Image synthesis, a technology combining computer vision and artificial intelligence, is gaining increasing importance with the rapid advancement of large models. Among various image synthesis scenarios, generative adversarial network (GAN) is considered the most effective and widely applied technique. Currently the most popular solution is to deploy it on cloud rather than locally, as it not only saves local computing resources but also significantly improves the efficiency of data processing. Nevertheless, the convenience of this outsourcing method is counterbalanced by significant challenges to privacy protection, as there is a risk that sensitive original data or trained models could be exposed within cloud environments. To solve these issues, this work introduces a privacy-preserving self-attention GAN framework (PPAGAN) for image synthesis based on Secure Multi-Party Computation (SMPC) technology. To do that, we innovatively design a suite of efficient and secure protocols and modules using additive secret sharing technology, achieving high-quality image synthesis and ensuring user privacy. The correctness and security of our approach are rigorously established through theoretical examination. This work conducts extensive experiments on real datasets, revealing that the output of our proposed PPAGAN scheme is consistent with those of non-privacy-preserving methods. Moreover, the images synthesized by our scheme outperform those from other privacy-preserving methods in terms of common evaluation metrics, thereby validating the accuracy and effectiveness of our approach.