Abstract

Network technologies are becoming more digitalized and vulnerable to various cyberattacks, therefore creation of effective Intrusion Detection Systems (IDSs) is essential, particularly for high network traffic volumes and need to distinguish between normal and abnormal activities. In this study, four case models of IDS with varying feature values, base classifiers (Naïve Bayes, k-Nearest Neighbor, Logistic Regression) and) Meta classifier (Random Forest) trained and tested in a stack ensemble method with UNSWNB-15 dataset are examined. The Particle Swarm Optimization algorithm (PSO) serves as the foundation for the two sets of selecting and extracting features with enhanced with Residue Number System (RNS) forward conversion. The performance of the model is evaluated using classification accuracy, error rate, precision, specificity, F-score, sensitivity, and training time. Case D (NB + LR + KNN with RF) model performs best with PSO+RNS selected features, as evidenced by its accuracy of 97.47%, compared to 95.36% for PSO-based selected features respectively.