Escalating cyber threats demand predictive mechanisms capable of anticipating attacks before they materialize. Conventional machine learning (ML) approaches — including decision trees, support vector machines, and random forests — have provided a foundation for anomaly detection and attack classification but exhibit critical weaknesses when confronted with dynamic adversarial strategies, imbalanced data distributions, and previously unseen zero-day exploits. This paper investigates the evolutionary transition from traditional ML to Generative Artificial Intelligence (Gen AI), encompassing Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Large Language Models (LLMs). A hybrid predictive framework is proposed that integrates the statistical robustness of classical ML with the simulation capacity of Gen AI. The architecture spans layered modules for data acquisition, preprocessing, ML-based anomaly detection, adversarial simulation, predictive analytics, and real-time dashboard visualization. Experimental assessment on the NSL-KDD benchmark reveals that the hybrid approach achieves 96.4% detection accuracy, reduces false-negative rates on rare attack classes from 18.3% to 4.7%, and enables proactive zero-day threat anticipation — capabilities that exceed standalone ML by a significant margin. Ethical governance, computational considerations, and explainability requirements are systematically addressed.
Kumar et al. (Mon,) studied this question.