Feature Relevance Analysis and Feature Reduction of UNSW NB-15 Using Neural Networks on MAMLS

Abstract

Feature relevance is often investigated in classification problems to determine the contribution of each feature, especially when a dataset comprises of numerous features. Feature selection or variable selection aids in creating an accurate predictive model because fewer attributes tend to reduce computational complexity, thereby promising better performance. Machine learning, a preferred approach to intrusion detection, manifests on the appropriate usage of features to improve attack detection rate. A new benchmark dataset, UNSW NB-15, has been used in the study which comprises of five classes of features. This work attempts to demonstrate the relevance of each feature class along with the importance of various combinations of feature classes. During the course of this analysis, 31 possible combinations of features were taken into consideration and their relevance was examined. Empirical results pertaining to feature reduction have shown that an accuracy of 97% could be obtained by using only 23 features. The entire sequence of experimentation was conducted on Microsoft Azure machine learning studio (MAMLS), a scalable machine learning platform. Two-class neural network was used to perform the classification task. Since UNSW NB-15 is a contemporary dataset with modern attack vectors, the research community is still in the process of exploring various facets of this dataset. This article thus intends to offer valuable insights on the significance of features found in UNSW NB-15 dataset.