shifted_interpolation_dp

A repository for reproducing the numerical results in the paper "Shifted Interpolation for Differential Privacy"

Files

gdp.py : utils for calculating Gaussian tradeoff functions
plot_intro.py : code for generating Figure 1 (exemplary numerics on NoisyGD) and Figure 2 (illustration of $f$-DP and GDP)
plot_nsgd_fdp.py : code for generating Figure 10 (numerics on NoisySGD)
plot_nsgd_opt.py : code for generating Figure 4 (one-step optimality of privacy bound of NoisySGD)
plot_table_ncgd_fdp.py : code for generating Figure 8-9 and Table 9-12 (numerics on NoisyCGD)
plot_table_ngd_fdp.py : code for generating Figure 6-7 and Table 7-8 (numerics on NoisyGD)
prv_SymmPoissonSubsampledGaussianMechanism.py : code for calculating privacy of subsampled Gaussian mechanism (see Appendix D.5)
table_acc_lr.py : code for generating Table 2, 5-6 (train and test accuracy of the experiment on regularized logistic regression)
table_privacy_lr.py : code for generating Table 1, 3-4 (privacy of the experiment on regularized logistic regression)
MNISTdata.hdf5 : MNIST data
requirements.txt : required packages

The numbers on tables and figures are based on the latest arXiv version of the paper.
The formats of tables and figures obtained from the code may have been modified for better presentation in the paper.

The code for regularized logistic regression table_acc_lr.py is adapted and modified from https://github.com/yawen-d/Logistic-Regression-on-MNIST-with-NumPy-from-Scratch.
The code for calculating the privacy of subsampled Gaussian mechanism prv_SymmPoissonSubsampledGaussianMechanism.py is designed to work within the PRV framework of https://github.com/microsoft/prv_accountant.