milkigit / differential-privacy-vs-fairness

Code for "Differential Privacy Has Disparate Impact on Model Accuracy" NeurIPS'19

Readme

The paper discusses how Differential Privacy (specifically DPSGD from [1]) impacts model performance for underrepresented groups.

Usage

Configure environment by running: pip install -r requirements.txt

We use Python3.7 and GPU Nvidia TitanX.

File playing.py allows run the code. It uses utils/params.yaml to set parameters from the paper and builds a graph on Tensorboard. For Sentiment prediction we use playing_nlp.py.

Datasets:

1. MNIST (part of PyTorch)
2. Diversity in Faces (obtained from IBM here)
3. iNaturalist (download from here)
4. UTKFace (from here)
5. AAE Twitter corpus (from here)

We use compute_dp_sgd_privacy.py copied from public repo

DP-FedAvg implementation is taken from public repo

Implementation of DPSGD is based on TF Privacy repo and papers:

[1] M. Abadi, A. Chu, I. Goodfellow, H. B. McMahan, I. Mironov, K. Talwar, and L. Zhang. Deep learning with differential privacy. In CCS, 2016.

[2] H. B. McMahan and G. Andrew. A general approach to adding differential privacy to iterative training procedures. arXiv:1812.06210, 2018

[3] H. B. McMahan, D. Ramage, K. Talwar, and L. Zhang. Learning differentially private recurrent language models. In ICLR, 2018