Identifying Malicious URLs Using Deep Learning based VGG16 Architecture with Transfer Learning Method

Recently, a sharp rise in cybersecurity attacks, including ransomware, phishing, malware injection, etc., has been seen on many websites worldwide. Hence, numerous commercial institutions, e-commerce companies, and individuals suffered substantial monetary losses. Experts in cyber security need help in this situation because new varieties of attacks are emerging daily. Deep learning-based algorithms have outperformed compared to other traditional machine learning algorithms in a variety of applications. However, classifying the Uniform Resource Locators (URLs) into malicious and non-malicious URLs is complex using feature extraction-based traditional machine learning algorithms. Therefore, there is some scope for further improvement. This article suggests a VGG16-based transfer learning approach to develop a model to detect malicious URLs. The malicious URL dataset, comprising 651,191 URLs, where 428,103 are benign or safe, 96,457 defacements, 94,111 phishing, and 32,520 malware, is used to develop the suggested model. The efficacy of the suggested model is measured using the loss, accuracy, precision, recall, and f1-score, and the suggested model attained efficacy of 0.152, 95.15%, 95.27, 95.03%, and 95%, respectively. The efficacy of the suggested model is also compared with the other existing models, and it is observed that the proposed model outperformed.