ReFaceX: donor-driven reversible face anonymisation with detached recovery.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Current facial image anonymization methods struggle to balance identity protection with data utility.
• Reconstruction-focused methods blur important details, while latent editing techniques often fail to remove identity cues effectively.

Purpose of the Study:

• To develop a novel reversible anonymization framework, ReFaceX, that effectively separates identity information from facial imagery.
• To preserve non-identity content like pose, background, and expression while ensuring robust identity protection.

Main Methods:

• ReFaceX utilizes a U-Net anonymizer with Identity Feature Fusion, guided by a donor identity code.
• A steganographic channel embeds a compact recovery payload, with blocked reconstruction gradients to prevent identity leakage.
• The framework explicitly defines threat models and audits outcomes using strong facial recognition algorithms.

Main Results:

• ReFaceX significantly reduces identity similarity across multiple facial recognition models (FaceNet, ArcFace, AdaFace) on LFW and CelebA-HQ datasets.
• The framework demonstrates improved recovered image quality, with high SSIM, LPIPS, and PSNR scores.
• ReFaceX operates in real-time on a single GPU and shows robustness to JPEG re-encoding.

Conclusions:

• ReFaceX offers a practical and auditable solution for responsible facial imagery release, balancing privacy and utility.
• The framework serves as a foundation for future extensions to video, higher resolutions, and enhanced recovery guarantees.