Identity-Consistent Video De-identification via Diffusion Autoencoders

With the rise of deep learning and the widespread use of face recognition, face image privacy has become a critical research issue. Face de-identification is acknowledged as effective for protecting identity privacy. As media formats diversify, it is imperative to extend privacy protection to videos. Addressing the core problem of identity consistency between frames, we propose a video de-identification approach based on the diffusion model. We disentangle video features with a diffusion autoencoder, where the identity and motion features are encoded into high-level semantic spaces while background and other facial identity-independent features into low-dimensional random subcodes. The unified time-independent identity representation is used to achieve coherent video de-identification results. Compared to existing methods, our proposed approach demonstrates superior performance in terms of privacy protection effectiveness, identity consistency between frames, and utility.

I. Introduction

The privacy protection of face images is an emerging research issue in computer vision, dedicated to addressing privacy infringements arising from the rise of deep learning technologies and the widespread application of face recognition. Face images are considered one of the most privacy-sensitive biometrics, closely linked to personal identity.