Topological consistency with low imperceptibility for graph adversarial attacks

Abstract

Recent research shows that graph neural networks (GNNs) are easy to receive disruptions due to the lack of robustness, the phenomenon that poses a serious security threat. Currently, most efforts to attack GNNs mainly use gradient information to guide the attacks. However, the unreliability of gradient information, and the perceptibility of adversarial examples pose challenges that impede further progress in researching on graph adversarial attacks. From the unreliability of gradient information, we propose a Graph Distance Topological Consistency (GDTC). The scheme introduces graph connectivity, geodesic distance, cosine similarity, and Minkowski distance to construct the similarity matrices of the input space and embedding space of the surrogate model. The difference between the two similarities matrices is constrained during the training process of the surrogate model so that the surrogate model fully learns the topology of the original graph. From adversarial examples perceptibility, we propose attack loss with homogeneity restriction. Experiments show that GDTC can study the topological information of the original graph, enhance the reliability of gradient information, and significantly boost attack performance.