DiffDeID: a Multi-conditional Diffusion-based Method for High Fidelity Face De-indentification with Diversity

Yanzhuo Wei, Yu Pan

Keywords: Face De-identification, Data privacy, Diffusion Model

Abstract:

Face de-identification is a critical task that aims to obscure true identities while preserving other facial attributes. Current methodologies typically involve disentangling identity features within a latent space and leveraging adversarial training to balance privacy with utility, often at the cost of a trade-off between two. To surmount these limitations, we introduce DiffDeID, a novel approach grounded in diffusion models. This method incrementally safeguards identity and sustains utility, all while ensuring enhanced interpretability.
Our method employs a Latent Diffusion-based ID Sample to generate authentic identity embeddings that are obfuscated from the original identity, thereby providing users with diverse options. Additionally, a multi-condition diffusion model is utilized for facial images, ensuring the retention of image utility. We further introduce a novel training and inference paradigm, utilizing the unified architecture tailored for video facial de-identification tasks.
The robustness of our method is attributed to its powerful 3D prior and meticulous generation design, enabling natural identity protection, generation of high-quality details, and robustness across various attributes. Through extensive experimentation, we demonstrate that DiffDeID surpasses previous methodologies.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9230