Research map / Adversarial robustness and security / Attacks on text and graphs
THE SELF-RE-WATERMARKING TRAP: FROM EXPLOIT TO RESILIENCE
Vithurabiman Senthuran, Yong Xiang, Iynkaran Natgunanathan, Uthayasanker Thayasivam
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 re-watermarking adversarial training and spectral normalization (full model) vs existing SOTA | self-re-watermarking attack · COCO | ACC_orig (%): 49.90 -> 100.00 | Improved |
| add spectral normalization (Lipschitz constraint) (ablation) | Proposed · COCO | ACC_orig after self overwrite (%): 76.33 -> 100.0 | Improved |
| add spectral normalization (ablation) | Proposed · COCO | PSNR (dB): 30.40 -> 34.03 | Improved |
| add spectral normalization (ablation) | Proposed · COCO | ACC_orig after strong PGD (%): 98.90 -> 99.37 | Improved |
| add post-processing module (ablation) | Proposed · COCO | PSNR (dB): 31.82 -> 34.03 | Improved |
| add post-processing module (ablation) | Proposed · COCO | ACC_orig after strong PGD (%): 99.99 -> 99.37 | No clear effect |
| increase lambda_LPIPS from 0.5 to 0.7 (ablation) | Proposed · COCO | PSNR (dB) / JPEG(50) ACC: 34.03 -> 39.51; 95.06 -> 88.08 | Mixed |
| decrease lambda_LPIPS from 0.5 to 0.3 (ablation) | Proposed · COCO | PSNR (dB) / JPEG(50) ACC: 34.03 -> 32.39; 95.06 -> 99.63 | Mixed |
| increase removal attack step size before re-embedding | Proposed · COCO | ACC_orig (%): 68.11 -> 53.00 | Hurt |
| train on omitting robustness loss in adversary model (AM) vs BAM | adversary overwriting model · MIRFLICKR | ACC_orig of attacker's output: 1.00 vs 0.99 | No clear effect |
| increase PGD perturbation budget epsilon above 0.05 | Proposed · COCO | bit accuracy | Hurt |
| train and test on other datasets (MIRFLICKR, CelebA, ImageNet) | Proposed · MIRFLICKR | ACC_orig self re-embed (%): 100.00 -> 100.00 | No clear effect |
Closest papers by meaning
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.90)
- Revealing Weaknesses in Text Watermarking Through Self-Information Rewrite Attacks (ICML 2025 · similarity 0.89)
- Leveraging Optimization for Adaptive Attacks on Image Watermarks (ICLR 2024 · similarity 0.88)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.88)
- Decoder Gradient Shield: Provable and High-Fidelity Prevention of Gradient-Based Box-Free Watermark Removal (CVPR 2025 · similarity 0.88)
- END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions (AAAI 2025 · similarity 0.87)
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.
