ADVANCED HYBRID FEATURE SELECTION AND BAYESIAN OPTIMIZATION FOR PHISHING URL DETECTION

Abstract

Phishing attacks are a major cybersecurity threat that aim to steal sensitive information by redirecting users to fraudulent websites. Traditional blacklist- and rule-based methods often remain insufficient

especially against zero-day attacks. This study presents a systematic methodological benchmark framework for phishing URL detection

integrating hybrid feature selection

Bayesian hyperparameter optimization

model comparison

ensemble learning

explainability analysis

and additional validation strategies. Experiments were conducted on a dataset of 88

647 URLs with 111 features

evaluating 7 hybrid feature selection methods

3 Bayesian optimization techniques

and 252 optimized model configurations

complemented by deep learning baselines. The findings show that hybrid feature selection and optimization provide a clear contribution

particularly for tree-based models. The LightGBM model optimized on the L1+Boruta feature set with Scikit-Optimize produced the strongest single-model performance (Test Accuracy: 97.28%

F1: 95.99%

AUC: 99.54%)

while the best ensemble (CAT+XGB+LGBM

Soft Voting) reached 97.34% accuracy

96.17% F1

and 99.60% AUC — a comparable level of performance. Deep learning models yielded lower performance

while additional validation experiments demonstrated that the ensemble structure provided a consistent improvement over the baseline across different data splits.