Detection of malicious URLs using Temporal Convolutional Network and Multi-Head Self-Attention mechanism

Natural Language Processing (NLP) and Deep Learning (DL) have achieved remarkable results in various fields and have also been proven to be effective in detecting phishing webpages. Inspired by the great success of NLP and DL models in phishing detection-related tasks, we examined the application of these techniques in classifying malicious and benign URLs (Uniform Resource Locator). We found that the existing NLPbased solutions mainly used Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) for phishing URL detection. However, CNN performs poorly when handling non-spatial data, while RNN cannot capture the long-distance dependency and has higher computational complexity. To overcome these issues, this paper proposes a phishing detection model based on Temporal Convolutional Network (TCN) to address the limitations of the conventional CNN and/or RNN algorithms. The proposed model used character-level and word-level embedding methods to obtain the feature representations of the input URLs. Then, TCN was employed for feature extraction and further enhanced with Multi-Head Self-Attention (MHSA) mechanism to classify legitimate and phishing websites. We conducted several experiments to validate the performance of the proposed model and measured various evaluation metrics. The obtained results showed that our solution performed better than other baseline models in classifying malicious URLs, achieving an accuracy of 98.78%. This implied the proposed approach provided an effective and efficient solution for detecting phishing URLs.