Federated Learning with Kalman Filter for Intrusion Detection in IoT Environment

Abstract

Enhancing the IoT (Internet of Things) network for reliability prompted a heightened focus on device security, given their diverse characteristics and sensitive data exchange. Federated Learning (FL) gained attention for its collaborative approach, sharing only local models, not data, among devices in IoT networks. Additionally, the non-i.i.d. data distribution complexity adds another layer to the network's challenges. In this article, we propose FedKF as a model approach for a federated learning algorithm with KF (Kalman Filter). This approach improves the performance of the IDS (Intrusion Detection System), especially for IoT data. In this model, each edge client trains the data locally to form a local model, which is then aggregated on a central server to create a global model. The FedKF aggregation algorithm employs a KF to predict and estimate the aggregation weight, where the prediction is based on the current measured weight of aggregated global models and the previous model weight. Furthermore, by selecting clients and allowing only selected devices to participate in the training process, the overall energy consumption can be reduced. Therefore, it's essential to balance energy savings with the performance of the federated learning model, ensuring that the model remains accurate and effective despite reduced participation. The experimental results demonstrate the model's performance across different IoT datasets, compare the results with the average FL model, and verify the noticeable improvement in accuracy and communication loss. The model also shows the effect of client selection on the model's performance.