Employing Bayesian Inference Models to Bolster The Robustness of Graph Neural Networks

Abstract

Graph Neural Networks (GNNs) have become critical in the realm of node classification tasks. Nevertheless, they exhibit significant vulnerabilities to adversarial perturbations, such as adversarial attacks. Traditional approaches attempt to address this issue but have various shortcomings. For example, Bayesian approaches may suffer slow convergence during inference. To solve this issue, in this study, we leverage Bayesian methods to enhance the robustness of GNNs. Specifically, we propose a novel framework, named RobustGraph, that integrates Bayesian methods to defend GNNs on perturbed graphs. Our empirical results demonstrate that our framework can substantially outperform competing models in classification tasks.