DP-TabImage: a differentially private framework for paired table-image multimodal synthesis

DP-TabImage is a differentially private framework for paired table-image multimodal synthesis, using a privacy probabilistic graphical model to model the table distribution and a table-conditioned diffusion model trained with DP-SGD to generate images. To assist conditional learning under clipped and noised gradients, the framework is pretrained on private table-image prototypes without adding privacy cost. Experiments on three real datasets show that the framework achieves a good balance among table fidelity, image fidelity, and cross-modal alignment.