Research map / Adversarial robustness and security / Black-box and transfer attacks
Class-feature Watermark: A Resilient Black-box Watermark Against Model Extraction Attacks
Yaxin Xiao, Qingqing Ye, Zi Liang, Haoyang Li, RongHua Li, Huadi Zheng, Haibo Hu
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 |
|---|---|---|---|
| use WRK (boundary reshaping + feature shifting) to remove watermarks | entangled black-box watermarks · CIFAR-10 ResNet18 | watermark success rate (%): EWE 99.88 -> 8.75; Blend 99.95 -> 9.80 | Improved |
| apply existing removal methods against entangled watermarks | NC, I-BAU, CLP, FP, NAD, SEAM, FST · CIFAR-10 ResNet18 | watermark success rate (%) | No clear effect |
| use class-level artifact (OOD synthetic class) instead of sample-level artifacts | CFW · CIFAR-10 to Speech Commands, copy model after MEA+WRK | WSR_LC (%): WSR_LC >= 70.15 under MEA and WRK | Improved |
| add RepS loss (ablation) | CFW · CIFAR-10 copy model | WSR_LC (%): 70.00 -> 94.00 (w/ RepS only) | Improved |
| use SNNL instead of RepS for entanglement (ablation) | CFW · CIFAR-10 victim | accuracy (%): 93.70 -> 90.13 | Hurt |
| add CD2 loss (ablation) | CFW · CIFAR-10 copy model | WSR_LC (%): 84.00 -> 94.00 (CD2 only vs full) | Improved |
| add CD2 loss (ablation) | CFW · CIFAR-10 copy model | intra-class variance (x10^2): 10.30 -> 2.74 | Improved |
| add RepS and CD2 together vs CD2 only (ablation) | CFW · CIFAR-10 copy model | representation entanglement RE: 0.04 -> 0.75 | Improved |
| verify via label clustering instead of WSR | CFW · CIFAR-10 under WRK | WSR vs clustering: WSR 100% -> 19.60% while representations stay clustered | Improved |
| apply CFW on ImageNette | CFW · ImageNette | accuracy: about 2% utility drop | Hurt |
Closest papers by meaning
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.88)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.87)
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models (ICML 2024 · similarity 0.87)
- StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Defense against Model Extraction Attack by Bayesian Active Watermarking (ICML 2024 · similarity 0.87)
- Robustness of AI-Image Detectors: Fundamental Limits and Practical Attacks (ICLR 2024 · 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.
