Research map / Adversarial robustness and security / Attacks on text and graphs
Robust Watermarking for Deep Neural Networks via Bi-Level Optimization
Peng Yang, Yingjie Lao, Ping Li
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 |
|---|---|---|---|
| increase number of embedded keys from 5 to 20 on ImageNet (ablation) | Proposed bi-level watermarking · ImageNet (Inception-V3) | Rauth: 100% -> 97.5% | Hurt |
| increase number of embedded keys from 5 to 20 on ImageNet (ablation) | Proposed bi-level watermarking · ImageNet (Inception-V3) | Rloss: 0.07% -> 0.15% | Hurt |
| increase number of embedded keys from 5 to 20 on MNIST (ablation) | Proposed bi-level watermarking · MNIST (LeNet5) | Rloss: 0.00% -> 0.07% | Hurt |
| use bi-level robust watermarking instead of white-box method [9] | DNN watermarking on CIFAR10 · CIFAR10 overwriting, 20 keys | Rloss: ∼0.5% -> ∼0.05% | Improved |
| use bi-level robust watermarking instead of backdoor black-box method [1] | DNN watermarking on CIFAR10 · CIFAR10 overwriting, 20 keys | Rpres: 95% -> 100% | Improved |
| use bi-level robust watermarking instead of white-box method [9] | DNN watermarking on CIFAR10 · CIFAR10 overwriting, 20 keys | Rpres: 58% -> 100% | Improved |
| apply pruning up to 50% rate on the watermarked model | Proposed bi-level watermarking · CIFAR10 pruning | identification accuracy: 50% | No clear effect |
| embed keys one by one instead of all at once under sequential overwriting | Proposed bi-level watermarking · CIFAR10/100 and ImageNet overwriting | signature preserve rate: 95% and 100% | Mixed |
| embed 60 keys instead of fewer keys (capacity test) | Proposed bi-level watermarking · MNIST, CIFAR10, CIFAR100 capacity | authentication rate / function preserve rate: more than 94% auth, around 99.0% function preserve at 60 keys | Hurt |
Closest papers by meaning
- Margin-based Neural Network Watermarking (ICML 2023 · similarity 0.90)
- Deep Model Intellectual Property Protection via Deep Watermarking (TPAMI 2021 · similarity 0.89)
- Lightweight-Mark: Rethinking Deep Learning-Based Watermarking (ICML 2025 · similarity 0.89)
- Safe and Robust Watermark Injection with a Single OoD Image (ICLR 2024 · similarity 0.88)
- END^2: Robust Dual-Decoder Watermarking Framework Against Non-Differentiable Distortions (AAAI 2025 · similarity 0.87)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · similarity 0.87)
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.
