A lightweight Federated Learning-based Intrusion detection model for Internet of Things

The advancement of sophistication in botnet intrusions on Internet of Things (IoT) systems and the resources required by traditional IDS make strong security solutions necessary, which reduce resource usage while enhancing detection capability. Therefore, this paper tests the performance and performance comparison of classification models based on lightweight Federated Learning (FL) for IoT botnet intrusion detection using Random Forest (RF), XGBoost, and LightGBM classifiers. Utilizing a large-scale dataset from IoT traffic, the three models were tested in metrics such as confusion matrices, ROC curves, and Precision-Recall curves. The results show that the RF and XGBoost classifiers achieve high AUCs of 1.00, thereby, there is no trade-off between sensitivity and specificity. On the other hand, LightGBM scored a low of 0.58 AUC. The results show that robust ensemble methods can be applied to handle complex and imbalanced datasets commonly presented in IoT traffic in FL environments.