Abstract
An Intrusion Detection System (IDS) is essential to identify cyber-attacks and implement appropriate measures for each risk. The efficiency of the Machine Learning (ML) techniques is compromised in the presence of irrelevant features and class imbalance. In this research, an efficient data pre-processing strategy was proposed to enhance the model’s generalizability. The class dissimilarity is addressed using k-Means SMOTE. After this, we furnish a hybrid feature selection method that combines filters and wrappers. Further, a hyperparameter-tuned Light Gradient Boosting Machine (LGBM) is analyzed by varying the optimal feature subsets. The experiments used the datasets – UNSW-NB15 and CICIDS-2017, yielding an accuracy of 90.71% and 99.98%, respectively. As the transparency and generalizability of the model depend significantly on understanding each component of the prediction, we employed the eXplainable Artificial Intelligence (XAI) method, SHapley Additive exPlanation (SHAP), to improve the comprehension of forecasted results.