Reduced Stacking Classifier Model with Optimized feature Selection for Efficient Phishing Website Detection

Abstract

Phishing websites remain an effective vehicle for credential theft and financial fraud, while machine-learning detectors must increasingly balance predictive performance against the cost of deployment. Stacking ensembles can improve classification by combining diverse learners, but large ensembles operating on full feature sets increase inference cost. This paper examines an HHO–GWO reduced stacking model for phishing website classification. It uses Harris Hawks Optimisation (HHO) for feature selection, Random Forest, XGBoost, and Gradient Boosting classifiers, and Grey Wolf Optimiser (GWO) weights for combining classifier probabilities. The model was evaluated on the UCI Phishing Websites dataset and the Mendeley Web Page Phishing Detection dataset using five random seeds, and compared with individual classifiers, unoptimised stacking, a replicated nine-classifier stacking baseline, and ablated configurations. It obtained mean accuracies of 94.40% and 96.38% and mean F1-scores of 94.53% and 96.31% on UCI and Mendeley, respectively. It used an average of 23.2 of 30 UCI features and 65.0 of 87 Mendeley features. Relative to the nine-classifier baseline, mean inference latency declined from 1.924 ms to 0.465 ms on UCI and from 1.457 ms to 0.271 ms on Mendeley, corresponding to reductions of 75.8% and 81.4%. The associated accuracy differences were 0.47 and 0.39 percentage points. The findings position HHO–GWO reduced stacking as a practical accuracy–efficiency alternative for offline phishing classification experiments. Its suitability for live URL screening or evolving attack streams requires further validation.