The currently used URL identification methods require a large number of tags. The malicious URLs update quickly, and it is difficult to collect enough comprehensive URL tags, resulting in unstable identification accuracy. After calculating the boundary similarity of the URL string, the Skip-Gram model is used to embed the URL. The processed word vector is used as the generator input of the semi-supervised learning GAN to obtain the malicious URL identification result. The experimental results show that the accuracy of the URL recognition using GAN is higher than 96%, the fluctuation of the F1 value is small, and the recognition results are more reliable.