Enhancing Intrusion Detection: Leveraging Federated Learning and Hybrid Deep Learning Models

Abstract

An efficient and productive IoT application may ease some real-time tasks however, it is at risk of cyber-attack. Intrusion Detection Systems (IDS) are of significant importance for the security measures of IoT applications. Anomaly-based intrusion detection systems perform more efficaciously than other methods. IoT/IIoT devices that deal with large data volumes are at risk of malicious attacks and as a result, anomaly-based IDS are developed. But, the question that arises is whether the performance of models meets the required standards and accuracy. For research the Telemetry data of IoT/IIoT services from the ToN_IoT dataset collected at UNSW Canberra Cyber IoT lab, SEIT (Australia), is used. It includes data about seven IoT/IIoT sensors. Federated Learning based on Deep auto-encoder is adopted to efficiently identifying attacks while solvingthe issue of data leakage and the privacy of users . Federated models handle the non-IID data efficiently. Hybrid models use Machines and Deep Learning algorithms for efficient model design with increased detection rates. The algorithms used for the Hybrid model are Random Forest, Decision Tree and XGBoost. The XGBoost algorithm improves the accuracy of the Hybrid model with better predictions. Both Federated and Hybrid model ensures efficient pre-processing and feature selection. The results of the Federated model are dependent on device datasets while the Hybrid model outperforms on the same data.