Background: Network intrusion detection systems (NIDS) are a critical line of defence against cyberattacks, and machine learning approaches have increasingly replaced signature-based methods due to their ability to detect previously unseen attack patterns, though performance varies substantially across algorithms and is strongly affected by class imbalance in benchmark datasets.

Objective: This study evaluates and compares the performance of five machine learning classifiers, Naive Bayes, Logistic Regression, Random Forest, XGBoost, and a Deep Neural Network (DNN), for binary network intrusion detection (normal vs. attack traffic), with the Synthetic Minority Oversampling Technique (SMOTE) applied to address class imbalance.

Method: A simulated dataset, modelled on patterns reported in published research using the NSL-KDD and CICIDS2017 benchmark datasets, was used to evaluate accuracy, precision, recall, F1-score, and area under the ROC curve (AUC) across thirty simulated train-test runs (six replicates per classifier). Data were analysed using descriptive statistics and one-way analysis of variance (ANOVA) in Python (scikit-learn, XGBoost) and SPSS (version 27).

Key Results: XGBoost achieved the highest overall performance (accuracy = 98.6%, F1-score = 98.4%, AUC = 0.994), followed closely by the Deep Neural Network (accuracy = 98.2%) and Random Forest (accuracy = 97.8%). Naive Bayes recorded the weakest performance (accuracy = 89.4%, F1-score = 85.2%), consistent with its known sensitivity to feature independence assumptions. SMOTE-based class balancing improved minority-class (attack) recall across all classifiers, with the largest relative gain observed for Logistic Regression.