Research map / Adversarial robustness and security / Attacks on text and graphs
Adversarial Attack on Graph Structured Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 RL-S2V edge-modification attacks | structure2vec graph classifier · Synthetic graph classification, test set I | accuracy: 93.20% -> 42.93% (15-20 nodes, K=2); 94.60% -> 66.07% (90-100 nodes, K=2) | Hurt |
| Apply RL-S2V policy learned on test set I without querying the classifier | structure2vec graph classifier · Synthetic graph classification, test set II | accuracy: 94.67% -> 44.00% (15-20 nodes, K=2); 96.67% -> 62.67% (90-100 nodes, K=2) | Hurt |
| Use GeneticAlg instead of RandSampling | structure2vec graph classifier · Synthetic graph classification, test set I, 15-20 nodes, K=2 | accuracy: 78.73% -> 39.87% | Hurt |
| Use GradArgmax edge selection | structure2vec graph classifier · Synthetic graph classification, test set I | accuracy: K=2, 15-20 nodes: 93.20% -> 69.47%; K=4, 15-20 nodes: 98.87% -> 95.80% | Mixed |
| Apply RL-S2V edge-modification attacks | GCN node classifier · Citeseer, Cora, Pubmed, Finance node classification, test set I | accuracy: 71.60% -> 62.70%; 81.00% -> 71.20%; 79.90% -> 72.80%; 88.67% -> 85.43% | Hurt |
| Apply RL-S2V policy without access to the target classifier | GCN node classifier · Citeseer, Cora, Pubmed, Finance node classification, test set II | accuracy: 72.60% -> 66.00%; 80.20% -> 75.00%; 80.40% -> 74.00%; 91.88% -> 89.10% | Hurt |
| Use GradArgmax rather than RL-S2V | GCN node classifier · Pubmed node classification, test set I | accuracy: 79.90% -> 72.4% (GradArgmax), compared with 72.80% (RL-S2V) | Hurt |
| Randomly drop edges during training | GCN node classifier · Citeseer, Cora, Pubmed, Finance node classification | accuracy: Unattacked: 71.30%, 81.70%, 79.50%, 88.55%; RL-S2V attack: 63.90%, 72.80%, 72.90%, 85.80% | Mixed |
| Increase the propagation steps K | structure2vec graph classifier · Synthetic graph classification, unattacked test set I | accuracy: 15-20 nodes: 93.20% (K=2) -> 99.07% (K=5); 40-50 nodes: 92.60% (K=2) -> 97.93% (K=5); 90-100 nodes: 94.60% (K=2) -> 98.2 | Mixed |
Closest papers by meaning
- Value at Adversarial Risk: A Graph Defense Strategy against Cost-Aware Attacks (AAAI 2024 · similarity 0.91)
- Revisiting Graph Adversarial Attack and Defense From a Data Distribution Perspective (ICLR 2023 · similarity 0.91)
- Adversarial Attacks on Graph Classifiers via Bayesian Optimisation (NeurIPS 2021 · similarity 0.90)
- Highly Imperceptible Black-Box Graph Injection Attacks with Reinforcement Learning (AAAI 2025 · similarity 0.89)
- Adversarial Robust Generalization of Graph Neural Networks (ICML 2025 · similarity 0.87)
- A Simple and Yet Fairly Effective Defense for Graph Neural Networks (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.