PhishNet 1.0: optuna-optimized stacking ensemble with Boruta-based feature selection for phishing URL detection

Abstract

The objective of this research is to enhance phishing detection through ensemble learning integrated with well-structured metaheuristic algorithms. Various classifiers, including Logistic Regression, Nearest Neighbors, Support Vector Machine, Decision Tree, Naïve Bayes, and Gradient Boosting, were evaluated using features selected via the Boruta method. Among these, Gradient Boosting, KNN, and Decision Tree achieved the highest performance. These models were subsequently incorporated into two ensemble classifiers, namely Soft Voting and Stacking, with Logistic Regression selected as the final estimator in the stacking model, which outperformed alternative estimators. Furthermore, several metaheuristic optimization algorithms, such as genetic algorithm (GA), ant colony optimization (ACO), particle swarm optimization (PSO), bayesian optimization, and optuna, were employed to optimize the hyperparameters of the stacking model, thereby improving classification performance. Among these, the optuna-optimized stacking classifier achieved the best results, with an accuracy of 96.15%, precision of 96.45%, recall of 96.68%, and F1 score of 96.56%. This work makes a novel contribution by presenting an integrated framework, PhishNet 1.0, which combines ensemble learning with diverse metaheuristic optimizers and demonstrates its empirical effectiveness on phishing URL datasets. Its practical and experimental focus differentiates it from prior research and establishes a reproducible benchmark for cybersecurity applications. The findings conclusively demonstrate that ensemble learning combined with metaheuristic optimization substantially enhances phishing webpage detection and provides a reliable approach for real-world cybersecurity systems.