This repository contains the implementation for: Differentially Private GAN for Time Series
Generative Adversarial Networks (GANs) are a modern solution aiming to encourage public sharing of data, even if the data contains inherently private information, by generating synthetic data that looks like, but is not equal to, the data the GAN was trained on. However, GANs are prone to remembering samples from the training data, therefore additional care is needed to guarantee privacy. Differentially Private (DP) GANs offer a solution to this problem by protecting user privacy through a mathematical guarantee, achieved by adding carefully constructed noise at specific points in the training process. A state-of-the-art example of such a GAN is Gradient Sanitized Wasserstein GAN (GS-WGAN) [1] . This model is shown to create higher quality synthetic images than other DP GANs. To extend the applicability of GS-WGAN we first reproduce and extend the evaluation, verifying that the model outperforms DP-CGAN by an average of 40% when assessed across three qualitative metrics and two datasets. Secondly we propose improvements to the architecture and training procedure to make GS-WGAN applicable on timeseries data. The experimental results show that GS-WGAN is fit for generating synthetic timeseries through promising experimental results.
[1] D. Chen, T. Orekondy, and M. Fritz, “Gs-wgan: A gradient-sanitized approach for learning differentially private generators,” 2021.
This repo contains code from GS-WGAN, and DP-CGAN
Implementation taken from the author's implemenation with minor changes to the model to accept images with different sizes than 28x28 pixels, and added support for PTBand MIT-BIH datasets, both available on kaggle.
Code taken from DP-MERF, which is the same implementation used in the GS-WGAN evaluation. (link to git repo)
Contains a script that can compute Inception Score, Frechet Inception Distance and downstream classifier accuracy for a given GS-WGAN generator (.pth file) or a numpy archive containing generated (Fashion-)MNIST samples
To train GAN training either follow instructions in the corresponding README's in the model subfolders, or use the docker configurations supplied in docker During training intermediate samples and generators are saved.
They can be run using docker as follows:
- Make sure Docker and docker-compose are installed, then do:
cd docker/[DP-CGAN|GS-WGAN|eval]/
docker-compose up
This starts a docker container which automatically install all dependencies for a run on CPU.
For running without Docker it is easiest to create a new conda environment and install dependencies for the model you want to run. (can be found in docker/[GS-WGAN|DP-CGAN|eval]/requirements.txt) Or in the respective model's README.
!
During a training run intermediate samples (for DP-CGAN) or generators (for GS-WGAN) will be saved.
Once training is finished the evaluator can be run via docker or via commandline, an example command:
python evaluator.py --generator_dir ../../resources/gswgan/fashionmnist/gen --gen_data_size 10000 --dataset fashionmnist --save_location ../../resources/gswgan/fashion/ --num_batches 10
This computes Inception Score, Frechet Inception Distance and downstream classifier accuracy for all generator or sample files in the --generator_dir
- Refer to
eval/evaluator/evaluator.py, or runpython evaluator.py -hfor the complete list of arguments.
For the IS you first need to train a classifier (seperate for MNIST & FashionMNIST, change in the MLP_for_Inception.py file) by running:
python eval/evaluator/MLP_for_Inception.py This trains a classifier (depending on the selected dataset) with test set accuracy of 98% or 91% accuracy, respectively for MNIST and Fashion-MNIST.
Line 26 & 27 (FASHION_MNIST_PATH&MNIST_PATH) Should point to these classifiers if you want to compute the IS.
This GS-WGAN implementation supports two new datasets: PTB & MIT-BIH. For PTB & MIT-BIH set Kaggle keys in the environment:
export KAGGLE_USERNAME=username
export KAGGLE_KEY=key
The datasets will then be automatically downloaded upon training start.
or download the datasets from kaggle: https://www.kaggle.com/shayanfazeli/heartbeat, and create a new folder: resources/data/ecg/ and put the .csv's there.