Classifying Benign and Malicious Websites through URL Feature Extraction Using Transfer Learning

With the rapid development of digital and communication infrastructures, Internet usage is expanding rapidly. As the global number of Internet users increases, so does the prevalence of cybercriminals and websites infected with various types of malware. Cybercriminals often exploit these infected websites to carry out their malicious activities. Therefore, identifying these websites is crucial to safeguarding Internet users. However, traditional blocklists provided by web browsing services remain ineffective in detecting zero-day attacks, highlighting the need for machine learning techniques in this area. In this paper, we utilize features extracted from URLs of both benign and malicious sites–including defacement, phishing, malware, and spam – to differentiate between the two categories. First, we transform the feature space, then perform feature extraction using pre-trained models such as ResNet50, SqueezeNet, and AlexNet. Finally, we classify the URLs using a long short-term memory (LSTM) deep neural network. This proposed method effectively mitigates the risk posed by four types of malicious websites. The simulation, implemented in Python, achieved an accuracy rate of 97.13%.