Attributed Network Anomaly Detection via Feature Masking and Anomaly Enhancement

Abstract

Anomaly detection in attributed networks has extremely important application value, and graph neural networks (GNNs) based methods for it have been extensively proposed. However, the presence of anomalous nodes breaks the homophily assumption of GNNs, causing two unresolved problems: anomaly contamination and low anomaly discriminability, which often lead to suboptimal anomaly detection performance. In view of this, we propose a reconstruction-based attributed network anomaly detection method via feature masking and anomaly enhancement (FMAE). Specifically, to adaptively enhance anomaly discriminability, FMAE introduces dual-channel spectral filtering and an attention-based feature fusion mechanism. For anomaly contamination, FMAE designs a fine-grained masking module, which can automatically suppress the feature propagation of anomalous nodes. To further enhance the detection capability, FMAE utilizes a structure-aware module to measure the structural consistency of each node. Meanwhile, FMAE devises a reconstruction loss, which can alleviate the problem that abnormal nodes easily dominate the model optimization process. Compared with state-of-the-art baselines, extensive experiments conducted on benchmark datasets demonstrate the superiority of FMAE.