Enhanced IoT security: privacy-preserving federated learning model for accurate, real-time intrusion detection across devices

Federated learning (FL) has been proposed as an effective solution in the context of intrusion detection in IoT networks, where models can be trained collaboratively with the security of raw data protection. In this paper we present a privacy-preserving FL framework based on light weight neural network, differential privacy (DP) and homomorphic encryption (HE). With a dataset of 1,191,264 instances and 47 attributes, the proposed model conducted on the IoT Intrusion Detection Dataset available on Kaggle produces overall accuracy (93.5), precision (94.2), recall (93.4), and the F1-score (94.2), with the detection time of 90–130 ms and no distinction between the attacks, where detection latency was considered in this study. At the attack level the model delivered 94.1 %, 92.5 %, and 93.6 % accuracies on DoS, DDoS, and Mirai respectively, and above 85 % accuracy on Malware and Web-based attacks. DP experiments showed that augmenting the privacy budget parameter 0.5 to 20.0 increased the levels of accuracy by 2.6 % to 94.0 %, and decreased the computational time 150 ms to 121 ms, depicting a compromise between privacy and performance. HE experiments likewise exhibited a negligible accuracy reduction (94.1 % to 93.5 %) between no encryption to complete homomorphic encryption, but required more computation time (120 ms to 200 ms). Devices-level testing demonstrated that the model had > 91 % accuracy at the low-end (0.5 GHz CPU, 128 MB memory) and up to 94.5 % accuracy with 110 ms inference time on powerful processors, irrespective of whether or not the sensor was heterogeneous, demonstrating a viable solution to the heterogeneous IT situation. Audit mechanisms further enhanced greater compliance of 0 % to 99 % with minimal reduction in accuracy (< 0.8 %). The results show that privacy-preserving intrusion detection specifically can be performed with real-time intrusion detection, high detection gene, and privacy guarantees in resource-constrained IoT networks.