A novel feature selection-driven ensemble learning approach for accurate botnet attack detection
Md. Alamgir Hossain, Md. Saiful Islam
Abstract
The rapid expansion of Internet of Things (IoT) networks has made securing devices against botnet attacks a critical challenge. This research introduces a novel feature selection-driven ensemble learning approach that uniquely combines advanced feature selection techniques—mutual information, correlation analysis, and Principal Component Analysis (PCA)—with powerful ensemble classifiers to enhance detection accuracy and efficiency. Unlike previous methods, our approach rigorously identifies and utilizes the most relevant features, addressing the challenges of high-dimensional IoT data. The Extra Trees classifier, as part of the ensemble, achieved over 99.99% accuracy in anomaly and multiclass botnet attack detection, significantly outperforming traditional methods. Furthermore, the inclusion of Explainable AI techniques, such as Local Interpretable Model-Agnostic Explanations (LIME), improved the interpretability of the detection process, providing valuable insights into model decisions. This method demonstrates a clear advantage in scalability, accuracy, and interpretability, offering a robust solution for detecting Mirai and Gafgyt botnets, thereby advancing IoT network security.