Detection of Malicious URL Based on BERT-CNN

Disguising the content of the website service often lures users to visit the website URL, and malicious attacks are carried out on the user's computer, exposing the user to many security risks when using the Internet. At present, methods based on machine learning and some methods based on deep learning are inefficient and cannot accurately detect malicious URLs. To address the problem, we propose a detection method that combines BERT and CNN, termed as Malicious URL Detection based on BERT-CNN, which firstly segments the URL based on the WordPiece model, then obtains the embedding based on the pretrained model BERT, and finally connects CNNs with three convolution kernels after the BERT output layer to construct a BERT-CNN-based detection model. Experimental results show that the method outperforms the random forest, CNN, LSTM and BERT methods, indicating the proposed method can effectively detect malicious URLs.