Indian Journal of Science and Technology

Abstract

Objectives: To improve accuracy and reduce the computational overheads of Machine Learning (ML) classifiers to identify web server threats, develop a feature selection strategy that extracts pertinent and important features from the network dataset. Methods: This research work progressed in three phases i) Mutual Information (MI) was used first for the feature ranking and selection to reduce the dimension of feature space; ii) Genetic Algorithms (GA) were used to pick significant features for boosting the accuracy of ML classifiers; and iii) evaluates the performance of four ML classifiers; Naive Bayes (NB), k-Nearest Neighbor (k-NN), Random Forest (RF), and Support Vector Machine (SVM), using the selected features. The evaluation is conducted on the UNSW-NB 15 dataset, measuring accuracy, False Positive Rate (FPR), and computational time. Findings: The results indicate that the proposed feature selection method remarkably improves the accuracy of ML classifiers, reducing the number of features to just four. The accuracy of ML classifiers improved by 13.11%, resulting in a reduction of about 99% in computational time compared to the results reported in the literature. Novelty: A novel hybrid feature selection method is proposed, which combines feature reduction by MI, a filter-based method and further feature extraction by GA, a wrapper-based method. This approach effectively identifies essential features for enhancing the accuracy of ML classifiers.