Network Intrusion Detection System Using Stacking of Heterogeneous Base Learners

Abstract

In this electronic era, the importance of computer networks for social communication has increased. Consequently, organizations' internal and external intruders with new attacks are growing exponentially. Bagging, boosting, and stacked ensemble methods deal with excellent accuracy and fewer false positives. In this paper, a novel stacked method of ensemble is proposed for a network intrusion detection system. Selecting suitable base learners for the meta-classifier is a critical process. For receiving higher classification accuracy, four strong heterogeneous base learners have been selected to construct a stacked classifier. Two decision trees, a Naïve Bayes and one Rule learner, have stacked using the Logistic Regression Meta classifier. The performances of base learners and the proposed stacked classifier have been measured in terms of false positive, accuracy, model-building time, precision and recall. Base learners and meta-classifiers have been trained and tested on NSL-KDD datasets. The experimental results show that the proposed stacked classifier offers accuracies of 83.20%, 99.95% and 99.89 %on test, training datasets and cross-validation, respectively. The proposed stacked classifier outperforms its base learners and some existing intrusion detection systems. It also offers better false positive, precision and recall values than its base learners and existing intrusion detection system.