A Multi-layer Stack Ensemble Approach to Improve Intrusion Detection System's Prediction Accuracy

Abstract

Intrusion is a series of activities that violate an organisation's existing security goals and procedures. Hence, an Intrusion Detection System (IDS) should be capable of analysing incoming network traffic (packet) and determining if it is an attack or otherwise. Lack of recent and up to date data sets for the training of IDS is a critical issue in the development of effective IDS. This paper focuses on creating a more realistic data set in our case UMaT-OD-20 using ONDaSCA and also the building a Multi-layer Stack Ensemble (MLS) IDS Model for Intrusion Detection Systems. Multi-layer Stacked Ensemble exploits the strengths of various base-level model predictions to build a more robust meta-classifier that meliorate classification accuracy and reduces False Alarm Rate (FAR). Five (5) Supervised Machine Learning (ML) based algorithms videlicet K Nearest Neighbor (KNN), Decision Tree (DT), Logistic Regression (LR), Random Forest (RF) and Naive Bayes' (NB) were employed to generate predictive models for all features. The Python programming language was used for the entire research and all programming and evaluation of data was done with an Inter Core i7, 16GB RAM and 1TB HDD Windows 10 Pro Laptop computer. The predictions of the Multi-layer stacked ensemble showed an improvement of 0.97% over the best base model. This improvement reduced the FAR during the classification of network connections types. Again, the evaluation of our work shows a significant improvement over similar works in literature.