A Hybrid Intrusion Detection Framework Integrating Mutual Information-based Feature Selection with Stacking-based Ensemble Learning

Abstract

Intrusion Detection Systems (IDS) are crucial for protecting modern computer networks against the 
ever-evolving landscape of cyber threats. But there are certain challenges that are hard for many machine learning 
based intrusion detection systems, such as high dimensional data, class imbalance and the detection of unknown 
attacks. To solve these problems, this study suggests a hybrid stacking-based intrusion detection system combining 
the feature selection method named Mutual Information (MI) and supervised and anomaly-based learning. In the 
proposed framework, logistic regression (LR) is proposed to be used as a meta-classifier, which takes predictions 
from Random Forest (RF) and Isolation Forest (IF) as base learners. Once the data has been processed, the features 
have been selected, and the model has been trained, the UNSW-NB15 data set is used to test the framework. The 
results of the experiments demonstrate that the proposed approach can achieve an accuracy of 97.72%, with 
precision and recall (and F1-scores) close to 0.98 for both attack and normal classes. The proposed system has been 
proved to achieve improved performance on a regular basis compared to individual classifiers and existing intrusion 
detection systems. The outcomes illustrate that feature selection combined with a hybrid ensemble learning 
technique boosts the accuracy of intrusion detection, its robustness, and its practical usability, thus showing the 
viability of the proposed framework in the context of modern network security scenarios.