Exploring the Effectiveness of Synthetic Data in Network Intrusion Detection through XAI

Machine learning based Intrusion Detection Systems often utilize synthesized samples to address class imbalance in datasets and improve the detection of rare malicious activities. It is commonly assumed that models trained on synthetic data exhibit the same interpretability as those trained on real data. In this study, we challenge this assumption by integrating Conditional Generative Adversarial Networks (CGANs) with eXplainable AI (XAI) techniques to analyze differences in model behavior, specifically in feature importance and decision boundaries, relative to overall performance. We train multiple Multilayer Perceptron classifiers using both synthetic and real samples and employ SHapley Additive exPlanations to investigate variations in interpretability. Our experimental results demonstrate comparable performance achieved by models trained using synthetic and real data, however there are significant differences in interpretive characteristics suggesting that synthetic data can influence model behavior in distinct ways.