Federated Graph Anomaly Detection via Contrastive Self-Supervised Learning.

Area of Science:

• Graph machine learning
• Data mining
• Network security

Background:

• Attribute graph anomaly detection is crucial for identifying outliers in complex graph data.
• Centralized methods pose privacy risks, while federated learning offers a privacy-preserving alternative.
• Distributed graph data limitations hinder direct application of federated learning for anomaly detection.

Purpose of the Study:

• To propose a federated graph anomaly detection framework (FedCAD) using contrastive self-supervised learning (CSSL).
• To address privacy concerns and data limitations in distributed graph anomaly detection.
• To enhance the efficacy and precision of anomaly detection in federated settings.

Main Methods:

• Federated graph anomaly detection framework (FedCAD) utilizing contrastive self-supervised learning (CSSL).
• Pseudo-label discovery for preliminary anomaly node identification.
• Local anomaly neighbor embedding aggregation strategy for enhanced distinction.

Main Results:

• FedCAD effectively updates anomaly node information across clients via federated learning (FL).
• The aggregation strategy amplifies distinctions between anomaly and neighbor nodes.
• Experimental results on four real graph datasets demonstrate FedCAD's efficiency.

Conclusions:

• FedCAD offers a privacy-preserving and effective solution for distributed graph anomaly detection.
• Contrastive learning, enhanced by neighbor embedding aggregation, improves anomaly detection efficacy.
• The proposed framework overcomes limitations of limited client data in federated settings.