Distributed GAN with Swift Learning Mechanism for Scalable Multi-Party Tabular Data Synthesis

In modern contexts, collaborative efforts among multiple parties generate large datasets, enhancing information and accuracy. However, privacy preservation becomes a critical issue when handling sensitive data. Additionally, the development of generative models in distributed environments requires significant computational resources. Generating synthetic data from such collaborations is a major challenge. This study introduces a fast learning mechanism in a distributed generative adversarial network (GAN) for data synthesis in multi-party settings, where all parties share a common table schema. The model uses a single generator and distributed discriminators, each aligned with its dataset. During each epoch, discriminators are randomly selected to update the generator gradient, similar to a GAN with a single discriminator. We tested our model on two benchmark datasets for multi-party data synthesis. Results show that our model is significantly more efficient than existing distributed GAN models, being about 10 times faster than those using federated learning and 2-3 times faster than models with distributed discriminators, without compromising machine learning performance. These results highlight the model’s efficiency and potential for practical multi-party data synthesis applications.