Cross-Domain Graph Anomaly Detection.

Area of Science:

• Graph Machine Learning
• Cybersecurity Analytics
• Data Mining

Background:

• Anomaly detection on attributed graphs is crucial for cybersecurity, finance, and healthcare.
• Unsupervised methods dominate due to the high cost of labeled data, especially in new domains.
• Leveraging labeled data from one domain to detect anomalies in another (cross-domain) is under-explored.

Purpose of the Study:

• To address the challenge of cross-domain graph anomaly detection using domain adaptation.
• To develop a framework capable of handling data heterogeneity in attributed graphs (structure and attributes).
• To effectively capture both domain-invariant and domain-specific anomalies.

Main Methods:

• Proposed a novel framework named COMMANDER.
• Employed a graph attentive encoder to compress attributed graphs into a low-dimensional space.
• Utilized a domain discriminator and an anomaly classifier for cross-domain anomaly detection.
• Developed an attribute decoder to identify target-specific anomalies.

Main Results:

• COMMANDER demonstrates efficacy in cross-domain anomaly detection.
• The framework successfully handles heterogeneity in graph topology and node attributes.
• Experiments on real-world datasets validate the proposed approach.

Conclusions:

• COMMANDER provides an effective solution for cross-domain attributed graph anomaly detection.
• The framework successfully adapts knowledge from labeled domains to unlabeled ones.
• This work advances anomaly detection capabilities in heterogeneous graph data.