Optimizing Feature Selection for IoT Intrusion Detection: A Hybrid Bio-Inspired Method

Abstract

With the rapid adoption of Internet of Things (IoT) applications, the risk of cyber-attacks has increased significantly, underscoring the need for efficient intrusion detection systems (IDS). Machine learning models have proven effective for detecting malicious traffic, but the high-dimensional feature space often introduces computational challenges. To address this, we propose a lightweight IDS combining machine learning with a novel twostage hybrid feature selection approach. In the first stage, filterbased methods reduce the feature space to lower computational demands. The second stage applies a hybrid of genetic algorithms and whale optimization to select an optimal subset of features, leveraging the strengths of both algorithms for enhanced search and exploration. For classification, we employ three ensemble learning models, with LightGBM achieving accuracy scores of $\mathbf{9 9. 9 9 \%}$ on the N-BaIoT dataset. Our approach surpasses current methods in performance and classification efficiency.