Research map / Graph learning / GNN architectures
GraphTextack: A Realistic Black-Box Node Injection Attack on LLM-Enhanced GNNs
Jiaji Ma, Puja Trivedi, Danai Koutra
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 |
|---|---|---|---|
| inject nodes with jointly optimized structure and text via evolutionary search (GraphTextack) rather than best uni-modal text attack | One-for-all · Cora, r=0.03 | post-attack accuracy: 74.36 -> 65.75 | Improved |
| GraphTextack vs best prior baselines across budgets | One-for-all · ogbn-products, r=0.01 | post-attack accuracy: 78.99 vs GANI 78.93 | No clear effect |
| GraphTextack vs WTGIA on WikiCS at r=0.05 | One-for-all · WikiCS, r=0.05 | post-attack accuracy: 61.35 vs 60.95 | No clear effect |
| add crossover to the evolutionary search (ablation) | GraphTextack · WikiCS, r=0.05 | post-attack accuracy: 64.42 -> 61.35 | Improved |
| add mutation to the evolutionary search (ablation) | GraphTextack · WikiCS, r=0.05 | post-attack accuracy: 65.19 -> 61.35 | Improved |
| add local prediction shift term to fitness (ablation) | GraphTextack · WikiCS, r=0.05 | post-attack accuracy: 66.83 -> 61.35 | Improved |
| add PageRank influence term to fitness (ablation) | GraphTextack · WikiCS, r=0.05 | post-attack accuracy: 61.97 -> 61.35 | No clear effect |
| add PageRank influence term to fitness (ablation) | GraphTextack · WikiCS, r=0.01 | post-attack accuracy: 71.92 -> 71.69 | No clear effect |
| query target model directly instead of surrogate models | GraphTextack · runtime per injection (ogbn-arxiv, ogbn-products) | runtime | Improved |
| uni-modal attacks (text-only) on LLM-enhanced GNNs | LLM-enhanced GNNs · node classification | accuracy drop: less than 5% drop | Hurt |
Closest papers by meaning
- Intruding with Words: Towards Understanding Graph Injection Attacks at the Text Level (NeurIPS 2024 · similarity 0.91)
- Highly Imperceptible Black-Box Graph Injection Attacks with Reinforcement Learning (AAAI 2025 · similarity 0.88)
- LLM Enhancers for GNNs: An Analysis from the Perspective of Causal Mechanism Identification (ICML 2025 · similarity 0.87)
- Large Language Model Meets Graph Neural Network in Knowledge Distillation (AAAI 2025 · similarity 0.87)
- Label-free Node Classification on Graphs with Large Language Models (LLMs) (ICLR 2024 · similarity 0.87)
- Stronger and Transferable Node Injection Attacks (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.
