Research and Implementation of a Machine Learning-Based URL Phishing Detection Project

Abstract

This thesis focuses on the research and implementation of a machine learning-based URL phishing detection system to address the limitations of traditional detection methods and the black-box problem of single machine learning models. A multi-dimensional URL feature system is constructed, incorporating Web3-specific features such as brand misspelling detection and free hosting domain identification to adapt to emerging phishing attack patterns. A Random Forest classifier is trained and combined with a rule-based reasoning module to form a quantitative fusion decision mechanism, balancing detection accuracy and result interpretability. Experiments based on a balanced public dataset show that the system achieves 96.4% overall accuracy via 5-fold cross-validation and independent test set verification, outperforming most existing representative methods. The fusion framework outputs interpretable risk scores and suspicious reasons, solving the transparency issue of traditional machine learning models. Finally, the system is deployed on Hugging Face as a practical detection tool, and future research directions are proposed to optimize features, weights, and engineering applications.