Large-Scale Adversarial Attacks on Graph Neural Networks via Graph Coarsening
Graph Neural Networks (GNNs) are fragile to adversarial attacks. However, existing state-of-the-art adversarial attack methods against GNNs are typically constrained by the graph's scale, failing to attack large graphs effectively. In this paper, we propose a novel attack method that attacks the graph in a divide-and-conquer manner to tackle large-scale adversarial attacks on GNNs. Specifically, the nodes are clustered based on node embeddings, coarsened graphs are constructed using the node clusters, and attacks are conducted on the coarsened graphs. Perturbations are selected starting with smaller coarsened graphs and progressing to larger detailed graphs while most of the irrelative nodes remain clustered, significantly reducing the complexity of generating adversarial graphs. Extensive empirical results show that the proposed method can greatly save the computational resources required to attack GNNs on large graphs while maintaining comparable performance on small graphs.
PDF Abstract
Cora
