Malicious Website Detection Method Based on BGResNet Multi-feature Fusion

Website security detection is important for Internet security. Existing machine learning-based malicious URL detection methods have a low accuracy and weak generalization ability. Thus, we proposed a new multi-feature fusion malicious website detection model BGResNet by integrating the advantages of Bidirectional Encoder Representations from Transformers (BERT), Graph Convolutional Network (GCN), and Residual neural network (ResNet). The method integrates three features: Uniform Resource Locator (URL) creator, rule features, and website titles. First, we used BGResNet to process URL characters and website titles separately, transforming them into vector representations. Then, these two vectors were fused with URL rule vectors. Finally, we employed the softmax function to realize the detection of malicious websites. The experimental results showed that the proposed model exhibited significant superiority in detecting malicious websites, providing a new and effective method for malicious URL detection in Internet security.