FADngs: Federated Learning for Anomaly Detection

Area of Science:

• Machine Learning
• Data Privacy
• Cybersecurity

Background:

• Federated learning (FL) is increasingly vital for privacy-preserving data analysis.
• Existing FL methods primarily address classification, neglecting privacy-preserving anomaly detection.
• Traditional anomaly detection algorithms face challenges in FL settings, including detection inaccuracies and performance degradation due to local data distribution variations.

Purpose of the Study:

• To develop a novel federated anomaly detection method that addresses privacy concerns and performance limitations.
• To enable effective anomaly detection in decentralized environments without compromising data privacy.
• To improve the global model's ability to detect anomalies deviating from local data distributions.

Main Methods:

• Federated Anomaly Detection with Noisy Global Density Estimation, and Self-supervised Ensemble Distillation (FADngs) is proposed.
• Clients share processed density functions to align data distribution knowledge.
• Local models are trained using contrastive learning enhanced by shared density functions.
• Ensemble distillation is employed to aggregate knowledge from diverse local models into a global model.

Main Results:

• FADngs significantly outperforms existing state-of-the-art federated anomaly detection methods.
• The proposed method demonstrates effective anomaly detection capabilities while preserving data privacy.
• Empirical evidence confirms the privacy-preserving nature of the shared density functions.

Conclusions:

• FADngs offers a robust solution for privacy-preserving federated anomaly detection.
• The method successfully integrates density function sharing, contrastive learning, and ensemble distillation for enhanced performance.
• The approach maintains local model specificity while building a capable global anomaly detection model.