Deep Learning Project: Differentially Private Releasing via Deep Generative Model

Alexandre Huat (INSA Rouen Normandie, Dept. Information Systems Architectures, Data Science MSc by Research)

---

This repository contains my implementation of dp-GAN (Differentially Private Generative Adversarial Network) for my deep learning project assessment. It is organized as follows:

• Directory summary contains the report of the project in French only. This document consists in (i) a summary of the original paper of dp-GAN and (ii) a report on my implementation. Its PDF version has been precompiled, but run ./compile.sh from summary if needed.
• Directory dpgan contains the implementation of the project. See the report in summary or the docstrings of each files to understand it.
• Directory data contains all relevant data for the use of the implemented neural network.

All other useful information can be found in summary/summary.pdf.

Requirements

In a Python 3 virtual environment, run pip install -r requirements.txt.

The test of dp-GAN requires the MNIST dataset. Since it needs a prior download from Keras, its first run could be long.

Reference

X. Zhang, S. Ji and T. Wang, "Differentially Private Releasing via Deep Generative Model", ArXiv e-prints, jan. 2018. arXiv : 1801.01594 [cs.CR].