Loss-Based Decentralized Federated Learning for Robust IoT Intrusion Detection System

This research paper proposes a loss-based decentralized federated learning (FL) that supports collaborative learning mechanisms for Internet of Things (IoT) intrusion detection systems (IDS). Recent works in a similar field mainly focus on developing IDS systems via centralized learning approaches of Artificial Intelligence (AI), which require participants to forward their information. Another solution that has been introduced is utilizing FL with a centralized server, which successfully preserves participant privacy and reduces communication overhead. However, the threat of adopting centralized FL still remains. To mitigate the single point of failure, decentralized FL can be considered. We propose a mechanism that controls the aggregation flow in decentralized FL by introducing a loss-based approach in an independent and identically distributed (IID) environment. This method allows the system to train the model based on participants' losses. By utilizing this, the performance of the decentralized FL system can be improved as the bias from the learning process is reduced. The results show that the proposed loss-based decentralized FL is capable of preserving participant privacy while also improving the overall model performance with an F1-score of 80.04% under the CICIoT2023 dataset. It is worth noting that compared with traditional decentralized FL, the proposed system is able to deliver better performance, up to 7.83%.