Research map / Adversarial robustness and security / Attacks on text and graphs
Contrastive Adversarial Training for Robust Graph Neural Networks under Label Poisoning
Manshika C Bissessur, Melis Ilayda Bal, Michael Muehlebach
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a contrastive adversarial training defense for graph neural networks against label poisoning attacks on training labels.
Problem
GNNs are vulnerable to label poisoning, where a few corrupted training labels propagate errors across the graph through message passing; existing defenses target label noise and rely on robust losses or heuristic cleaning.
Approach
Combines contrastive learning with adversarial training to make GNNs robust to label poisoning; the abstract is truncated so the specific mechanism beyond this framing is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace the contrastive detection model with a logit/loss-based node selection heuristic (ablation) | CoLAT · Cora ML, 30% NTK poisoning | test accuracy gain over clean training: 0.4% gain vs 13% with detection model | Hurt |
| Replace the contrastive detection model with a logit-based heuristic (ablation) | CoLAT · Cora ML, SMulti | test accuracy gain over clean training: about 5% vs up to 31% | Hurt |
| Use adversarial iterative training instead of one-step label filtering with the detection model (ablation) | CoLAT · Cora ML, 20% NTK | test accuracy gain over clean training: about 13% vs roughly 1% for filtering | Improved |
| Use adversarial iterative training instead of one-step label filtering (ablation) | CoLAT · Cora ML, 20% SMulti | test accuracy gain over clean training: around 26% vs about 8% | Improved |
| Select the most inconsistent nodes rather than the most consistent nodes (ablation) | CoLAT · Cora ML, NTK and SMulti | test accuracy: 0.4978 -> 0.6689 (NTK 20%) | Improved |
| Apply CoLAT with GraphSAGE backbone | CoLAT · Cora ML, SMulti 30% | test accuracy gain: up to 31% gain at 30% poisoning | Improved |
| Update the detection model less often (interval 1, 5, 10, 20) (ablation) | CoLAT · OGBN-ArXiv | test accuracy and training time | Improved |
| Apply CoLAT on clean labels | GCN · Cora ML, clean labels | test accuracy: 0.8270 -> 0.8377 | Improved |
| Apply CoLAT on a heterophilic graph | GCN · Actor, NTK low poisoning | test accuracy | No clear effect |
| Apply CoLAT on a heterophilic graph under SMulti | GCN · Actor, SMulti 30% | test accuracy: 0.1079 -> 0.1444 | Improved |
| Apply CoLAT on Citeseer under SMulti 20% | GCN · Citeseer, SMulti 20% | test accuracy: 0.6167 -> 0.6200 | No clear effect |
Closest papers by meaning
- Rethinking Label Poisoning for GNNs: Pitfalls and Attacks (ICLR 2024 · similarity 0.90)
- Exact Certification of (Graph) Neural Networks Against Label Poisoning (ICLR 2025 · similarity 0.89)
- Refine then Classify: Robust Graph Neural Networks with Reliable Neighborhood Contrastive Refinement (AAAI 2025 · similarity 0.88)
- Adversarial Contrastive Graph Masked AutoEncoder Against Graph Structure and Feature Dual Attacks (AAAI 2025 · similarity 0.87)
- Clean-Label Graph Backdoor Attack in the Node Classification Task (AAAI 2025 · similarity 0.87)
- Bounding the Expected Robustness of Graph Neural Networks Subject to Node Feature Attacks (ICLR 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.
