Adaptive Machine Learning Strategies for Detecting Malicious URLs

The proliferation of online services has brought even greater exposure to cyber-attacks, especially in the form of phishing and malicious URL-based threats that impersonate legitimate websites to steal user credentials and financial data. Traditional blacklisting and rule-based security solutions are not able to keep up with the changing phishing tactics, posing a challenge for developing intelligent and adaptive detection solutions. Here, a holistic machine learning based framework for malicious URL classification using supervised learning models is being proposed. Four classifiers, namely K-Nearest Neighbor (KNN), Kernel Support Vector Machine (SVM), Decision Tree, Random Forest have been trained and tested on a comprehensive dataset which contains lexical, domain based and host based URL attributes. Model efficiency is improved using the preprocessing based on standardization and the automated feature extraction. A comparative analysis based on confusion matrices and accuracy indicates that Random Forest classifier gives the best results among other classifiers with highest accuracy of 96.82%. The results demonstrate that some method is more robust to different phishing scenarios. In this work, we present an efficient, scalable and low cost detection approach to facilitate real-time cyber security system, which constitutes a practical solution to enhancing online security against emerging malicious URL attacks.