/fair-unlearning

Fair Machine Unlearning: Data Removal while Mitigating Disparities

Primary LanguagePythonApache License 2.0Apache-2.0

Fair Machine Unlearning: Data Removal while Mitigating Disparities

Paper

The official code to replicate the results found in Fair Machine Unlearning: Data Removal while Mitigating Disparities

Usage

To unlearn at random, execute the following script. --protected-attribute determines what feature to measure fairness metrics over, and --std deterrmines the noise added to make the method unlearnable.

python unlearning_disparity_experiments.py --dataset "COMPAS" --protected-attribute "race" --reg 1e-4 --std 1e0

To unlearn from a specific subgroup, include the following flags specifying --unlearned-attribute and --value:

python unlearning_disparity_experiments.py --dataset "Adult" --protected-attribute "race" --reg 1e-4 --std 1e0 --unlearned-attribute "race" --value 0

Finally, to test on fair or unfair loss functions, include the --fair flag and a corresponding fairness regularizer value:

python unlearning_disparity_experiments.py --dataset "Adult" --protected-attribute "race" --reg 1e-4 --std 1e0 --unlearned-attribute "race" --value 0 --fair --fairreg 1e0

eps_delta_experiments.py computes accuracy while sweeping over pairs of $\epsilon, \delta$ to construct unlearning trade-off curves.

fairness_tradeoff_experiments.py computes accuracy while sweeping over $\gamma$ to construct fairness trade-off curves.

API

The main functions for fair unlearning, and replication of previous methods are all in fair_retraining.py. The other scripts are infrastructure to run and replicate experiments.s