Cross-layer Federated Heterogeneous Ensemble Learning for Lightweight IoT Intrusion Detection System

This paper presents a heterogeneous federated ensemble model for intrusion detection system, employing a semisupervised novelty detection technique - the baseline K-means. The technique learns normal traffic from baseline data and utilizes the Mahalanobis distance to detect anomalous packets. To mitigate the false-positive rate inherent in anomaly-based intrusion detection system, we propose an ensemble approach that integrates local novelty detection models dedicated to each worker in both weighed and voting-based strategies. The federated design augments each worker’s detection capability without increasing the false positive rate. Our extensive experiments showcase the system’s robustness and adaptability over traditional standalone IDS, with marked improvements in precision, recall, and F1score under varying sampling rates. We made this project’s code publicly available on Github for replicability.