Fusion of statistical importance for feature selection in Deep Neural Network-based Intrusion Detection System

Abstract

Intrusion Detection System (IDS) is an essential part of network as it contributes towards securing the network against various vulnerabilities and threats. Over the past decades, there has been comprehensive study in the field of IDS and various approaches have been developed to design intrusion detection and classification system. With the proliferation in the usage of Deep Learning (DL) techniques and their ability to learn data extensively, we aim to design Deep Neural Network (DNN)-based IDS. In this study, we aim to focus on enhancing the performance of DNN-based IDS by proposing a novel feature selection technique that selects features via fusion of statistical importance using Standard Deviation and Difference of Mean and Median. Here, in the proposed approach, features are pruned based on their rank derived using fusion of statistical importance. Moreover, fusion of statistical importance aims to derive relevant features that possess high discernibility and deviation, that assists in better learning of data. The performance of the proposed approach is evaluated using three intrusion detection datasets, namely, NSL-KDD, UNSW_NB-15, and CIC-IDS-2017. Performance analysis is presented in terms of different evaluation metrics such as accuracy, precision, recall, f-score, and False Positive Rate (FPR) and the results are compared with existing feature selection techniques. Apart from evaluation metrics, performance comparison is also presented in terms of execution time. Moreover, results achieved are also statistically tested using Wilcoxon Signed Rank test.