基于CharacterBERT的恶意URL检测模型OA

Traditional URL detection methods relying on blacklists and heuristic rules exhibit limitations when confronting new URL variants.Although the BERT(bidirectional encoder representations from transformers)model has been introduced into the field of malicious URL detection,it still faces issues like vocabulary dependence,poor handling ability of unlisted vocabulary terms,and insufficient semantic granularity.In view of the above,this paper introduces a novel malicious URL detection model integrating CharacterBERT with URL structural features.In the model,a character-level convolutional neural network(CharacterCNN)is employed to eliminate the dependency on predefined vocabularies,and deformable convolution kernels are used to extract finer semantic information.Additionally,a gated fusion network unit is developed to integrate structural features such as sub-domain quantity,sensitive word,and URL length,so as to enhance the ability to identify malicious URLs.Experimental results show that the datasets Grambeddings and kaggle_1 demonstrate the superior performance of the model,with F1-scores of 97.88%and 99.83%,respectively.To sum up,the proposed model shows outstanding detection performance and has high application value in practical security scenarios.