zliangak/dp-fed-ls

Differentially Private Federated Learning

Exploring Private Federated Learning with Laplacian Smoothing

This repository contains the source code of the paper

Liang, Z., Wang, B., Gu, Q., Osher, S. and Yao, Y., 2020. Exploring private federated learning with laplacian smoothing. arXiv preprint arXiv:2005.00218

It includes training differentially-private federated logistic regression over MNIST, CNN over SVHN and LSTM over Shakespeare dataset, with uniform or Poisson subsampling.

Dependency

The code is run with Python==3.8.8, torch==1.8.0, torchvison==0.9.0

Dataset

For Shakespeare dataset, we went through the preprocessing steps and placed the result into a json file. It is a python dict object with keys ['users', 'user_data', 'num_samples'], which represent roles' names, their corresponding samples (lines) and sample numbers.

Moment Accountants

Moment Accountant for uniform subsampling in dp/autodp is borrowed from this repo. And the one for poisson subsampling in dp/google_accountant is borrowed from this repo.

Experiment in the paper

For the experiments mentioned in the paper, one can find the bash script in bash/bash.sh.