Research map / Adversarial robustness and security / Attacks on text and graphs
Watermarking Deep Neural Networks with Greedy Residuals
Hanwen Liu, Zhenyu Weng, Yuesheng Zhu
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 |
|---|---|---|---|
| select only half of extracted weights (eta=0.5, greedy residuals) instead of all weights (eta=1) (ablation) | greedy residuals watermark · CIFAR-10 AlexNet pruning | watermark detection under pruning | Improved |
| select half (eta=0.5) instead of all weights (eta=1) (ablation) | greedy residuals watermark · CIFAR-10 AlexNet clean accuracy | accuracy (%): accuracy drops by up to 0.22% when all weights are selected | No clear effect |
| embed greedy residuals watermark (vs. passport/backdoor) | AlexNet/ResNet-18 image classification · Table 1 fidelity | accuracy drop vs baseline: maximum drop 1.8% (ResNet-18 with BN on Caltech-256) | No clear effect |
| embed greedy residuals watermark | LSTM/TextCNN text classification · IMDB-2 and TREC-6 | accuracy: IMDB-2 LSTM -0.31 points; TREC-6 TextCNN +1.00 point | Mixed |
| embed greedy residuals watermark | ImageNet subset AlexNet · ImageNet subset (100 classes) | accuracy (%): 54.09 to 53.33 | Hurt |
| greedy residual watermark vs. Passport under fine-tuning (worst case, last layer replaced) | greedy residuals watermark · CIFAR-10 to CIFAR-100 fine-tuning | watermark detection rate (%): 100% in all experiments vs Passport slight decline | Improved |
| Backdoor watermark under fine-tuning with replaced output layer | backdoor baseline · fine-tuning resistance | watermark detection rate (%): drops to 0 | Hurt |
| greedy residuals watermark vs. backdoor/passport under forging attacks | greedy residuals watermark · forging attacks (200 random signatures) | detection accuracy of forged watermark: near 50% vs 100% for Passport/Backdoor | Improved |
| greedy residuals watermark under overwriting attacks | greedy residuals watermark · CIFAR-10/100 overwriting attacks | detection rate of real watermark: real watermark ~100%; Backdoor drops to 13% without data | Improved |
| greedy residuals watermark under ell1 weight pruning | greedy residuals watermark · CIFAR-10 ResNet-18 pruning | watermark detection rate: drops only at ~75% sparsity where accuracy ~44.74% | Improved |
| greedy residuals watermark under permutation and scaling attacks | greedy residuals watermark · permutation and scaling attacks | watermark detection rate: all remain above 95% (permutation) and 100% (scaling) | No clear effect |
Closest papers by meaning
- ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark (NeurIPS 2024 · similarity 0.90)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · similarity 0.89)
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.87)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.86)
- Safe and Robust Watermark Injection with a Single OoD Image (ICLR 2024 · similarity 0.86)
- Resource Efficient Deep Learning Hardware Watermarks with Signature Alignment (AAAI 2024 · similarity 0.86)
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.
