Temporal Weighted Averaging for Asynchronous Federated Intrusion Detection Systems.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Cybersecurity

Background:

• Federated learning (FL) enables distributed, privacy-preserving machine learning.
• Traditional FL faces deadlocks due to synchronous model averaging, especially with low-power IoT devices.
• Efficient global model aggregation is crucial for FL performance.

Purpose of the Study:

• To propose a novel temporal model averaging algorithm for asynchronous federated learning (AFL).
• To address deadlocks in FL systems and improve server/client throughput.
• To evaluate the algorithm's effectiveness in cybersecurity anomaly detection.

Main Methods:

• Developed a dynamic expectation function to estimate expected client models per round.
• Implemented a weighted averaging algorithm for continuous global model updates.
• Tested the AFL algorithm on the NSL-KDD intrusion detection dataset.

Main Results:

• Achieved a global model accuracy of 99.5% in anomaly detection.
• Demonstrated a throughput increase of approximately 10.17% every 30 timesteps.
• Outperformed traditional federated learning models in cybersecurity applications.

Conclusions:

• The proposed asynchronous federated learning algorithm effectively prevents system deadlocks.
• The algorithm enhances overall system throughput and maintains high accuracy in intrusion detection.
• Asynchronicity in FL is a viable strategy for improving cybersecurity performance.