<![CDATA[Phishing websites remain a major cybersecurity threat, while conventional blacklist-based detection systems often fail to identify newly emerging and zero-day attacks. In addition, phishing datasets are frequently imbalanced, causing machine learning models to exhibit poor minority-class detection performance. To address these challenges, this study proposes PhishTect, a hybrid phishing detection framework that integrates Synthetic Minority Oversampling for Nominal Data (SMOTEN), Feature Tokenizer Transformer (FT-Transformer), and Particle Swarm Optimization (PSO). Unlike existing approaches that typically focus on balancing, deep learning, or optimization techniques separately, the proposed framework combines these components within a unified architecture. SMOTEN is employed to balance categorical phishing data, FT-Transformer learns contextual representations from tabular features, and PSO optimizes model hyperparameters to improve predictive capability. Experiments conducted on the PhishTank dataset evaluated three scenarios: baseline FT-Transformer, FT-Transformer with SMOTEN, and the proposed SMOTEN–FT-Transformer–PSO framework. The proposed model achieved the best performance, obtaining 96.62% accuracy, 0.9696 F1-score, 0.9963 ROC-AUC, and 0.997 PR-AUC. The results demonstrate that integrating oversampling, Transformer-based feature learning, and swarm intelligence optimization significantly improves phishing detection effectiveness, robustness, and generalization on imbalanced cybersecurity datasets.]]>
Copyrights © 2026
