Research map / Graph learning / GNN architectures
Randomized Message-Interception Smoothing: Gray-box Certificates for Graph Neural Networks
Yan Scholten, Jan Schuchardt, Simon Geisler, Aleksandar Bojchevski, Stephan Günnemann
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 |
|---|---|---|---|
| apply GDC graph sparsification at test time (ablation) | Message-interception smoothing (GAT) · Cora-ML, second-hop attacks | certified ratio: 15,962 -> 14,606 | Improved |
| use Monte-Carlo samples more efficiently than prior smoothing certificate | Message-interception smoothing · Cora-ML, pd=0.3, pa=0.85 | certification time: 25 -> 17 | Improved |
| add skip-connection that forwards the non-ablated graph without edges (ablation) | Smoothed GNN, node feature ablation · Cora-ML, second-hop attacks | robustness-accuracy tradeoff (AUCRC vs clean accuracy) | Improved |
| ablate 10% less during training than at inference (pt = pe-0.1) (ablation) | Smoothed GAT · Cora-ML, second-hop attacks | robustness-accuracy tradeoff | Improved |
| use attention-based architectures (GAT, GATv2) instead of GCN (ablation) | Smoothed GNN · Cora-ML, second-hop attacks | robustness-accuracy tradeoff | Improved |
| combine edge deletion and feature ablation instead of either alone (ablation) | Smoothed GAT (no skip) · Cora-ML and Citeseer | robustness-accuracy tradeoff | Improved |
| combine edge deletion and feature ablation instead of edge deletion alone (ablation) | Smoothed GAT (no skip) · PubMed | robustness-accuracy tradeoff | Hurt |
| increase confidence level alpha (ablation) | Message-interception smoothing · Cora-ML | certificate strength | No clear effect |
| increase number of features d per node for naive baseline (certify rho = rho~/d) | Bojchevski et al. (2020) naive baseline · Cora-ML | AUCRC | Hurt |
Closest papers by meaning
- Deterministic Certification of Graph Neural Networks against Graph Poisoning Attacks with Arbitrary Perturbations (CVPR 2025 · similarity 0.88)
- Collective Certified Robustness against Graph Injection Attacks (ICML 2024 · similarity 0.87)
- Incremental Randomized Smoothing Certification (ICLR 2024 · similarity 0.87)
- Exact Certification of (Graph) Neural Networks Against Label Poisoning (ICLR 2025 · similarity 0.87)
- A Simple and Yet Fairly Effective Defense for Graph Neural Networks (AAAI 2024 · similarity 0.86)
- Moderate Message Passing Improves Calibration: A Universal Way to Mitigate Confidence Bias in Graph Neural Networks (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.
