Research map / Label-efficient and robust learning / Recognition and long-tail
Rethinking Data Bias: Dataset Copyright Protection via Embedding Class-wise Hidden Bias
Jinhyeok Jang, ByungOk Han, Jaehong Kim, Chan-Hyun Youn
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 mean of target dataset mu(X) to the noise watermark when testing (ablation) | Noise patch watermark · CIFAR10 (ResNet18, no flip) | val acc on watermark (%): 62.14 -> 71.69 | Improved |
| apply horizontal flip augmentation to noise placement watermark (ablation) | Noise patch watermark · CIFAR10 (ResNet18) | val acc on watermark (%): 62.14 -> 20.26 | Hurt |
| overlay auxiliary dataset instead of noise patches (ablation) | Class-wise bias embedding · CIFAR10 (ResNet18, both flips) | val acc on watermark (%): 9.15 -> 69.05 | Improved |
| overlay auxiliary data at higher lambda (0.3 to 0.7) (ablation) | Overlay watermark · CIFAR10 (ResNet18, no flip) | val acc on benign (%): 90.47 -> 88.52 | Hurt |
| use steganography-based DWN invisible watermark rather than visible overlay | Undercover bias · CIFAR10 | avg. SSIM: 0.9785 vs 0.9775 (Blended), 0.9202 (Radioactive) | Improved |
| watermark training data with undercover bias | DenseNet-BC · CIFAR100 | val acc on benign (%): 71.79 -> 68.27 | Hurt |
| watermarked training of MobileNetV2 vs clean | MobileNetV2 · ImageNet | val mAcc on watermark (%): 11.90 -> 43.95 | Improved |
| use modulo pairing when target has more classes than auxiliary | Undercover bias · Tiny ImageNet | val mAcc on watermark (%): 8.84 -> 34.37 | Improved |
| use 2/N_cls threshold for verification | Undercover bias · CIFAR10, CIFAR100 histogram analysis | verification separation | Improved |
Closest papers by meaning
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.88)
- ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark (NeurIPS 2024 · similarity 0.88)
- Data Taggants: Dataset Ownership Verification Via Harmless Targeted Data Poisoning (ICLR 2025 · similarity 0.87)
- STAMP Your Content: Proving Dataset Membership via Watermarked Rephrasings (ICML 2025 · similarity 0.85)
- Disentangled Style Domain for Implicit $z$-Watermark Towards Copyright Protection (NeurIPS 2024 · similarity 0.85)
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (CVPR 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.
