In recent years, differential privacy techniques have become the de facto standard for privacy protection. With rigorous mathematical proofs, it only requires to add a small amount of noise to the deterministic functions to achieve privacy protection. However, most existing data generation methods based on differentially private GANs (Generative Adversarial Networks) are very difficult to train. Especially for privacy-preserving image generation, existing methods are difficult to generate high-quality results. We rethink the noise addition in differentially private GAN and design a self-attention differentially private GAN (DP-SAGAN). Even working in the case of privacy preservation, our DP-SAGAN still scores comparable to the current mainstream models in IS and FID metrics, and also generates higher resolution images stably.