Multiclass Intrusion Detection in IoT Using Boosting and Feature Selection

The proliferation of Internet of Things (IoT) devices has resulted from the continuous evolution of interconnected computing devices and emergence of novel network technologies. Intrusion Detection Systems (IDSs) play a pivotal role in ensuring IoT network security. Nonetheless, the development of a robust IDS with high performance for detecting specific anomalies to prevent attacks remains a challenging task. This study systematically evaluated and compared two classifies XGBoost and LightGBM employing five filter-based feature selection methods (ANOVA, Kendall's test, Mutual Information, Maximum Relevance Minimum Redundancy, and Chi-square) with distinct selection thresholds. The evaluation was conducted over the NF-ToN-IoT-v2 dataset using the Scott-Knott test and Borda Count system. A total of 152 classifier variants were assessed to determine the most effective model. Results demonstrated that utilizing XGBoost in conjunction with 200 estimators, ANOVA and MI as the feature selection combination, and selecting the top 18 features, constituted an effective and powerful model compared to alternative methods.