SecURL: Design and Implementation of a Machine Learning-Based Web Extension for Malicious URL Detection

An ideal malicious URL detection service is one that can provide reliable detection during a user's browsing session without introducing significant latency in loading benign pages. We designed and implemented a machine learning-based malicious URL detection system that minimizes detection delay while maximizing accuracy through feature reduction techniques. A comparative analysis of feature reduction techniques revealed that the XGBoost algorithm trained with the reduced feature set from wrapper-based feature selection provides the best performance with an accuracy of up to 97.94% and minimum score time of 12.8 ms, utilizing lexical and content-based features of a URL. The detection service was packaged as an installable browser extension which analyzes visited URLs in the browser with an average end-to-end delay ranging from 60ms to 70ms.