Research map / Adversarial robustness and security / Attacks on text and graphs
Hashed Watermark as a Filter: A Unified Defense Against Forging and Overwriting Attacks in Neural Network Watermarking
Yuan Yao, Jin Song, Jian Jin
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 hashed watermark as a parameter filter | NeuralMark · CIFAR-10/100 forging attack, ResNet-18 | counterfeit detection rate (%): 100.00 (VanillaMark) -> 48.56 | Improved |
| use hashed watermark as a filter under strong overwriting (lambda 1000) | NeuralMark · CIFAR-100 to CIFAR-10, ResNet-18 | original watermark detection rate (%): 53.90 (VanillaMark) -> 100 | Improved |
| use hashed watermark filter at lr 0.005 overwriting | NeuralMark · CIFAR-100 to CIFAR-10, ResNet-18 | original watermark detection rate (%): 73.04 (VanillaMark) -> 99.60 | Improved |
| increase overwriting learning rate to 0.1 or more | NeuralMark · CIFAR-100 to CIFAR-10, ResNet-18 | detection rate and accuracy: detection 50.78, accuracy 75.2 at lr 0.1 | Hurt |
| add average pooling to filtered parameters | NeuralMark · fine-tuning and pruning attacks | watermark detection rate: 100% under fine-tuning | Improved |
| embed watermark with NeuralMark | NeuralMark · CIFAR-100, 9 architectures | accuracy (%): WideResNet-50 59.67 -> 58.41; ViT-B/16 39.07 -> 39.22 | No clear effect |
| embed watermark with NeuralMark | NeuralMark · E2E text generation | BLEU: GPT-2-S 69.36 -> 69.59; GPT-2-M 68.7 -> 67.73 | Mixed |
| increase number of filtering rounds (ablation) | hashed watermark filter · CIFAR-100 ResNet-18 | parameter overlap ratio | Improved |
| use 6 or 8 filtering rounds instead of 4 (ablation) | NeuralMark · robustness against attacks | detection rate | No clear effect |
| use a plain private filter instead of hashed watermark filter (ablation) | private-filter baseline · forging attack | forging success | Hurt |
Closest papers by meaning
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.89)
- Robust Watermarking for Deep Neural Networks via Bi-Level Optimization (ICCV 2021 · similarity 0.89)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.87)
- Safe and Robust Watermark Injection with a Single OoD Image (ICLR 2024 · similarity 0.87)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · similarity 0.87)
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models (ICML 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.
