Abstract

These days phishing attacks are a significant source of cybersecurity threat. They defraud others to provide confidential information such as usernames, passwords, and money by using fraudulent websites or malicious links. Most of the common techniques of identifying these attacks such as blacklist or predetermined rules mostly fail to recognize new phishing web addresses as they emerge. We propose a new solution that combines machine learning and natural language processing (NLP) models to identify phishing URLs more in an effective and efficient way in this paper. We use the top domains on Tranco to extract several characteristics of URLs, including lexical patterns, character statistics, entropy, suspicious keywords, and reputation of the domain. To enhance the accuracy of detection with the minimum computational costs, we assembled an ensemble model based on the Random Forest and XGBoost classifiers. Besides, the system has the benefit of providing real time URL analysis in an interactive interface, which is conveniently applicable in both web browsers and security systems. As our results reveal, our method as a way of tackling this issue beats some of the existing methods in terms of accuracy, preciseness, and recall. On a larger scale, we have an opportunity to enhance the security of the Internet by offering a fast, reliable, and scalable method of detecting acts of phishing.