Research map / Adversarial robustness and security / Attacks on text and graphs
Margin-based Neural Network Watermarking
Byungjoo Kim, Suyoung Lee, Seanie Lee, Sooel Son, Sung Ju Hwang
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 |
|---|---|---|---|
| margin-based watermarking (maximize trigger-sample margins via projected gradient ascent) | source model trained with trigger set · CIFAR-10 / CIFAR-100 | trigger set accuracy of source model: 100.00 vs 95.67 (RS) and 0 vanilla | Improved |
| margin-based watermarking instead of EWE/RS | soft-label surrogate · CIFAR-10 | trigger set accuracy of surrogate: 82.00 vs 26.88 (EWE) and 7.67 (RS) | Improved |
| margin-based watermarking instead of EWE/RS | surrogates with hard-label and regularized-GT distillation · CIFAR-10 | trigger set accuracy of surrogate: 51.33 (hard) and 72.67 (reg w. GT) vs baselines at most 51.01 and 36.05 (EWE) or 7.67 / 6.33 / 3.00 (RS) | Improved |
| margin-based watermarking instead of EWE/RS | surrogate models on CIFAR-100 · CIFAR-100 | trigger set accuracy of surrogate: 70.67 (soft-label) vs 30.90 (EWE) and 2.67 (RS) | Improved |
| margin-based watermarking instead of EWE/RS | source model accuracy · CIFAR-10 | clean accuracy of source model: 87.81 vs EWE 86.10 and RS 84.17 but below DI 92.03 and EEF 91.86 | Mixed |
| margin-based watermarking | source model p-value under statistical testing · CIFAR-10 | p-value of ownership test: 10^-12 vs 10^-7 (EEF) and 10^-3 (DI) | Improved |
| SVHN as surrogate dataset (heterogeneous) (ablation) | margin-based watermarked model · CIFAR-10 surrogate with SVHN | trigger set accuracy of surrogate: 72.00 vs 82.00 with CIFAR-10 surrogate data | Hurt |
| VGG11 surrogate architecture instead of ResNet34 (ablation) | margin-based watermarked model · CIFAR-10 surrogate with VGG11 | trigger set accuracy of surrogate: 32.00 vs 82.00 with same-architecture surrogate | Hurt |
| randomly chosen trigger-set labels vs highest/lowest confidence labels (ablation) | margin-based watermarking · CIFAR-10 trigger labeling | source model trigger accuracy: perfect accuracy on both extreme trigger sets | No clear effect |
| regularizing surrogate with ground-truth loss (alpha=0.3) (ablation) | surrogate model under margin-based watermark · CIFAR-10 | trigger set accuracy of surrogate: 72.67 vs 82.00 soft-label | No clear effect |
Closest papers by meaning
- Towards Robust Model Watermark via Reducing Parametric Vulnerability (ICCV 2023 · similarity 0.90)
- ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark (NeurIPS 2024 · similarity 0.87)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.87)
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.86)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · similarity 0.85)
- Task-Agnostic Language Model Watermarking via High Entropy Passthrough Layers (AAAI 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.
