Deep Approaches on Malicious URL Classification

Abstract

Malicious URLs are one of the biggest threats to this digital world and preventing it is one of the challenging tasks in the domain of cyber security. Previous research to tackle malicious URLs using hard-coded features have proven good indeed, but it comes with the limitation that these features are non-exhaustive and therefore detection algorithms fail to recognize new or unseen malicious URLs. However, with the deep learning revolution, this problem can be easily solved, since deep learning models extract features of their own by learning from patterns occurring in such URLs. In this paper, we have shown a comparative study of deep learning based architectures - simple RNN, simple LSTM and CNN-LSTM and how these methods can be effective for classifying URLs as malicious or benign. We have compared their performances on the basis of accuracy, precision and recall, where CNN-LSTM architecture outperforms the other two with an accuracy of 93.59%.