Differentially Private Generative Adversarial Networks for Time Series, Continuous, and Discrete Open Data

Abstract

Open data plays a fundamental role in the 21th century by stimulating economic growth and by enabling more transparent and inclusive societies. However, it is always difficult to create new high-quality datasets with the required privacy guarantees for many use cases. This paper aims at creating a framework for releasing new open data while protecting the individuality of the users through a strict definition of privacy called differential privacy. Unlike previous work, this paper provides a framework for privacy preserving data publishing that can be easily adapted to different use cases, from the generation of time-series to continuous data, and discrete data; no previous work has focused on the later class. Indeed, many use cases expose discrete data or at least a combination between categorical and numerical values. Thanks to the latest developments in deep learning and generative models, it is now possible to model rich-semantic data maintaining both the original distribution of the features and the correlations between them. The output of this framework is a deep network, namely a generator, able to create new data on demand. We demonstrate the efficiency of our approach on real datasets from the French public administration and classic benchmark datasets.

DP-CGAN: Differentially Private Synthetic Data and Label Generation

“…Moreover, the datasets used to evaluate the model are small. The other work is a DPGAN framework for time series, continuous, and discrete data [29]. This framework is alike the previous DPGAN work [28] except it employs moments accountant approach to account the privacy budget and clips the discriminator gradients while reducing the clipping parameter over time (adaptive clipping).…”

“…In contrast to PATE-GAN [30] which generates only binary labels, our model generates multi-class labels. Finally, in DPGAN frameworks [28,29] the discriminator gradients are clipped and perturbed by adding Gaussian noise to gradients of the discriminator loss, while in our framework, Gaussian noise is added to the accumulation of clipped gradients of discriminator loss on real data and clipped gradients of discriminator loss on fake data.…”

“…Differential Privacy (DP) [26,27] is a common technique to protect the privacy of ML models trained on sensitive data. However, in spite of its popularity, there have been very few recent studies on training GANs in a differentially private way [28][29][30][31][32]. The standard procedure leveraged by these recent studies to enforce DP is to first clip the l2 norm of the gradients of the sum of the discriminator's loss on real and fake data and then add Gaussian noise to the clipped gradients.…”

Privacy considerations for sharing genomics data

“…Other work has focussed on private deep learning in general (Abadi et al, 2016[ 1 ]) and GANs in particular (Xie et al, 2018[ 79 ]; Beaulieu-Jones et al, 2019[ 4 ]; Frigerio et al, 2019[ 31 ]; Torkzadehmahani et al, 2019[ 72 ]) over the past years, among those a recent publication that addresses privacy issues when sharing sensitive data or when sharing generators trained on such data and proposes the training and release of differentially private generators instead (Chen et al, 2020[ 12 ]). They further illustrate how the approach can naturally adapt to a federated learning setting.…”

Deepfake Forensics, an AI-synthesized Detection with Deep Convolutional Generative Adversarial Networks

“…Jones proposed a differential privacy assisted classification generation countermeasure network combining GANs and differential privacy to generate medical clinical data [50]. Frigerio proposed a data publishing framework to protect privacy through the definition of differential privacy [51], to ensure the protection of user personality while publishing new open data. Abadi used GANs to replace the communication parties and adversaries in the traditional symmetric encryption system with neural networks to protect the communication process [52].…”