Reproduce “Unveiling the Vulnerability of Graph-LLMs: An Interpretable Multi-Dimensional Adversarial Attack on TAGs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bowen Fan, Zhilin Guo, Xunkai Li, Yihan Zhou, Bing Zhou, Zhenjun Li, Rong-Hua Li, Guoren Wang
IMDGA attacks text-attributed graph node classifiers by coordinating word substitutions with graph-edge pruning. It uses Topological SHAP to identify influential words, tests contextually plausible replacements across a target and its neighbors, and prunes influential links only when text edits fail. On BERT node classification on Cora, it reports 94.42 ± 0.87% attack success rate, a local untargeted measure on tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.