Assessing the Effectiveness of Synthetic Data Generation for Multi-Class Cyber-Attacks Detection using Generative Adversarial Networks

This paper presents a new intrusion detection system using Conditional Tabular Generative Adversarial Networks (CTGAN) with Consistency-based feature selection (CON) on the CIC-IDS2018 dataset. The performance of CTGAN-generated balanced data was compared with traditional oversampling methods like SMOTE and ADASYN. A mix of synthetic and real data, in varying proportions, was then used to train and evaluate the effectiveness of Random Forest (RF), XGBoost (XGB), LightGBM (LGBM), and Deep Neural Networks (DNN) in detecting network intrusions. Our XGB classifier with CTGAN-generated data outperformed other methods with 99.99% accuracy, recall, and kappa, using only 13 of 81 features. At least 70% synthetic data is needed to maintain optimal performance.