Improving Cybersecurity: A Comparative Analysis of Machine Learning-Based Uniform Resource Locator (URL) Classification

Threats on the internet are a major issue, as cyberattacks continue to pose a security challenge due to the increasing number of malicious sites. Therefore, it is crucial to maintain and improve cybersecurity to prevent these malicious attacks. This study utilized machine learning and different algorithms to accurately classify Uniform Resource Locator (URL) as either benign or malicious. To address the issue of an imbalanced dataset, the Synthetic Minority Over-Sampling Technique was applied. Additionally, three methods have been used to select features: univariate selection, recursive feature elimination, and feature importance. The performance of twelve models, namely Logistic Regression, Decision Tree, Linear Discriminant Analysis, Gradient Boosting, Extreme Gradient Boosting, Adaptive Boosting, Gaussian Naïve Bayes, Bernoulli Naïve Bayes, Support Vector Classifier, Random Forest, Extra Trees, and K-Nearest Neighbors, has been evaluated using accuracy, precision, F1 score, recall, and ROC-AUC. Hyperparameter tuning was also employed to generate better results. Gradient boosting outperformed the other machine learning models with an accuracy score of ${9 8 . 4 7 \%}$. Overall, this study aims to create a hardware device capable of identifying the nature of URLs, which would improve cybersecurity practices.