A General Unified Graph Neural Network Framework Against Adversarial Attacks

Graph Neural Networks (GNNs) are powerful tools in representation learning for graphs. However, they are reported to be vulnerable to adversarial attacks, raising numerous concerns for applying it in some risk-sensitive domains. Therefore, it is essential to develop a robust GNN model to defend against adversarial attacks. Existing studies address this issue only considering cleaning perturbed graph structure, and almost none of them simultaneously consider denoising features. As the graph and features are interrelated and influence each other, we propose a General Unified Graph Neural Network (GUGNN) framework to jointly clean the graph and denoise features of data. On this basis, we further extend it by introducing two operations and develop a robust GNN model(R-GUGNN) to defend against adversarial attacks. One operation is reconstructing the graph with its intrinsic properties, including similarity of two adjacent nodes’ features, sparsity of real-world graphs and many slight noises having small eigenvalues in perturbed graphs. The other is the convolution operation for features to find the optimal solution adopting the Laplacian smoothness and the prior knowledge that nodes with many neighbors are difficult to attack. Experiments on four real-world datasets demonstrate that R-GUGNN has greatly improved the overall robustness over the state-of-the-art baselines.