End-to-End Federated Diffusion Generative Models for Tabular Data

Tabular data is widely used in various fields and applications, making the synthesis of such data an active area of research. One important aspect of this research is the development of methods for privacy-preserving data synthesis, which aims to generate synthetic data that retains statistical properties while protecting the privacy of individuals in the dataset. Recently, Diffusion Generative Models, such as Gaussian Diffusion Model, have significantly improved image synthesis, but their effectiveness in synthesizing tables is limited, because of using One-Hot encoding for representing categorical attributes with many categories. Furthermore, it needs the private data to be centrally collected for training, thus violating the privacy-preserving criteria. In this paper, we propose a new decentralized tabular synthesizing framework, which has three key features: (i) a decentralized Autoencoder comprised of an encoder and a decoder to map discrete features into the continuous space and back, (ii) a tabular diffusion model trained in a decentralized manner and (iii) incorporating differential privacy on central stochastic gradient training. We conduct extensive experimental studies that focus on sampling quality and diversity, using 9 tabular datasets and 4 state-of-the-art synthesizers. The results show that our method outperforms existing central methods by 10.7% and 31.4% in data quality and diversity on average, and 6.8% and 21.1% in data quality and diversity in scenarios facing non-IID data.Computer Scienc