Differentially private data synthesis method based on nonlinear mapping embedding

This paper introduces a synthetic data generation method Privnpgd based on private-PGD. private-PGD is a differential data generation method that utilizes optimal transmission and particle gradient descent to effectively fit the probability distribution of noisy data. It is superior to other synthetic data models in handling large data sets. Our research focuses on the embedding phase and particle gradient descent phase of attribute fields, where the percentiles of attribute fields are calculated and the discrete properties are mapped to these percentiles using nonlinear functions, and the stability and convergence of particle renewal are improved using energy function-based slice distributions. This method effectively captures the non-uniform distribution of data while preserving the sequence information of discrete attributes. It generates synthetic data on some discretized datasets and outperforms both private-pgd and existing methods.