A Union-Driven Feature Selection Framework for Robust Phishing URL Classification

Abstract

Phishing attacks continue to exploit deceptive URL structures to compromise sensitive information and mislead users. This paper proposes a phishing detection framework that integrates a union-based feature selection strategy combining the Whale Optimization Algorithm (WOA) and the Dragonfly Algorithm (DA). The proposed WOA∪DA approach merges complementary feature subsets selected by both optimizers and removes redundancy to construct an informative and consistent feature space. The optimized features are evaluated using Extra Trees (ET) and K-Nearest Neighbor (KNN) classifiers on the ISCX-URL2016 dataset. Experimental results demonstrate that ET achieved 98.70% accuracy, 98.87% precision, 98.48% recall, and 98.68% F1-score, while KNN achieved 97.46% accuracy with competitive precision and recall values. Comparative analysis against related works shows measurable improvements in overall accuracy. The findings confirm that combining ensemble learning with union-based metaheuristic feature selection enhances classification stability, improves generalization, and strengthens phishing detection performance.