DP-GAN: Differentially private consecutive data publishing using generative adversarial nets

Abstract In the era of big data, increasingly massive volumes of data is generated and published consecutively for both research and commercial purposes. The potential value of sensitive information also attracts interest from adversaries and thereby arises public concern. Current research mostly focuses on privacy-preserving data publishing in a statistic manner rather than taking the dynamics and correlation of context into consideration. Motivated by this, we propose a novel idea that combining differential privacy and generative adversarial nets. Generative adversarial nets and its extensions are used to generate a synthetic data set with indistinguishable statistic features while differential privacy guarantees a trade-off between the privacy protection and data utility. By employing a min-max game with three players, we devise a deep generative model, namely, DP-GAN model, for synthetic data generation while fulfilling the privacy constraints in a differentially private manner. Extensive simulation results on a real-world data set testify the superiority of the proposed model in terms of privacy protection, data utility, and efficiency.