Facial Identity Anonymization via Intrinsic and Extrinsic Attention Distraction

TL;DR

This work tackles face anonymization by distracting identity-related attention in both intrinsic feature space and extrinsic visual cues. It introduces a two-step pipeline: first preprocess faces with intrinsic identity feature anonymization (IFA) and visual clue anonymization (VCA), then synthesize anonymous faces via a conditional generator that preserves data utility and supports diverse, user-controllable outputs. The approach combines Grad-CAM-based attention manipulation, instance-level DP-inspired visual clue delegation, and a GAN-based synthesis framework with AdaIN conditioning, optimized by a multi-task loss that balances realism, identity obfuscation, and utility preservation. Experimental results across CelebA-HQ, VggFace2, and LFW demonstrate competitive privacy protection (low ReID/IDS) while maintaining facial attribute integrity and image quality, with supporting ablation and user studies confirming practical effectiveness and controllability.

Abstract

The unprecedented capture and application of face images raise increasing concerns on anonymization to fight against privacy disclosure. Most existing methods may suffer from the problem of excessive change of the identity-independent information or insufficient identity protection. In this paper, we present a new face anonymization approach by distracting the intrinsic and extrinsic identity attentions. On the one hand, we anonymize the identity information in the feature space by distracting the intrinsic identity attention. On the other, we anonymize the visual clues (i.e. appearance and geometry structure) by distracting the extrinsic identity attention. Our approach allows for flexible and intuitive manipulation of face appearance and geometry structure to produce diverse results, and it can also be used to instruct users to perform personalized anonymization. We conduct extensive experiments on multiple datasets and demonstrate that our approach outperforms state-of-the-art methods.

Figures (11)

• Figure 1: Demonstration of our approach for face anonymization.
• Figure 2: Overview of our approach: (a) the flowchart, (b) identity feature anonymization (IFA), and (c) visual clue anonymization (VCA).
• Figure 3: Intuitive comparison of our approach with the existing anonymization methods, where the first column presents the original faces.
• Figure 4: Demonstration of our diverse anonymous results.
• Figure 5: Demonstration of our results on controlling the geometry structures and visual appearances. Each of the top row shows the employed geometry structures and visual appearance images.