Research map / Adversarial robustness and security / Black-box and transfer attacks
Removing Box-Free Watermarks for Image-to-Image Models via Query-Based Reverse Engineering
Haonan An, Guang Hua, Hangcheng Cao, Zhengru Fang, Guowen Xu, Susanto Rahardja, Yuguang Fang
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 |
|---|---|---|---|
| train a forward surrogate of HNet instead of an inverse of HNet | query-based watermark removal · Deraining, victim VZhang | PSNR (dB): 33.98 -> 34.69 | Improved |
| train a forward surrogate of HNet instead of an inverse of HNet | query-based watermark removal · Image generation, victim VWu | PSNR (dB): 31.75 -> 32.05 | Improved |
| train a forward surrogate of HNet instead of an inverse of HNet | query-based watermark removal · Image generation, victim VZhang | MS-SSIM: 0.991 -> 0.990 | No clear effect |
| query with identity-transformation images that bypass GNet (rather than images from domain A) (ablation) | HNet inverse attack · Image generation | image quality | Improved |
| query with identity-transformation images that bypass GNet (rather than images from domain A) (ablation) | Forward HNet attack · Image generation | watermark removal | Improved |
| apply Forward HNet attack versus generic JPEG compression | VWu deraining · Deraining, victim VWu | PSNR (dB): 26.53 -> 33.75 (JPEG-20% vs Forward HNet) | Improved |
| apply Forward HNet attack versus WEvade-B-Q | VWu deraining · Deraining, victim VWu | removal success rate: 0.380 -> 1.000 | Improved |
| apply Forward HNet attack versus AWGN-30dB noise | VWu image generation · Image generation, victim VWu | removal success rate: 0.000 -> 1.000 | Improved |
| apply attack versus Regeneration-VAE | VZhang image generation · Image generation, victim VZhang | PSNR (dB): 32.54 -> 32.60 | No clear effect |
| approximate delta of ai by delta of bi (additive equivalence) | Forward HNet attack · Deraining, victim VZhang | correlation: 0.635 (deraining VZhang) vs 0.894 (image generation VZhang) | Mixed |
Closest papers by meaning
- Decoder Gradient Shield: Provable and High-Fidelity Prevention of Gradient-Based Box-Free Watermark Removal (CVPR 2025 · similarity 0.87)
- WMCopier: Forging Invisible Watermarks on Arbitrary Images (NeurIPS 2025 · similarity 0.87)
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.87)
- Invisible Image Watermarks Are Provably Removable Using Generative AI (NeurIPS 2024 · similarity 0.86)
- Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles (NeurIPS 2026 · similarity 0.86)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.86)
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.
