Cure-GNN: A Robust Curvature-Enhanced Graph Neural Network Against Adversarial Attacks

Abstract

Graph neural networks (GNNs) are a specialized type of deep learning models on graphs by learning aggregations over neighbor nodes. However, recent studies reveal that the performance of GNNs are severely deteriorated by injecting adversarial examples. Hence, improving the robustness of GNNs is of significant importance. Prior works are devoted to reducing the influence of direct adversaries which are adversarial attacks by positioning a node's one-hop neighbors, yet these approaches are limited in protecting GNNs from indirect adversarial attacks within a node's multi-hop neighbors. In this work, we approach this problem from a new angle by exploring the graph Ricci curvature, which can characterize the relationships of both direct and indirect links from any two nodes’ neighborhoods in the Riemannian space. We first investigate the distinguishable properties of adversarial attacks with graph Ricci curvature distribution. Then, a novel defense framework called Cure-GNN is proposed to detect and mitigate adversarial effects. Cure-GNN discerns the distinction between adversarial edges and normal edges via computing curvature, and merges it into the node features reconstructed by a residual learning framework. Extensive experiments over real-world datasets on node classification task demonstrate the efficacy of Cure-GNN and achieves superiority to the state-of-the-arts without incurring high complexity.