D-PGAN: a differentially private generative adversarial network for multi-party publishing of high-dimensional tabular data

The rapid growth of data-driven applications has produced massive high-dimensional tabular data horizontally partitioned across independent institutions. Under strict privacy regulations and limited trust in third parties, jointly publishing such data while preserving both individual privacy and analytical utility remains an open challenge. Existing differentially private generative adversarial networks for multi-party publishing typically aggregate discriminator feedback with uniform weights and apply static gradient perturbation, leading to imbalanced fitting under heterogeneous distributions and excessive utility loss under tight privacy budgets. This paper proposes D-PGAN, a differentially private generative adversarial framework for multi-party publishing of high-dimensional data. D-PGAN adopts a single central generator with per-party local discriminators, keeping raw data within each party. A global attention mechanism adaptively weights each party’s feedback so the generator focuses on harder-to-fit distributions. Via chain-rule decomposition of the generator gradient, only the discriminator-side sensitive factor is perturbed; the discriminator is then updated by a differentially private Adam optimizer with density-clustering-based adaptive clipping and a square-root decaying noise schedule. At publishing, a Bayesian-posterior cooperative filter rejects samples on which the discriminators disagree, improving utility at no additional privacy cost. Theoretical analysis shows that D-PGAN satisfies \((\epsilon ,\delta )\) -differential privacy with respect to each party’s local data. Experiments on the Adult, TPC-E, KDD99, and NLTCS datasets confirm its superior statistical similarity, downstream accuracy and cross-party fairness, and its robustness under tight privacy budgets and larger numbers of parties.