Revisiting Model Fairness via Adversarial Examples

Authors: Tao Zhang, Tianqing Zhu, Jing Li, Wanlei Zhou

This is the code for my paper 'Revisiting Model Fairness via Adversarial Examples'.

Requirements

To install requirements:

pip install -r requirements.txt

Datasets

• Download .csv files for 4 datasets used in the paper from data directory

Preprocess

• Data preprocessing is implemented in the prepare_data.py.

Training

• Two adversarial attack methods, LowProFool and DeepFool, are implemented in the Adverse.py.

• Training and evaluation are implemented in the train_model.py.

• Demographic parity is implemented in the Fairness_metrics.py.

Evaluation

• All evaluation metrics are implemented in the Metrics.py.

Results

• We provide an example to train models and obtain experimental results on the German dataset with sensitive attribute age in the Playground_German.ipynb.