Research map / Graph learning / GNN architectures
Watermarking Graph Neural Networks via Explanations for Ownership Protection
Jane Downer, Yingdan Shi, Ziyan Liu, Ren Wang, Binghui Wang
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 |
|---|---|---|---|
| embed explanation watermark via dual-objective loss | GraphSAGE · Photo, PubMed, CS test accuracy | test accuracy: Photo 88.2 -> 90.8; PubMed 81.2 -> 86.0; CS 88.9 -> 88.4 | No clear effect |
| embed explanation watermark in Graph Transformer on Reddit2 | Graph Transformer · Reddit2 test accuracy | test accuracy: 80.4 -> 79.4 | Hurt |
| embed explanation watermark (T=4 subgraphs) | GNNs on Photo, PubMed, CS, Reddit2 · ownership verification | MI p-value: <0.001 | Improved |
| increase number of watermarked subgraphs T (ablation) | explanation watermark · Photo, PubMed, CS | alignment and p-value | Improved |
| increase watermarked subgraph size s (ablation) | explanation watermark · Photo, PubMed, CS | alignment, p-value, robustness | Improved |
| apply structured pruning attack | GraphSAGE watermarked model · Photo, PubMed, CS, Reddit2 | MI p-value | No clear effect |
| fine-tune at 0.1x learning rate for 49 epochs | GraphSAGE watermarked model · Photo, PubMed, CS, Reddit2 | MI p-value | No clear effect |
| fine-tune at 1x or 10x learning rate | GraphSAGE watermarked model · fine-tuning attack | watermark presence | Hurt |
| model merging of watermarked and fine-tuned models | watermarked GCN/SAGE/Transformer · merged model verification | MI p-value: GCN 9.633e-7; SAGE 0.007; Transformer 0.008 | No clear effect |
| knowledge distillation attack without defense | watermarked GNN · removal attack | MI p-value: p > 0.05 | Hurt |
| use GCN or SGC instead of GraphSAGE (ablation) | watermarked model under fine-tuning and pruning · Photo, PubMed, CS | robustness | Mixed |
| add watermarking to training | GCN/SAGE training · training time | wall time (s): Photo SAGE 109.06 -> 167.28 | Hurt |
Closest papers by meaning
- GraphIP–Bench: How Hard Is It to Steal a Graph Neural Network, and Can We Stop It? (NeurIPS 2026 · similarity 0.89)
- ZeroMark: Towards Dataset Ownership Verification without Disclosing Watermark (NeurIPS 2024 · similarity 0.88)
- EntropyMark: Towards More Harmless Backdoor Watermark via Entropy-based Constraint for Open-source Dataset Copyright Protection (CVPR 2025 · similarity 0.87)
- FedGMark: Certifiably Robust Watermarking for Federated Graph Learning (NeurIPS 2024 · similarity 0.86)
- Can Graph Neural Networks Expose Training Data Properties? An Efficient Risk Assessment Approach (NeurIPS 2024 · similarity 0.86)
- Revisiting the Information Capacity of Neural Network Watermarks: Upper Bound Estimation and Beyond (AAAI 2024 · 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.
