Adversarial Robustness of Graph Neural Networks Under Feature Distribution Shift

Graph Neural Networks (GNNs) have demonstrated remarkable performance across various node- and graph-level prediction tasks. However, their sensitivity to adversarial attacks and shifts in the input data distribution remains a significant concern, particularly in real-world deployment scenarios. This paper investigates the adversarial robustness of GNNs under feature distribution shift. We analyze the impact of subtle, targeted perturbations to node features on the prediction accuracy of GNNs trained on a source distribution and evaluated on a shifted target distribution. Our empirical results demonstrate a significant degradation in performance, highlighting the vulnerability of GNNs to such shifts. We further explore potential mitigation strategies, including adversarial training and distributionally robust optimization, to enhance the robustness of GNNs in the presence of feature distribution shift.