Structure-Aware Attack on Graph Neural Networks via Imperceptible Node Injection

Graph neural networks (GNNs) excel in various graph-based tasks due to their exceptional ability to process non-Euclidean data. However, recent research indicates that their predictive performance is highly vulnerable to perturbations from intentionally manipulated data. Current node injection attack methods disrupt GNN training by injecting numerous nodes, often in excessive amounts, making them easily detectable. To address this issue, this study introduces a structure-aware node injection attack (SNIA), which enables effective and subtle attacks under extreme budget constraints. The scheme leverages the network topology to construct an attack candidate set and applies homogeneity constraints to regulate the generation of perturbed features. By eliminating the dependency on surrogate models for generating perturbed data, SNIA effectively diminishes the global classification performance of GNNs based on the network's inherent structure. We conducted attack experiments with the SNIA scheme on real-world network datasets against both general and defensive GNNs. The experimental findings reveal that it significantly surpasses the existing state-of-the-art methods in attack efficiency, while also showcasing outstanding generalization and resilience.