequate detail, distortions in extreme poses, and inconsistent temporal patterns. This study introduces an identity diffuser based on a dual-conditional diffusion model that efficiently anonymizes facial images while preserving task-relevant features for diverse applications. Our approach ensures a clear separation from the original identity by utilizing synthetic identities and an optimized identity feature space derived from three state-of-the-art models. It maintains consistency across frames for video anonymization. Unlike existing methods, our approach eliminates the need for task-relevant feature extractors, such as those for pose and expression. Instead, it employs a dual-condition diffusion model to integrate both identity and non-identity information, offering improved anonymization without compromising data usefulness. Our technique enables seamless transitions from real to synthetic identities by incorporating a time-step-dependent ID loss, providing controllable identity anonymization. Extensive studies demonstrate that our method achieves superior de-identification rates and consistency compared to state-of-the-art techniques, preserving non-identity features with a 20% improvement in emotion recognition, handling extreme poses with enhanced image quality, output diversity, and temporal consistency. This makes it a valuable tool for privacy-preserving computer vision applications.