A New High-Performance Feature Selection Method for Machine Learning-Based IOT Intrusion Detection

In recent years, there is a significant growth experienced in both data traffic as well as dimensionality in Internet of Things (IoT) environment. In parallel, IoT networks are often threatened by high number of cyber-attacks which mandate the need for self-protective tools like Intrusion Detection Systems (IDSs). In spite of its importance, IDS systems undergo notable challenges in terms of physical and functional diversity. It is challenging for IoT characteristics to exploit the features and attributes of IDS to ensure self-protection. In account of this scenario, the current study develops a new method for feature extraction and selection for anomaly-based IDS. In this approach, the authors made use of Gain Ratio (GR) in addition to Information Gain (IG) i.e., two entropy-based approaches to choose and retrieve the coherent features in different ratios. The best features are then extracted using mathematical set theory. The model framework presented in this study was trained and assessed with the help of four Machine Learning (ML) methods, including bagging, ANN, KNN, and J48 Algorithm, using the IoT intrusion dataset, IoTID20 followed by NSL-KDD datasets. The suggested method produced 15 and 25 relevant features (out of 41) on the IoTID20 dataset, while it identified 11 and 28 relevant features (out of 86) on the NSL-KDD dataset by utilising intersection and union, respectively. Based on the outcomes attained from the comparison analysis conducted between the presented model framework with that of the previous state-of-the-art techniques, the study arrives at a conclusion that the presented model is better than others because it was able to obtain the highest classification accuracy of 89.70%.