Phishing Detection Using Random Forest-Based Weighted Bootstrap Sampling and LASSO+ Feature Selection

Abstract

Phishing attacks are becoming more complex and harder to differentiate from legitimate websites. This poses serious risks to users and organizations. This study introduces a phishing detection framework that combines LASSO-based feature selection and a Random Forest classifier enhanced by Weighted Bootstrap Sampling (WBS). The framework addresses two key challenges: optimizing feature selection for high-dimensional data and managing datasets with over 70% outliers. LASSO+ extends the traditional LASSO (Least Absolute Shrinkage and Selection Operator) by integrating Pearson Correlation and Grid Search. This combination improves feature selection by identifying the most relevant features, reducing redundancy, and ensuring efficient processing without compromising accuracy. WBS further enhances Random Forest by prioritizing uncertain samples during training, enabling the model to effectively handle outlier-heavy datasets and improve recall. The proposed framework was evaluated on four diverse datasets with distinct challenges. Results demonstrated high recall rates of 99.59% for Dataset A, 98.76% for Dataset B, 100.00% for Dataset C, and 98.99% for Dataset D. The method also achieved competitive execution times. Compared to existing approaches, the framework delivered better predictive accuracy, robustness, and efficiency. This study highlights the advantages of combining LASSO+ and WBS to improve feature selection and manage outliers in phishing detection. The proposed method provides a reliable solution for addressing cybersecurity challenges in practical applications.