Stability-Enhanced Feature Selection and Evolutionary Ensemble Learning: Addressing Adversarial Attacks in IoT Intrusion Detection
Waleed Almughalles, Jianming Yong, Xiaohui Tao, Yuefeng Li, Jun Shen · Tsinghua Science & Technology · 2026
Abstract The rapid growth of Internet of Things (IoT) networks has introduced significant challenges for machine learning-based intrusion detection systems (ML-based IDSs), particularly the threat of malicious feature injection through extensible data serialization formats. Such attacks can corrupt training data, compromise feature selection processes, and undermine system reliability. To address this issue, we propose a secure framework that applies recursive feature elimination with cross-validation (RFECV) across multiple models, followed by stability analysis to ensure robust feature selection. Additionally, we enhance the classification accuracy of ML-based IDSs through a weighted ensemble learning approach optimized using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Extensive experiments on two benchmark IoT intrusion detection datasets, namely UNSW-NB15 and TON-IoT, demonstrate the effectiveness of the proposed framework in mitigating adversarial feature injection. The proposed approach outperforms state-of-the-art methods across multiple performance metrics, significantly improving the robustness and accuracy of ML-based IDSs.