Malicious URL detection with distributed representation and deep learning

Abstract

There exist numerous solutions to detect malicious URLs based on Natural Language Processing and machine learning technologies. However, there is a lack of comparative analysis among approaches using distributed representation and deep learning. To solve this problem, this paper performs a comparative study on phishing URL detection based on text embedding and deep learning algorithms. Specifically, character-level and word-level embedding were combined to learn the feature representations from the webpage URLs. In addition, three deep learning models, including Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Bidirectional Long Short-Term Memory (BiLSTM), were constructed for effective classification of phishing websites. Several experiments were conducted and various evaluation metrics were used to assess the performance of these deep learning models. The findings obtained from the experiments indicated that the combination of the character-level and word-level embedding approach produced better results than the individual text representation methods. Also, the CNN-based model outperformed the other two deep learning algorithms in terms of both detection accuracy and execution time.