Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Secure and reversible face anonymization with diffusion models

Pol Labarbarie, Vincent Itier, William Puech

TL;DR

This paper tackles the privacy risks of face imagery by proposing a diffusion-model–based reversible anonymization framework that incorporates a secret key and facial masking to securely de-anonymize only authorized users. By treating the forward diffusion latent as a Gaussian and deterministically steering the process with a secret-key–driven perturbation, the method achieves strong identity obfuscation while enabling exact recovery with the correct key, all without retraining a diffusion model. The approach yields high-quality anonymized faces and robust de-anonymization performance, significantly reducing identity leakage compared to prior methods. The proposed pipeline has practical implications for privacy-preserving analytics and for scenarios requiring controlled, auditable re-identification, with extensions to video and attack-resilience evaluation suggested for future work.

Abstract

Face images processed by computer vision algorithms contain sensitive personal information that malicious actors can capture without consent. These privacy and security risks highlight the need for effective face anonymization methods. Current methods struggle to propose a good trade-off between a secure scheme with high-quality image generation and reversibility for later person authentication. Diffusion-based approaches produce high-quality anonymized images but lack the secret key mechanism to ensure that only authorized parties can reverse the process. In this paper, we introduce, to our knowledge, the first secure, high-quality reversible anonymization method based on a diffusion model. We propose to combine the secret key with the latent faces representation of the diffusion model. To preserve identity-irrelevant features, generation is constrained by a facial mask, maintaining high-quality images. By using a deterministic forward and backward diffusion process, our approach enforces that the original face can be recovered with the correct secret key. We also show that the proposed method produces anonymized faces that are less visually similar to the original faces, compared to other previous work.

Secure and reversible face anonymization with diffusion models

TL;DR

This paper tackles the privacy risks of face imagery by proposing a diffusion-model–based reversible anonymization framework that incorporates a secret key and facial masking to securely de-anonymize only authorized users. By treating the forward diffusion latent as a Gaussian and deterministically steering the process with a secret-key–driven perturbation, the method achieves strong identity obfuscation while enabling exact recovery with the correct key, all without retraining a diffusion model. The approach yields high-quality anonymized faces and robust de-anonymization performance, significantly reducing identity leakage compared to prior methods. The proposed pipeline has practical implications for privacy-preserving analytics and for scenarios requiring controlled, auditable re-identification, with extensions to video and attack-resilience evaluation suggested for future work.

Abstract

Face images processed by computer vision algorithms contain sensitive personal information that malicious actors can capture without consent. These privacy and security risks highlight the need for effective face anonymization methods. Current methods struggle to propose a good trade-off between a secure scheme with high-quality image generation and reversibility for later person authentication. Diffusion-based approaches produce high-quality anonymized images but lack the secret key mechanism to ensure that only authorized parties can reverse the process. In this paper, we introduce, to our knowledge, the first secure, high-quality reversible anonymization method based on a diffusion model. We propose to combine the secret key with the latent faces representation of the diffusion model. To preserve identity-irrelevant features, generation is constrained by a facial mask, maintaining high-quality images. By using a deterministic forward and backward diffusion process, our approach enforces that the original face can be recovered with the correct secret key. We also show that the proposed method produces anonymized faces that are less visually similar to the original faces, compared to other previous work.

Paper Structure

This paper contains 12 sections, 7 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Reversible anonymization with diffusion models
  4. Using the Gaussian property for anonymization
  5. Deterministic forward and backward processes
  6. Overall pipeline
  7. Experiments & Results
  8. Experimental settings
  9. Face anonymization
  10. Original face recovery
  11. Wrong password de-anonymization
  12. Conclusion

Figures (3)

  • Figure 1: Diagram of our method for face anonymization using a diffusion model. On top in blue and on bottom in green are represented the anonymization and de-anonymization procedure, respectively. SD is the abbreviation for Stable Diffusion.
  • Figure 2: Qualitative comparison of face anonymization (second group of columns) and recovery (last group of columns, "method's name"-R) among different methods on the CelebA-HQ dataset.
  • Figure 3: Qualitative comparison of de-anonymized faces with wrong passwords on the CelebA-HQ dataset for our method. Faces are anonymized using $k_1$ and are then de-anonymized using either $k_2$ or $k_3$.