An Explainable Machine Learning-Based Phishing Website Detection using Gradient Boosting

As the need for achieving optimal performance in prediction models, several recent and complex models have been developed. However, many of these models operate as black boxes, providing little insight into their predictive results. In recent years, three gradient-boosting methods based on Decision Trees have been recommended, i.e., XGBoost, CatBoost, and LightGBM. These methods have proven to deliver competitive performance with fast training times. Nevertheless, in critical domains like security, there is a growing need for increased transparency among stakeholders. One pressing concern in the security field is the proliferation of phishing websites. To address this issue, we propose an explainable machine learning-based approach using gradient-boosting methods with hyperparameter optimization on three phishing website datasets. Our best methods surpass the state-of-the-art phishing website detection methods, achieving accuracy rates of 97.45%, 99.16%, and $\mathbf{9 7. 8 5 \%}$ for the UCI (2015), Mendeley (2018), and Mendeley (2020) datasets, respectively. Subsequently, we implement posthoc explainability using SHAP and LIME for the selected dataset. The experimental results indicate that three features, i.e., length_url, directory_length, and time_domain_activation, are consistently identified as the most influential features in the dataset. Moreover, our proposed approach demonstrates promising results for detecting phishing websites with both high accuracy and explainability.