Research shows that deep neural network models are susceptible to the influence of adversarial samples. It has attracted strong global attention strong global attention. This study takes the text classification model as the research object and focuses on the research on the generation method of text adversarial samples. An adversarial sample generation algorithm based on synonym replacement (ASBSR) is proposed to reveal the vulnerability of deep neural network models. This method effectively reduces the number of word substitutions and optimizes the substitution order. In the experiment, three data sets of IMDB, AG’s News and Yahoo! Answers were used, and comparative experiments were conducted on three models of WordCNN, WordLSTM and WordBi-LSTM. The results show that this method can significantly reduce the model’s classification accuracy performance under the premise of low perturbation rate and it has good transferability.