This is the codebase accompanying the paper Disparate Impact in Differential Privacy from Gradient Misalignment. It was accepted for a spotlight presentation in ICLR 2023 and you can check the open review.
- Install conda, pip
- Python 3.10
conda create -n FairDP python=3.10
conda activate FairDP- PyTorch 1.11.0
conda install pytorch=1.11.0 torchvision=0.12.0 numpy=1.22 -c pytorch- functorch 0.1.1
pip install functorch==0.1.1- opacus 1.1
conda install -c conda-forge opacus=1.1- matplotlib 3.4.3
conda install -c conda-forge matplotlib=3.4.3- Other requirements
conda install pandas tbb regex tqdm tensorboardX=2.2
pip install tensorboard==2.9
Scripts to reproduce experiments located at fair-dp/experiment_scripts, results saved to fair-dp/runs.
bash ./experiment_scripts/mnist_script.sh
tensorboard --logdir=runs
-
Download CelebA dataset from https://www.kaggle.com/datasets/jessicali9530/celeba-dataset and save files to fair-dp/data/celeba/
-
Download Adult dataset from https://archive.ics.uci.edu/ml/datasets/Adult and save files adult.data, adult.test to fair-dp/data/adult/
bash ./experiment_scripts/adult_script.sh
- Download Dutch dataset from https://easy.dans.knaw.nl/ui/datasets/id/easy-dataset:32357. Free registration is required on the website. Under the "Data Files" tab download all files. Unzip and save to fair-dp/data/dutch/. Full file path required is ./fair-dp/data/dutch/original/org/IPUMS2001.asc
bash ./experiment_scripts/dutch_script.sh