Reproduce “Adversarial Attacks and Defenses on Graphs: A Review, A Tool and Empirical Studies” on OpenResearch

About this paper

Wei Jin, Yaxin Li, Han Xu, Yiqi Wang, Shuiwang Ji, Charu Aggarwal, Jiliang Tang

Researchers from Michigan State University, Texas A&M, and IBM reviewed adversarial attacks and defenses on Graph Neural Networks (GNNs). Their work introduces DeepRobust, an open-source tool, and through empirical studies, demonstrates that attacks effectively degrade GNN performance by strategically disrupting graph homophily and local connectivity, while current defenses offer partial but not complete robustness.

Key insights

• Adversarial attacks on graphs primarily operate by disrupting local community structures and graph homophily, often by adding edges between dissimilar nodes and deleting edges between similar nodes.
• Attacks like Metattack can dramatically degrade GNN accuracy (e.g., from over 80% to under 40% on Cora with increasing perturbation) on node classification tasks.
• Existing defense mechanisms offer only partial effectiveness, maintaining higher accuracy than undefended models against minor perturbations but struggling with heavy poisoning.

Summary by alphaXiv.