Spiderweb Protector: Biomimetic Defense Against Targeted Attacks on Graph Neural Networks

Graph Neural Networks (GNNs) have emerged as powerful tools across diverse applications, including social networks, bioinformatics, and recommendation systems. However, their susceptibility to adversarial attacks, particularly targeted attacks, poses a significant challenge. Targeted attacks exploit the message-passing mechanism of GNNs to induce severe performance degradation, even with minimal perturbations. Inspired by the self-repair and adaptive behaviors of spider webs, this paper introduces SpiderWeb Protector, a biomimetic defense strategy designed to enhance GNN robustness. The proposed method integrates three core components: Community-Centric Graph Update, which selectively reinforces key graph structures; Topology Feature Merging, which improves feature representation by incorporating structural insights; and Edge Pruning, which dynamically removes adversarial connections. Comprehensive experiments on three datasets (Cora, Citeseer and Cora_ml) and two GNN models (GCN and GIN) demonstrate the superior performance of SpiderWeb Protector against both direct and influence-targeted attacks, achieving state-of-the-art results.