Research map / Image generation / Editing and style transfer
FaceArmor: A Universal Facial Image Protection Against Diffusion-Based Manipulations
Yiming Wang, Jiahao Chen, Qingming Li, Chunyi Zhou, Zhi Chen, Lingzhong Meng, Jinbao Li, Shouling Ji
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add semantic loss LC to latent loss LE (ablation) | FaceArmor · Loss ablation (mean over 4 tasks) | LPIPS / FR: LPIPS 0.5215 -> 0.5329; FR 0.7615 -> 0.7540 | Improved |
| add LC, LF and LR to LE-only optimization (ablation) | FaceArmor · Loss ablation (mean over 4 tasks) | FR: FR 0.7615 -> 0.4522 | Improved |
| add all four losses to LE-only optimization (ablation) | FaceArmor · Loss ablation (mean over 4 tasks) | LPIPS: LPIPS 0.5215 -> 0.5781 | Improved |
| add region-aware (attention soft weight + PCGrad) optimization to all losses (ablation) | FaceArmor · Mean over all scenarios | LPIPS, SSIM, CLIP, FR (not FID) | Improved |
| add multiple losses to LE-only baseline (ablation) | FaceArmor · Mean over all scenarios | FID | No clear effect |
| increase perturbation budget (ablation) | FaceArmor · Perturbation budget sweep | protection effectiveness | Improved |
| increase PGD iterations from 10 to 50 (ablation) | FaceArmor · Iteration sweep | protection effectiveness | Improved |
| increase PGD iterations beyond 100 (ablation) | FaceArmor · Iteration sweep | protection effectiveness | No clear effect |
| add EOT training with transformations (JPEG QF=75), compared with baselines | FaceArmor · Robustness, JPEG QF=75 | LPIPS: FaceArmor 0.557 -> 0.431 vs FaceLock 0.456 -> 0.240 | Improved |
| transfer FaceArmor perturbations (crafted on SD1.5) to SD 3 | FaceArmor vs AdvDM · Image Inpainting, SD3 transfer | LPIPS: 0.2830 -> 0.3131 | Improved |
| transfer FaceArmor perturbations to Flux | FaceArmor vs AdvDM · Image Inpainting, Flux transfer | LPIPS: 0.2824 -> 0.2906 | Improved |
| use multi-objective multi-tier optimization instead of single-model attack | FaceArmor · Computation time | seconds per image: 28.5 -> 90.2 | Hurt |
Closest papers by meaning
- Enhancing Facial Privacy Protection via Weakening Diffusion Purification (CVPR 2025 · similarity 0.86)
- IDProtector: An Adversarial Noise Encoder to Protect Against ID-Preserving Image Generation (CVPR 2025 · similarity 0.86)
- Adversarial Robust Safeguard for Evading Deep Facial Manipulation (AAAI 2024 · similarity 0.85)
- Edit Away and My Face Will not Stay: Personal Biometric Defense against Malicious Generative Editing (CVPR 2025 · similarity 0.85)
- Nearly Zero-Cost Protection Against Mimicry by Personalized Diffusion Models (CVPR 2025 · similarity 0.84)
- Harnessing Global-Local Collaborative Adversarial Perturbation for Anti-Customization (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
