Conformalized Adversarial Attack Detection for Graph Neural Networks

Graph Neural Networks (GNNs) have achieved remarkable performance on diverse graph representation learning tasks. However, recent studies have unveiled their susceptibility to adversarial attacks, leading to the development of various defense techniques to enhance their robustness. In this work, instead of improving the robustness, we propose a framework to detect adversarial attacks and provide an adversarial certainty score in the prediction. Our framework evaluates whether an input graph significantly deviates from the original data and provides a well-calibrated p-value based on this score through the conformal paradigm, therby controlling the false alarm rate. We demonstrate the effectiveness of our approach on various benchmark datasets. Although we focus on graph classification, the proposed framework can be readily adapted for other graph-related tasks, such as node classification.