An Ensemble Feature Selection Approach for Intrusion Detection Systems

The accelerated development of Internet technologies and online services worldwide has increased the number of intruders drastically. To keep sensitive and confidential data safe, a machine learning-based intrusion detection system (IDS) has attracted considerable industry and academic attention in recent times. However, the efficiency and effectiveness of machine learning (ML)-based IDSs depend on the quality and quantity of training data. Therefore, the presence of repetitive and irrelevant features in training data can cause a delay in training detection and reduce the accuracy of IDS. The selection of features plays an significant role in choosing the most promising features from network traffic and thus can address the issues of IDSs. This work proposes an ensemble feature selection (EFS) method for effective detection of intrusion. The suggested approach involves the combination of features from different methods for detecting intrusions accurately. In the present work, common features of recursive feature elimination (RFE), chi-square, embedded logistics regression (Embedded_LR), embedded random forest (Embedded_RF), and embedded light gradient boosting machine (Embedded_LGBM) have been used for detailed analysis of intrusion detection accuracy. The comparative analysis was carried out on the basis of various performance parameters such as time, accuracy, TPR, and FNR, for various intrusion detection methods. This work represents the proposed method for feature selection which exhibits 99.77% accuracy with a reduced execution time of 0.4631 s, 0.998 TPR, and 0.001 FNR by using the selected feature set on UNSW-NB15 dataset. The outcomes imply that the suggested ensemble method can be used as a practical method to improve the training time, detection time, and accuracy of IDS.