Analysis of feature selection algorithms for IDS using machine learning classification.

Abstract

The network and digital devices consumers are increasing rapidly, simultaneously the inevitability of its security and intruder detection. An intelligent intrusion detection system is needed to detect novel vulnerable attacks. The existing system practices the NS KDD, KDD Cup99, UNSW-NB15, and CICIDS2017 datasets that have old network traffic. The latest captured dataset UKM-IDS20 involves novel ARP Poisoning, DoS, Exploits, and Port Scan attacks. This article analyzed the filter-based feature selection(FS) method with rule based and tree based machine learning classifiers using multiclass classification: The Gain Ratio (GR), Chi-square, Info. Gain(IG), symmetric uncertainty(SU), and correlation(CR) filter based methods choose the vital feature from the UKM-IDS20 dataset. The highest accuracy and lesser model building time machine classifier are decided to select for the proposed framework. The preferred feature form accomplished IG feature evaluation method achieves superior accuracy on the Hoeffding tree based classifier compared with the JRip rule based classifier. The Hoeffding classifier proceeds the model within 0.13 seconds using 23 selected features. The proposed IDS framework compared to the existing systems.