Published Paper Details:

The increasing sophistication of cyber-attacks and the inherent class imbalance in cybersecurity datasets present significant challenges to the development of accurate and reliable intrusion detection systems (IDSs). Most benchmark intrusion detection datasets, including CIC-IDS2017, CIC-IDS2018, NSL-KDD, UNSW-NB15, and Bot-IoT, contain disproportionately distributed attack categories, causing machine learning and deep learning models to exhibit biased learning toward majority classes while performing poorly on minority attacks. Traditional data balancing techniques, such as Random Oversampling and Synthetic Minority Over-sampling Technique (SMOTE), often fail to preserve the complex statistical characteristics and nonlinear relationships of network traffic, thereby limiting their effectiveness in realistic cybersecurity environments. This study evaluates the effectiveness of Generative Adversarial Network (GAN)-based synthetic data generation for balancing imbalanced cybersecurity datasets and enhancing intrusion detection performance across multiple benchmark datasets. The proposed evaluation framework incorporates data preprocessing, GAN-based synthetic attack generation, dataset balancing, feature engineering, and comprehensive performance assessment using conventional machine learning and deep learning classifiers. The generated synthetic samples are analyzed for their ability to preserve the distribution of minority attack classes while improving classifier learning and generalization. Furthermore, the performance of GAN-based augmentation is comparatively evaluated against traditional imbalance handling techniques to assess its impact on attack detection accuracy, precision, recall, F1-score, Matthews Correlation Coefficient (MCC), and ROC-AUC. The study also investigates the robustness of GAN-generated data across diverse intrusion detection datasets containing heterogeneous attack categories and network traffic characteristics. The findings are expected to demonstrate that GAN-based synthetic data generation effectively mitigates class imbalance, enhances minority attack detection, reduces false-negative rates, and improves the overall reliability of intelligent intrusion detection systems. The proposed evaluation provides valuable insights into the applicability of adversarial learning for developing robust, scalable, and data-efficient cybersecurity solutions suitable for modern enterprise, cloud, and Internet of Things (IoT) environments.