In modern cybersecurity scenario, an intrusion detection system (IDS) should not only be highly predictive but also be dynamically responsive to new attack vectors. In this paper, the hybrid architecture is proposed, and it combines graph-based, sequential, and tabular learning into one architecture, which is backed by the Generative AI-based cycle of data augmentation. Such a design can which enhances the F1-score by 3.7 percent andreducesthe false-positive rate by 38 percent of the baseline deep and ensemble IDS models, such as CNN-LSTM, AE-XGBoost, and GAT-IDS. With adversarial perturbations implemented on both FGSM and PGD, the loss in detection performance is less than 5% reflecting a large adversarial robustness. The generative augmentation and unified embedding fusion are the key features that differentiate the suggested design compared to the previous hybrid IDS design, providing a scalable and reproducible way to guard against cyber-threats adaptively.
B.P. et al. (Thu,) studied this question.