Ensemble Diversity Pruning on Cybersecurity: Optimizing Intrusion Detection Systems

Abstract

Several recent studies demonstrate that Intrusion Detection Systems (IDS) leveraging Ensemble learning techniques can effectively reduce the misclassification of malicious traffic on computer networks. However, identifying an optimal combination of classifiers often presents a significant challenge characterized by high computational cost. This work proposes an application of Diversity Pruning to address this challenge, aiming to surpass the performance of prior works. This work extend the experimental analysis by introducing four datasets for process evaluation. The results demonstrate a substantial reduction in computational cost alongside significant improvements in detection rates. The proposed approach reduced the classification errors by 18.82% for KDD-Cup'99 dataset, 26.58% for NSL-KDD dataset, 22.93% for UNSW-NB15 dataset, and 52.34% for ISCX-IDS-2012 dataset and the training time reduced by an factor of 98 for all datasets.