CTGAN-Based Multi-View Learning (CTGAN-MVL) for Intrusion Detection Systems

Intrusion Detection Systems (IDSs) are a critical component of network security, yet their effectiveness is often undermined by severe class imbalance and the limited representation of rare attack types in real-world traffic datasets. Conditional Tabular Generative Adversarial Networks (CTGAN) can be an effective solution for generating realistic synthetic tabular data and mitigating minority-class scarcity in intrusion detection scenarios. Building upon this capability, this paper proposes CTGAN-Based Multi-View Learning (CTGAN-MVL), a novel IDS methodology that integrates CTGAN-driven data augmentation with complementary multi-view feature representations and ensemble learning. Multiple data views are constructed using diverse feature selection techniques to capture different characteristics of network traffic, and CTGANs are applied within each view to generate synthetic samples for underrepresented attack classes. View-specific tree-based classifiers are then trained and combined through a class-aware weighted ensemble optimized via AutoML. Exxperiments on the CIC-IDS 2017 and NSL-KDD benchmarks demonstrate that CTGAN-MVL achieves 99.98% precision, recall, F1-score, and accuracy on CIC-IDS 2017, and 99.59% across the same metrics on NSL-KDD, consistently surpassing state-of-the-art intrusion detection methods.