This project contains scripts to reproduce experiments from the paper Deep Learning with Gaussian Differential Privacy by Zhiqi Bu, Jinshuo Dong, Weijie Su and Qi Long.

Deep learning models are often trained on datasets that contain sensitive information such as individuals' shopping transactions, personal contacts, and medical records. Many differential privacy definitions arise for the study of trade-off between models' performance and privacy guarantees. We consider a recently proposed privacy definition termed f-differential privacy (https://arxiv.org/abs/1905.02383) for a refined privacy analysis of training neural networks. Using GDP instead of (epsilon,delta)-DP, we can get much better privacy guarantee and alternatively, trade off some privacy for better accuracy.

You need to install Pytorch privacy package pytorch-dp to run the following codes. This can be done easily by

pip install pytorch-dp

However, for latest version of Pytorch privacy package opacus, you should do

pip install opacus

---------------- Update on 2021/4/28-----------------

mnist.py: private CNN on MNIST

adult.py: private FFNN on Adult data

Note that the Embedding Layer can be used in private training in Tensorflow but not in Pytorch.

gdp_accountant.py computes the moments accountant (MA), central limit theorem (CLT) and dual relation (Dual) between \delta,\epsilon,\mu. This computation does not have any TensorFlow dependencies and is data-independent, and thus is extremely fast.

For example, if you run MNIST dataset (60,000 samples) for 15 epochs, batch size 256, noise level 1.3 and \delta 1e-5, then your \epsilon is

By Moments Accountant:

compute_epsilon(15,1.3,60000,256,1e-5)=1.1912

%% this has been improved by Tensorflow-privacy using new Moments Accountant to epsilon = 0.9545.

By GDP CLT (Uniform subsampling):

compute_epsU(15,1.3,60000,256,1e-5)=1.0685

By GDP CLT (Poisson subsampling):

compute_epsP(15,1.3,60000,256,1e-5)=0.8345

For example, MNIST in Figure 4 of our paper shows that our GDP CLT is both more accurate and more private then existing MA method.