An Evaluation of a Tree-based Federated Learning Approach for Internet of Things Anomaly Detection

The Internet of Things (IoT) has gained significant popularity in recent years, resulting in an increased focus from malicious individuals. This is particularly concerning as these networks tend to utilise sensitive data, resulting in the need for effective methods that detect and classify attacks. However, significant challenges must be overcome, including the limited computational resources available within IoT networks. To this end, this work evaluates the effectiveness of a lightweight approach based on Federated XGBoost in detecting and classifying attacks on IoT networks. To evaluate this approach, two datasets, Edge-IIoTset and NSL-KDD, are considered, along with their pre-defined feature set. Additionally, we explore a custom feature set of 30 features for Edge-IIoTset. Using these datasets, 10-fold stratified cross-validation is employed. From the evaluation, it was identified that our approach is generally effective when considering established datasets, including outperforming neural network-based FL approaches in terms of Accuracy for the attack classification problem. However, improvements are needed to ensure that adequate performance is achieved when considering recent datasets.