Unsupervised Anomaly Detection on Attributed Networks With Graph Contrastive Learning for Consumer Electronics Security

The proliferation of consumer electronic products has engendered a substantial surge in data generation and information exchange, concurrently escalating the potential for security threats. Detecting anomalies effectively on attributed networks has undeniable positive significance for consumer electronic security, such as fraudulent user detection, malicious consumption actions analysis, and network threat detection. However, the lack of real tags poses great challenges for detecting anomalies. Therefore, this paper introduces an unsupervised learning framework, namely ADVANCE, to jointly optimize the graph contrastive learning module and the network reconstruction module to accurately discover the anomalies on attributed networks in unsupervised scenarios. Specifically, ADVANCE first constructs the target view and the self-enhanced view to encode view representations by maximizing the consistency between the two views with graph contrastive learning. Subsequently, a network reconstruction module is introduced to assess the anomaly status of each node based on the degree of consistency from both topological structure and node attributes perspectives. The nodes with higher reconstruction errors are considered as anomalous nodes. Finally, the two complementary modules are jointly trained to enhance the accuracy of anomaly detection. Extensive experimental results on three benchmark datasets demonstrate the remarkable effectiveness of our proposed framework in unsupervised anomaly detection.