Revisiting Attack-Caused Structural Distribution Shift in Graph Anomaly Detection

Yuan Gao, Jinghan Li, Xiang Wang, Xiangnan He, Huamin Feng, Yongdong Zhang

Abstract: Graph anomaly detection (GAD) under semi-supervised setting poses a significant challenge due to the distinct structural distribution between anomalous and normal nodes. Specifically, anomalous nodes constitute a minority and exhibit high heterophily and low homophily compared to normal nodes. The distribution of neighbors of the two types of nodes is close, making them difficult to distinguish during aggregation. Furthermore, we discover that apart from various time factors and annotation preferences, graph adversarial attacks can amplify the heterophily difference across training and testing data, namely distribution shift (SDS) in this paper. Current methods for GAD tend to overlook SDS, resulting in poor generalization and limited effectiveness. This work solves the problem from a feature view. We observe that the degree of SDS varies between anomalies and normal nodes. Hence the key lies in (1) resisting high heterophily for anomalies and (2) benefiting the learning of normals from homophily. To this end, we design a Graph Decomposition Network (GDN), which not only teases out the anomaly features that make great contributions to GAD to mitigate the effect of heterophilous neighbors and make them invariant, but also constrain the remaining features for normal nodes to preserve the connectivity of nodes and reinforce the influence of the homophilous neighborhood. To further validate the effectiveness of our method, we illustrate the feature decomposition process in spectral domain, and we also conduct an adversarial attack to incur different heterophily degrees under SDS. Extensive experimental results demonstrate that our framework achieves both accuracy and robustness enhancement.