Rapid development of deep learning has made it one of the most effective tools for data utilization. However, the training process of deep learning algorithms often requires a large amount of user data, which brings great risks of privacy leakage. Due to the complexity and semantics of data features, traditional data desensitization techniques usually need to over-clarify the original data, resulting in low data availability and difficult to apply to deep learning models. To this end, a data desensitization algorithm of personalized differential privacy generative adversarial network (PDP-GAN-DD) is proposed. By adding noise to the gradient of SeqGAN model training, differential privacy is realized, which ensures that GAN can generate an unlimited amount of synthetic data that meets the statistical characteristics of the source data but does not leak privacy. Aiming at the problems of slow convergence and poor data quality of the existing differential privacy generative adversarial algorithm, a dynamic privacy budget allocation and adaptive gradient cutting threshold optimization strategy are designed, and comparative experiments are carried out on the public data set. The success rate of member inference attack is 51%–52%, which is close to the random guess probability. Compared with DP-SeqGAN, the classification accuracy of the downstream model is improved by 6%–9%.