Research map / Adversarial robustness and security / Black-box and transfer attacks
Finding a needle in a haystack: A Black-Box Approach to Invisible Watermark Detection
Minzhou Pan, Zhenting Wang, Xin Dong, Vikash Sehwag, Lingjuan Lyu, Xue Lin
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 asymmetric loss (softmax on clean data, linear on detection data) instead of symmetric linear loss (ablation) | WMD · ImageNet DctDwtSvd | AUC: 0.786 -> 0.968 | Improved |
| use asymmetric loss instead of symmetric BCE loss (ablation) | WMD · ImageNet DctDwtSvd | AUC: 0.757 -> 0.968 | Improved |
| use exponential loss on detection data (maximize side) (ablation) | WMD · ImageNet DctDwtSvd | AUC: 0.968 -> 0.513 | Hurt |
| add iterative pruning with model reinitialization to the detection dataset (ablation) | WMD · ImageNet | AUC | Improved |
| raise the pruning rate to a high value (ablation) | WMD · ImageNet | AUC | Hurt |
| lower the pruning rate from 50% to 1% (ablation) | WMD · ImageNet | compute overhead: approximately 66.2 times more overhead | Mixed |
| apply multiple watermark types at once instead of a single type | WMD · ImageNet LSB | AUC: 0.852 -> 0.732 | Hurt |
| replace the clean reference set with a mismatched-domain dataset (CelebA) (ablation) | WMD · Detection ImageNet | AUC: 0.968 -> 0.612 | Hurt |
| replace 80% of the clean set with synthesized images from detection-set captions (ablation) | WMD · ImageNet DctDwtSvd | AUC: 0.968 -> 0.684 | Hurt |
| replace 20% of the clean set with synthesized images (case 1) (ablation) | WMD · ImageNet DctDwtSvd | AUC: 0.968 -> 0.966 | No clear effect |
| use WMD detection output to build a pure watermarked set for PGD removal | black-box watermark removal · Tree-Ring | AUC after removal: AUC 0.504 after removal | Improved |
Closest papers by meaning
- WMarkGPT: Watermarked Image Understanding via Multimodal Large Language Models (ICML 2025 · similarity 0.87)
- Black-Box Detection of Language Model Watermarks (ICLR 2025 · similarity 0.87)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.86)
- Mask Image Watermarking (NeurIPS 2025 · similarity 0.86)
- WMCopier: Forging Invisible Watermarks on Arbitrary Images (NeurIPS 2025 · similarity 0.86)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.85)
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.
