Federated Learning for Privacy-Preserving Cyber Threat Detection in Distributed IoT Networks

Abstract

This study examines federated learning as a privacy-conscious approach to cyber-threat detection in distributed Internet of Things (IoT) environments, where centralized processing may expose sensitive network information. The study aims to establish a structured federated-learning framework for cyber-threat detection and to examine detection effectiveness using standard classification measures while considering the influence of aggregation strategies and privacy preservation. The analysis draws on published benchmark evaluations using the ToN_IoT cybersecurity dataset. Reported results indicate that federated learning can maintain consistent cyber-threat detection performance across distributed clients, with the aggregated federated model achieving competitive performance compared with centralized learning. The comparison of FedAvg, FedAvgM, FedAdam, and FedAdagrad further demonstrates that the choice of aggregation mechanism can substantially influence classification performance. Published privacy-preserving federated-learning results also indicate that collaborative threat detection can be supported within a distributed data architecture while avoiding direct centralization of underlying network observations. These findings highlight the potential of federated learning for privacy-conscious cybersecurity in distributed IoT networks. However, broader evaluation involving heterogeneous clients, non-IID data distributions, resource-constrained devices, dynamically changing attack patterns, and stronger privacy mechanisms is required. Future validation across independent IoT cybersecurity datasets and real-world deployment environments would further strengthen evidence regarding the robustness, scalability, and generalizability of federated cyber-threat detection.