Comparative Analysis of Popular Supervised Machine Learning Algorithms for Detecting Malicious Universal Resource Locators

Malicious Universal Resource Locators (URLs), also referred to as malicious websites have become a serious cause for concern for cyber security administrators of various organisations, institutions, Agencies, businesses and companies. These websites host malware, spam, drive by links and phishing. Unfortunately, Internet users worldwide visit such malicious sites and become the victims of cybercrimes like credit card credentials theft, theft of personal information, monetary savings or investments. Multitudes of researchers have embarked on attempts to design and implement response solutions to malicious URLs threat. The approaches are largely divided into two groups, the traditional approaches (Blacklising and Heuristics) and the data driven approaches (statistical methods, machine learning methods, data mining methods, and deep learning methods). In some instances, there are divergent views on which algorithm is the best to be used for building models. To our knowledge, there are still few works that have taken an initiative to comparatively analyse the performance of machine learning algorithms which have been identified by various authors as being the most suitable to use for building detection models. This study therefore focused on the Light Gradient Boost, Extreme Gradient Boost and the Random Forest algorithms. For the study's experiments, a malicious URLs dataset was downloaded from Kaggle.com databases. The study's results demonstrated that the hostname_length was the most important feature to focus on when building malicious URL detection models using the three above mentioned algorithms. The results also revealed two more features that had importance; the count_www and the count_dir, when using Extreme Gradient Boosting and the Random Forest. The study will in future explore hybrid models where advantages of various algorithms will be exploited to be combined in order to improve performance. Other models that will be considered include Support Vector Machine, Neural Networks and Deep learning models.