Enhancing Cybersecurity Against Phishing Attacks Using VGA-Hyphishnet: A Hybrid Generative–Discriminative Model

Traditional phishing detection systems based on machine learning (ML) and deep learning (DL) techniques largely rely on static patterns and manually engineered features. However, as phishing attacks continue to evolve in complexity and sophistication, these conventional approaches face significant limitations in accurately identifying emerging threats. To address this challenge, this study proposes VGA-HyPhishNet, a self-evolving generative artificial intelligence (AI) framework designed for real-time phishing URL detection. The proposed architecture integrates a generative encoder, variational generative modeling (VGM), a self-evolving generative threat engine (SE-GTE), and a hybrid generative–discriminative classifier to learn distribution-aware latent representations directly from raw URL strings. Unlike traditional discriminative models that rely solely on labeled patterns, the proposed framework leverages synthetic latent threat generation to dynamically adapt to evolving phishing behaviors and probabilistic structural patterns. The effectiveness of the proposed model was evaluated using the URL-Phish dataset, consisting of 111,660 URLs. Experimental results demonstrate that VGA-HyPhishNet achieves an accuracy of 98.73%, precision of 93.00%, recall of 94.97%, and an F1-score of 93.97%. Furthermore, comparative analysis with baseline models, including URL-transformer, BiLSTM-URL, and CharCNN-URL, shows that the proposed framework significantly reduces false positives (FP) and false negatives (FN) while maintaining low inference latency suitable for real-time deployment. Overall, the results indicate that the proposed work provides a robust and reliable solution for modern cybersecurity systems by leveraging generative representation learning and adaptive threat modeling, thereby improving resilience against zero-day phishing attacks and adversarial threats.