Differential Privacy has Bounded Impact on Fairness in Classification

This archive contains the code related to the paper "Differential Privacy has Bounded Impact on Fairness in Classification". It contains everything needed to reproduce all the figures from the paper.

Getting the Data

The following instructions describe how to download the data to reproduce our experiments. Note that celebA takes about 20MB and folktables about 80MB of space on the disk.

For celebA

Go to http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html and download the "Attributes Location" dataset. To use the file, you must first format it as required by

cat celebA.txt | sed "s/ /,/g" | sed "s/  */,/g" > celebA_preprocessed.csv

Then, save the celebA_preprocessed.csv file in a data directory.

For folktables

To get the folktables dataset, simply run

python folktables_download.py

Running the Experiments

The file compute_bounds.py contains the code that effectively computes the bounds we use in the paper. The two scripts expe_fairness_fct_epsilon.py and expe_fairness_fct_n.py run all experiments and save the results in a results directory. To run all experiments on all datasets, run

./run_all_experiments.sh

Making the Plots

Finally, to make the plots, simply run

./make_all_plots.sh

Plots will be saved in a plots directory. Note that Table 1 may be reproduced by running