Hybrid Learning Model for intrusion detection system: A combination of parametric and non-parametric classifiers

The growing digital transformation has increased the need for effective intrusion detection systems. Traditional intrusion detection systems face challenges in accurately classifying complex patterns. To address this issue, this study proposed a Hybrid Learning Model (HLM) that combines both parametric and non-parametric classifiers. The proposed HLM consist of two stages: the first stage employs a non-parametric Base Learner (np-BL) to analyze the data patterns and the second stage involves meta-modelling to generalize the overall performance of the model, named the Parametric Meta-Learning (PML) model. The proposed HLM blends the outcomes of np-BL and PML models using a stacking ensemble. As a base learning model K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Gradient Boosting Machine (GBM), and Support Vector Classification with Radial Basis Function (SVC-RBF), are adopted from a non-parametric classifier group. The parametric classifiers Logistic Regression (LR), Naïve Bayes Classifier (NBC), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA) and Support Vector Machine with linear kernel (Linear SVM) were used as meta-models. The HLM, as proposed, enhances the adaptability and robustness of the model by combining non-parametric and parametric models. To evaluate the competence of the proposed HLM, a performance analysis was conducted using the NSL-KDD, UNSW-NB15, and CICIDS2017 datasets. The effectiveness was assessed using various metrics, including classification accuracy, precision, recall, F1-Score (F1), Receiver Operating Characteristic (ROC) curve, Detection Rate (DR), and False Alarm Rate (FAR). The proposed HLM achieves a better accuracy rate across different datasets when compared with the existing models. The achieved accuracies are 99.02 %, 99.98 % and 99.63 % for the NSL-KDD, UNSW-NB15, and CICIDS2017 datasets respectively. Furthermore, the HLM gave a significant reduction in FAR, with values of 0.0126, 0.0001, and 0.0016 for the above-mentioned datasets.