DiPPS: Differentially Private Propensity Scores for Bias Correction

Introduction

This is the implementation of the differentially private data collection mechanism DiPPS proposed in the paper Differentially Private Propensity Scores for Bias Correction and the experiments to compare it with other methods.

Prerequisites

• PyTorch 1.7.1 conda install pytorch torchvision cudatoolkit=10.1 -c pytorch

• numpy 1.19.2 conda instal -c anaconda numpy

• pandas 1.2.2 conda install -c anaconda pandas

• scikit-learn 0.23.2 conda install -c conda-forge scikit-learn

• R package 'transport' wget https://cran.r-project.org/src/contrib/transport_0.12-2.tar.gz

Code structure

.
├── run.py                              # The main script to run the experiments to estimate mean/variance/median or other target functions on datasets
├── data/                               # The datasets we use for the experiments
├── preprocess/                         # The scripts to preprocess the original datasets and to obtain the data split into participant and non-participant data
├── config/                             # The configurations of the experiments, including the model parameters
├── exponential_mechanism.py            # The implementation of the *exponential mechanism*
├── invert_exponential_mechanism.py     # The implementation of the estimation of group statistics based on data collected under the *exponential mechanism*
├── hybrid_mechanism.py                	# The implementation of the *Hybrid mechanism*
├── laplace_mechanism.py                # The implementation of the *Laplace mechanism*
├── LCIW.py                     	# The implementation of *PS*, i.e., the propensity score based estimation method proposed in the paper. Combined with the exponential mechanism, it turns into DiPPS
├── estimation.py                    	# The summarization script to estimate the target variable based on all the methods
├── computeWassersteinDist.r            # The helper script to compute the Wasserstein distance. It is based on the R package 'transport', which is mentioned in *Prerequisites*
├── view_results.ipynb          	# Summarize statistics of experiments
└── 
...

Run experiments

• bash run.sh The experiments can be configured via parameters in the run script itself, which are documented therein.

• The experiments in the paper are obtained with the 5 random seeds [1111, 2222, 3333, 4444, 5555]. One can set the seed in run.sh directly.