Detection and Selection of Task-specific Features Algorithms for IoT-based Networks

Abstract

In IoT-based home/enterprise network applications, an advanced security system is desirable for resource-constrained devices. Feature selection significantly affects the performance of a Machine Learning-based Intrusion Detection System (ML-IDS) to which data of the highest quality should be fed. An appropriate feature selection with sufficient features increases the accuracy of the Intrusion Detection System (IDS) classification. In addition, the consistent use of the same metrics in feature selection and detection algorithms further enhances classification accuracy. First, this paper studies two feature selection algorithms, Information Gain, a metric of entropy, and PSO-based feature selection, a metric of misclassification, to select a minimum number of attack feature subsets for resource-constrained IoT devices. Then, the detection algorithms for multi-classifications, Tree and Ensemble, are evaluated regarding non-consistent and consistent metrics. For specific performance comparison, the same metrics for feature selection and detection algorithm are utilized and compared with non-consistent use of feature selection and detection algorithm, e.g., feature selection by Information Gain (entropy) and Tree detection algorithm by classification.