XPhish: An Interpretable Machine and Deep Learning Framework for Phishing URL Detection

Phishing remains a critical cybersecurity threat, requiring detection frameworks that combine accuracy and interpretability. This study presents a complete machine learning and deep learning framework for phishing URL detection, developed using a newly created dataset of 50,406 labeled instances. The dataset incorporates 30 lexical and network infrastructure features. Multiple models, including Random Forest (RF), XGBoost, and a 1D Convolutional Neural Network (CNN), were evaluated. Random Forest achieved the highest accuracy of 95.76%, followed by CNN 1D (95.17%) and XGBoost (94.48%), both of which delivered competitive performance. To address model transparency, SHAP and LIME were applied, providing feature-level explanations of predictions. The proposed framework integrates dataset contribution, predictive performance, and explainable AI, offering a practical foundation for real-world phishing defense.