Authors: Muhammad Hilmi Kamarudin; Carsten Maple; Tim Watson
Addresses: Cyber Security Centre, Warwick Management Manufacturing, University of Warwick, CV47AL, Coventry, UK ' Cyber Security Centre, Warwick Management Manufacturing, University of Warwick, CV47AL, Coventry, UK ' Cyber Security Centre, Warwick Management Manufacturing, University of Warwick, CV47AL, Coventry, UK
Abstract: High dimensionality's problems have make feature selection as one of the most important criteria in determining the efficiency of intrusion detection systems. In this study we have selected a hybrid feature selection model that potentially combines the strengths of both the filter and the wrapper selection procedure. The potential hybrid solution is expected to effectively select the optimal set of features in detecting intrusion. The proposed hybrid model was carried out using correlation feature selection (CFS) together with three different search techniques known as best-first, greedy stepwise and genetic algorithm. The wrapper-based subset evaluation uses a random forest (RF) classifier to evaluate each of the features that were first selected by the filter method. The reduced feature selection on both KDD99 and DARPA 1999 dataset was tested using RF algorithm with ten-fold cross-validation in a supervised environment. The experimental result shows that the hybrid feature selections had produced satisfactory outcome.
Keywords: machine learning; filter-subset evaluation; wrapper-subset evaluation; genetic algorithm; random forest.
International Journal of High Performance Computing and Networking, 2019 Vol.13 No.2, pp.232 - 240
Received: 24 Jun 2016
Accepted: 26 Aug 2016
Published online: 22 Jan 2019 *