Real-Time Anomaly Detection Using Federated Learning in Edge Devices

Background: The widespread expansion of IoT devices has significantly increased the scope for malicious activities on the part of cybercriminals. This has made edge computing highly vulnerable to various types of network intrusions, including DDoS floods and brute-force attacks. However, traditional centralized Intrusion Detection Systems (IDS) do not address these concerns due to their inherent latency and lack of consideration of data privacy. To overcome these drawbacks of traditional IDS, we have developed a distributed real-time anomaly detection system architecturally designed for deployment on edge computing hardware. Method: A monitoring agent collects system and network telemetry and real-time network flow data on a Raspberry Pi∼3 and sends it to a centralized server via a TCP socket. The server uses a dual-engine detection system consisting of rule-based heuristics and a DNN model. To address the concerns of data privacy and non-IID data distribution inherent to distributed IoT networks, our system uses a collaboratively trained DNN model using three different Federated Learning (FL) techniques: Clustered Federated Learning (CFL), FedDyn, and FedNova. Results: These techniques were evaluated on the CIC-IoT-DIAD∼2024 dataset with severe non-IID data distribution. The results show that CFL outperforms others with an F1-score of 0.89 and an AUC of 0.95. This work shows that production-grade, privacy-preserving intrusion detection is feasible in IoT networks at the edge.