Contrastive Federated Learning for Graph Anomaly Detection.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Graph anomaly detection (GAD) faces challenges with imbalanced data and privacy concerns.
• Existing GAD models struggle with optimizing node embeddings and detecting multiple anomaly types simultaneously, leading to reduced accuracy.

Purpose of the Study:

• To introduce FedGAD, a novel federated learning model for graph anomaly detection (GAD).
• To address limitations in existing GAD approaches concerning imbalanced data and privacy protection.

Main Methods:

• FedGAD employs collaborative unsupervised learning across decentralized data centers without direct subgraph access.
• It enhances node representations by masking and reconstructing neighborhood features.
• A cross-client node representation module facilitates neighbor reconstruction using information from other clients.
• Multiscale contrastive learning (structure-level and contextual-level) is utilized for anomaly detection.

Main Results:

• FedGAD demonstrated superior performance compared to baseline methods on seven benchmark datasets.
• The model effectively improves GAD performance, particularly in scenarios with imbalanced data distributions across clients.

Conclusions:

• FedGAD offers an effective solution for graph anomaly detection in decentralized environments.
• The model successfully overcomes challenges related to imbalanced data and privacy, enhancing GAD accuracy and robustness.