Optimization of Boosting-based Classification for Phishing Web Detection Using Ant Colony Optimization

Abstract

Background: Cybercriminals commonly use phishing attacks by manipulating domain and website characteristics to mislead users into revealing sensitive personal information. The increasing scale of phishing attacks demands automated detection mechanisms that are accurate, efficient, and reproducible. Objective: The purpose of this research is to compare and evaluate the performance of boosting-based machine learning models for phishing domain detection integrated with Ant Colony Optimization (ACO) for feature selection. This study also aims to analyze the impact of ACO-based feature selection on classification performance and feature efficiency under consistent experimental conditions. Methods: Experiments were conducted using a public phishing webpage dataset from Kaggle, comprising 11,430 samples and 87 numerical features extracted from URL structures and webpage characteristics. Two scenarios were evaluated: training models with all features and with a reduced feature subset selected by ACO. Performance was assessed using Accuracy, Precision, Recall, F1-score, and ROC–AUC under a fixed train–test split and predefined hyperparameters. Result: Experimental results showed that LightGBM without feature selection achieved the highest accuracy 97.24%. However, ACO reduced feature dimensionality and improved computational efficiency in some models, including faster execution and lower memory usage for LightGBM, while also slightly decreasing accuracy. These findings indicate that ACO effectiveness is model-dependent and involves a balance between predictive performance and computational efficiency. Conclusion: The results confirm that boosting-based models effectively detect phishing domains, while ACO showed model-dependent effects on feature efficiency and computational trade-offs. Future work should use diverse datasets and systematic hyperparameter optimization to improve generalizability and performance.