Impact of Balancing Techniques on NIDS Classification Performance: A Comparative Analysis

Network Intrusion Detection Systems (NIDS) are essential for cybersecurity, but their effectiveness is often compromised by imbalanced datasets. This imbalance stems from factors such as the rarity of certain attacks, data collection biases, and ethical constraints. This research investigates the impact of dataset balancing techniques, including undersampling, oversampling, and class weighting, on the performance of NIDS. Eight machine learning models (Logistic Regression, Linear Discriminant Analysis, Classification and Regression Tree, Random Forest, Naïve Bayes, XGBoost, Convolutional Neural Network, and Long Short-Term Memory) were evaluated on two imbalanced datasets (CSE-CIC-IDS2017 and CSE-CIC-IDS2018). The study assessed the impact of balancing techniques on detection accuracy and false positive rates. The findings demonstrate the strengths and limitations of these techniques in addressing data imbalance and improving NIDS performance, thereby contributing to more robust cybersecurity defenses in dynamic threat environments.