A HYBRID STATISTICAL FEATURE SELECTION FRAMEWORK FOR EFFICIENT IOT INTRUSION DETECTION

Abstract

With the development of Internet of Things (IoT) technology in recent years, there has been an increase in the amount of IoT-based gadgets that have applications which can be applied in many different industries. At the same time, the rise of IoT has led to an increase in the vulnerability to cybersecurity attacks, and therefore, Intrusion Detection Systems (IDSs) have become very important. In this study, we propose a statistically driven feature selection framework for intrusion detection using Cohen’s d, Mann–Whitney U test, and Kolmogorov–Smirnov test to rank features from a high-dimensional dataset. The considered dataset comprises 95 features and 685,671 instances of the CIC-IIoT 2025 dataset. Feature importance is computed on the training set, after which feature score normalization, composite ranking, and correlation-based filtering of the features take place to eliminate redundant ones. The chosen features are analyzed based on multiple machine learning classifiers with different numbers of features. The best classification accuracy was obtained using the XGBoost classifier with 40 features selected by the composite filter combining Cohen’s d test, Mann–Whitney U test, and Kolmogorov–Smirnov test, achieving an accuracy of 99.05%, an F1-score of 99.18%, and an AUC-ROC of 99.92%