Correlative Analysis of Combined Machine Learning Classifiers on Anomaly-based Intrusion Detection Systems

Abstract

A detailed study on the performance improvements brought about in intrusion detection by fusing the evidence from heterogeneous classifiers derived out of supervised Neural and Algorithmic models, which have been trained across datasets differing in size and attributes to filter anomalous data packets. Binary Classification models with varying Neural layers of distinct Deep Learning architectures along with diverse rule-based probabilistic and deterministic classifiers have been constructed to give a wide comprehension regarding the impact of each classifier on the efficacy of the other. Data samples have been extracted from each of the NSL-KDD, UNSW-NB15, and CICIDS-2017 datasets to simulate real-life data conducive to an extensive and non-biased packet stratification. Logical conjunction and Matthews Correlation Coefficient (MCC) have been availed as the combination and evaluation techniques respectively. The obtained results indicate that the Support Vector Machine and Gated Recurrent Unit ensemble shows the highest accuracy. Furthermore, Logistic Regression model coupled with Naïve Bayes is the most optimal combination with regards to the MCC score and detection time taken.