Dual-CVAE: Unified Data Augmentation and Feature Extraction for Resnext-GRU Phishing Detection

Phishing attacks represent a persistent and evolving threat to cybersecurity, leveraging sophisticated social engineering tactics that consistently circumvent traditional detection mechanisms. Existing solutions often struggle with the dynamic nature of these threats and the inherent class imbalance found in phishing datasets. Furthermore, state-of-the-art ensemble methods often involve substantial computational overhead, which makes them difficult to deploy in environments with limited computational resources. To overcome this limitation, this study proposes a lightweight hybrid deep learning framework aimed at improving both efficiency and detection performance. The proposed architecture synergistically combines a Conditional Variational Autoencoder (CVAE) with a ResNeXt-Gated Recurrent Unit (GRU) classifier. The CVAE serves a dual purpose: first, as a generative model to augment the training data with realistic, class-conditioned synthetic samples, thereby mitigating class imbalance without external algorithms like SMOTE; and second, as a streamlined feature extractor that transforms a highdimensional, sparse feature space into a dense, 16-dimensional latent representation. This compressed feature set is then processed by a custom ResNeXt-GRU model engineered to capture complex, non-linear patterns and sequential dependencies within the data. Experimental evaluation on a public phishing dataset demonstrates the efficacy of this approach, achieving a validation accuracy of 91.44% and a training accuracy of 97-98%. This work contributes a significant advancement in phishing detection by creating an integrated, lightweight framework that addresses data quality, feature extraction, and classification in a single end-to-end pipeline.