This repository contains implementations of the experiments presented in the paper, together with additional experiments for analyzing the algorithm and its hyperparameters.
Contains the one-dimensional experiments: These experiments evaluate the estimators and privacy mechanisms on several one-dimensional distributions and experimental configurations.
Contains the logistic regression experiments: These experiments evaluate the proposed methods in a higher-dimensional setting using logistic regression.
Contains experiments that examine the effect of varying the main algorithmic hyperparameters:
-
$T$ - the number of resampling iterations. -
$m(n))$ - the resampled subset size as a function of the original dataset size$n$ . -
$q$ - the fraction of the privacy budget allocated through the epsilon split.
These experiments are intended to illustrate how the hyperparameter choices affect the accuracy, stability, and privacy-utility trade-off of the estimators.
Contains the experiment configurations and scripts used to generate the figures reported in the paper. The experiments corresponding to all paper figures are organized and documented in this directory.