Fine-tuned hyperparameter optimization for phishing website detection: insights into efficiency and performance

The escalation of digital threats has made phishing-site identification a critical aspect of online protection. This study investigates how systematic hyperparameter adjustment through grid search influences both predictive precision and computational efficiency in phishing detection. Nine supervised classifiers from different algorithmic families were analyzed: tree-based models (DT, RF, GB, XGBoost), margin and distance-based learners (SVM, k-NN), probabilistic and neural approaches (NB, MLP), and a linear baseline using logistic regression (LR). Although machine learning (ML) approaches have demonstrated strong predictive capability, their reliability largely depends on precise parameter calibration. Through systematic exploration of parameter combinations, the grid-search approach identifies optimal settings for each model. Using the Kaggle phishing-URL dataset, tuned models achieved noticeable accuracy gains. DT, RF, and k-NN reached 99.1% accuracy with training times of 0.10 s, 1.55 s, and 0.01 s, respectively. MLP yielded 99.0% accuracy but required 2758 s, while SVM and LR achieved 97.8% and 92.9%. NB did the worst (62.7%). The results indicate that careful hyperparameter optimization enhances predictive ability, whereas model complexity heavily impacts runtime. This study’s novelty lies in a balanced assessment of accuracy and efficiency trade-offs, offering guidelines for selecting computationally efficient algorithms in practical phishing-detection systems.