This package implements numerical methods for calculating the differential privacy of variants of Gaussian mechanism in the shuffle model using Rényi differential privacy (RDP). The following implementations are available:
SubGaussRDPtoDP: Subsampled Gaussian mechanismShuffGaussRDPtoDP: Shuffle Gaussian mechanismSubShuffGaussRDPtoDP: Subsampled Shuffle Gaussian mechanism
The following faster (in the sense that, binary search is used instead of computing a pre-determined list of moment) implementations are available in beta version:
FastShuffGaussRDPtoDP: Shuffle Gaussian mechanismFastSubShuffGaussRDPtoDP: Subsampled Shuffle Gaussian mechanism
Our implementation particularly allows fast comparisons of (epsilon, delta) at different numbers of composition. See Usage.
To install,
git clone https://github.com/spliew/shuffgauss
cd shuffgauss
pip install -e .This code supports Python 3.8 and newer. See pyproject.toml for other requirements.
import shuffgauss as sg
# setting up parameters
sigma = 1 # gauss std deviation
n = 6e5 # total number of users
m = 6e4 # expected no. of subsampled users. we assume sampling_rate = m/n
delta = 1/n # differential privacy parameter
mxlmbda = 20 # maximum rdp moment to calculate
# shuffle gaussian mechanism
sf = sg.ShuffGaussRDPtoDP(sigma, n, mxlmbda)
sf.get_shuff() # prepare the calculation
print(sf.get_eps(delta, 1)) # calculate epsilon when no of composition is 1, return a tuple of (epsilon, lmbda_min)
print(sf.get_eps(delta, 10)) # calculate epsilon when no of composition is 10If you use this code in your work, please cite our paper:
@misc{liew2023privacy,
title={Privacy Amplification via Shuffled Check-Ins},
author={Seng Pei Liew and Satoshi Hasegawa and Tsubasa Takahashi},
year={2023},
eprint={2206.03151},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
This code is heavily influenced by autodp.