Subsampling is not Magic: Why Large Batch Sizes Work for Differentially Private Stochastic Optimisation

This repository contains the code and Mathematica notebooks for the "Subsampling is not Magic: Why Large Batch Sizes Work for Differentially Private Stochastic Optimisation" (ICML 2024) paper.

Installing dependencies

Install Poetry following the instructions here.

Run poetry install --without dev to install the dependencies.

Reproducing experiments

cd ./src

Run sh ./plot.sh