Comparative study on feature selection techniques in intrusion detection systems using ensemble classifiers

Abstract

Network usage has become a paramount aspect of life, therefore, securing our networks is crucial. The world is experiencing a rapid breakthrough of internet usage, most especially with the concept of internet of things (IoT), now internet of everything (IoE. ). Real network data is rowdy, noisy and inconsistent. These issues with the data influences the performance of intrusion detection systems (IDS) and develop manifold of false alarms. Feature selection technique is used to remove the inconsistent and rowdy data from a large data set and presents a refined set of data. This research work adopts the use of two distinct feature selection technique in parallel: ReliefF ranking and particle swarm optimization, using linear discriminant analysis (LDA) and logistic regression (LR) as the machine learners, to first clean the data, train the classifiers, and subsequently classify new instances. The results showed that, the combination of the ReliefF with the ensemble machine learning (Linear Discriminant Analysis and Logistic Regression) has a higher classification accuracy of 99.7% compared to the Particle swarm optimization (PSO) which attained an accuracy of 98.6%.