Detection of the Malicious URL Using the Language Models

Abstract Today, the internet has become an indispensable aspect of modern life, with many organizations providing services through web applications. At the same time, increasing technological advances also raise security concerns. One of the most common mechanisms for committing cybercrimes is using malicious URLs that host unwanted content such as spam, phishing, malware, etc. They trick unsuspecting users into falling victim to scams such as stealing money, stealing private information, and installing malware. Therefore, the problem of detecting malicious URLs is an important research issue to take timely action against such threats, and some researchers have proposed methods using machine learning and deep learning algorithms in this field. Although these investigations have achieved notable precision, most existing methods have only focused on the malicious URLs’ lexical and statistical features. They have neglected to pay attention to the meaning of the phrases in these URLs. Also, the semantic connection between these phrases and because of that, this research seeks to use language models for semantic modeling of phrases in the malicious URLs to enhance the precision of their multiclass classification. The language models analyzed in this paper are Word2vec, GloVe, LASER, and BERT. The experimental results proved that the language models achieved a significant improvement compared to the other methods. Also, the BERT language model was able to achieve the best accuracy up to 99.7%.