Variance Threshold as Early Screening to Boruta Feature Selection for Intrusion Detection System

Abstract

A rapid development of internet technology brings convenience to society and threat of exploitation at the same time. As a countermeasure, an Intrusion Detection System (IDS) was introduced. Research to improve its performance in differentiating normal traffic from malicious ones has been carried out by exploring machine learning. One of them implemented the Boruta algorithm, whose performance is still challenging in processing time to select appropriate features of the NSL-KDD dataset. Some studies work on this issue, which is then labeled as an “infinite loop” problem. However, the methods do not work on every scenario of the experiments, despite showing terrific results on classification using Random Forests. In this paper, we resolve this matter using a statistical approach, which in this case is Variance Threshold, to eliminate unnecessary features earlier so that Boruta would be able to identify all accepted and rejected features sooner while hoping with the same Random Forests that the classification result would not be too affected. It turned out that the proposed method does not work well, and surprisingly, the classification cannot reach 76% accuracy. Nevertheless, we might find a potential flaw in the former study and possibly rule out its result.