Enhancing Intrusion Detection: A Robust Ensemble Learning Approach with Explainable AI and Strategic Fused Voting-Stacking Mechanism

Abstract

In response to the escalating complexities of cyber threats, this study introduces a robust ensemble machine learning model to enhance network intrusion detection. Combining Decision Tree, Random Forest, Gaussian Naive Bayes, Perceptron, and Multinomial Logistic Regression, our model leverages the unique strengths of these diverse algorithms to effectively manage the vast network traffic data. Initially, we apply a majority voting system to integrate the predictive capabilities of the base models. Subsequently, a stacking ensemble incorporating XGBoost as a meta-classifier is utilized, refining prediction accuracy. This two-pronged ensemble strategy, verified against the UNSW-NB15 dataset, has demonstrated superior performance, achieving precision, recall, accuracy, and F1-scores of 85.1%, 86.3%, 83.3%, and 84.7% respectively, thus outperforming existing base models. Additionally, LIME-based interpretability is integrated, enhancing transparency and providing deeper insights into the decision-making processes of the model. This research not only advances the interpretability of intrusion detection systems but also significantly elevates their effectiveness in real-world applications.