Enhanced federated anomaly detection through autoencoders using summary statistics-based thresholding.

Area of Science:

• Machine Learning
• Data Science
• Cybersecurity

Background:

• Federated Learning (FL) presents unique challenges for Anomaly Detection (AD) due to decentralized data and Non-Independent and Identically Distributed (Non-IID) distributions.
• Existing methods struggle with privacy and accuracy in decentralized AD settings, particularly with heterogeneous data.

Purpose of the Study:

• To propose a novel federated threshold calculation method for Anomaly Detection in Federated Learning.
• To enhance the accuracy and robustness of Anomaly Detection under Non-IID data distributions while preserving data privacy.

Main Methods:

• A federated threshold calculation method is proposed, aggregating summary statistics of normal and anomalous data from clients.
• This global threshold is utilized with federated Autoencoders for Anomaly Detection.
• The method was evaluated on diverse datasets including Credit Card Fraud Detection, Shuttle, and Covertype.

Main Results:

• The proposed federated threshold calculation method consistently outperformed existing federated and local thresholding techniques.
• Demonstrated significant improvements in Anomaly Detection accuracy and robustness in Non-IID settings.
• Validated the effectiveness of using aggregated summary statistics for privacy-preserving federated Anomaly Detection.

Conclusions:

• Summary statistics aggregation is a viable strategy for improving federated Anomaly Detection performance under Non-IID conditions.
• The proposed method offers a privacy-preserving and effective solution for decentralized Anomaly Detection.
• This research paves the way for more robust and accurate anomaly detection systems in federated environments.