This repo provides the code for our paper "On Adversarial Bias and the Robustness of Fair Machine Learning".

We design and implement data poisoning attack algorithms (algorithm1 and algorithm2) against machine learning models trained with equalized odds as the fairness constraint to reduce the overall test accuracy. We assume the attacker who can control the sampling process and (in the stronger case, also) the labeling process for some of the training data. Using these algorithms, we show that in presence of adversarial bias, fairness of machine learning models can be in conflict with robustness.

• Require: python 3.6

  1. Install virtualenv: pip install virtualenv

  2. Create a virtual environment

     • mkdir ~/env/ # this directory contains all virtual environment
     • virtualenv -p python3 ~/env/(name) # replace (name) with your naming of the virtual environment

  3. Install packages

     • source ~/env/(name)/bin/activate # activate the environment
     • pip install -r requirements.txt

• Note: For reproducibility, we include fairlearn repo (Microsoft, v0.3.0).

• We evaluate the code on the COMPAS and Adult dataset. We generate 4 datasets in the dataset folder based on the preprocessing step which we mentioned in the paper. Each dataset includes:

• x_train,y_train clean training dataset.

• x_noise,y_noise hard examples from nature.

• x_attacker, y_attacker attacker dataset which is from the same distribution as clean training dataset but no overlap.

• x_test, y_test test dataset.

• Implementation of Algorithm 1 and 2 are in attacks.py. Call the corresponding functions and provide required arguments to run the attacks.
• Follow Example notebook to see an example.

To cite the arxiv version, please use the following bibtex

@article{chang2020adversarial,
  title={On Adversarial Bias and the Robustness of Fair Machine Learning},
  author={Chang, Hongyan and Nguyen, Ta Duy and Murakonda, Sasi Kumar and Kazemi, Ehsan and Shokri, Reza},
  journal={arXiv preprint arXiv:2006.08669},
  year={2020}
}