SABAF: Removing Strong Attribute Bias from Neural Networks with Adversarial Filtering

Find our paper at arXiv and the short-version workshop paper at Algorithmic Fairness through the Lens of Time, NeurIPS 2023 and arXiv. Please cite the following if using the code:

Long paper:

@article{li2023sabaf,
  title={SABAF: Removing Strong Attribute Bias from Neural Networks with Adversarial Filtering},
  author={Li, Jiazhi and Khayatkhoei, Mahyar and Zhu, Jiageng and Xie, Hanchen and Hussein, Mohamed E and AbdAlmageed, Wael},
  journal={arXiv preprint arXiv:2311.07141},
  year={2023}
}

Workshop paper at Algorithmic Fairness through the Lens of Time, NeurIPS 2023 (see poster and presentation at https://nips.cc/virtual/2023/77745):

@article{li2023information,
  title={Information-Theoretic Bounds on The Removal of Attribute-Specific Bias From Neural Networks},
  author={Li, Jiazhi and Khayatkhoei, Mahyar and Zhu, Jiageng and Xie, Hanchen and Hussein, Mohamed E and AbdAlmageed, Wael},
  journal={arXiv preprint arXiv:2310.04955},
  year={2023}
}

Setup

This repo uses the packages described in the requirements.txt file. To set the environment, simply run:

pip install -r requirements.txt

Datasets

• Colored MNIST dataset

  • Download or set --color_dataset generated to generate Colored MNIST dataset with arbitrary color variance --biased_var BIASED_VAR

• CelebA dataset

  • Download
  • Place images in ./data/CelebA/raw_data/img_align_celeba/*.jpg and other files in ./data/CelebA/raw_data/
  • Preprocess data by python prepare_data.py --dataset CelebA

• Adult Income dataset

  • Download
  • Place data in ./data/Adult/raw_data/adult.csv
  • Preprocess data by python prepare_data.py --dataset Adult

• IMDB Face dataset

  • Download
  • Place data in ./data/IMDB/raw_data/ and unzip auxiliary.zip into ./data/IMDB/raw_data/
  • Preprocess data by python prepare_data.py --dataset IMDB

• FFHQ dataset

  • Download
  • Place data in ./data/FFHQ/raw_data/images and labels in ./data/FFHQ/raw_data/labels

Custom

In addition to the datasets mentioned above, you can train and apply adversarial filter on other datasets of interest.

Dataset

Place data in ./data/Custom/raw_data. Write the methods __len__ and __getitem__ in ./dataloader/Custom.py, and set_data in ./models_train/Custom_filter.py and ./models_eval/set_Custom.py.

Training

To train adversarial dataset on the custom dataset, simply run:

python train.py --experiment Custom_filter --name NAME 

Applying

To apply adversarial filter on any interested downstream tasks, simply run:

python eval.py --experiment Custom_downstream_our --name NAME

Experiments

The pre-defined experiments can be found in run.sh. You can either directly run these pre-defined experiments with chosen hyper-parameters to reproduce results in the paper, or train and evaluate your model.

Training

To train adversarial filter, please use train.py. To see all command-line accessible arguments run:

python train.py --help

Colored MNIST dataset

To train adversarial filter on universal dataset, simply run:

python train.py --experiment CMNIST_filter --name NAME --biased_var -1 \
    --mi 10.0 --gc 100.0 --dc 100.0 --gr 100.0 

CelebA dataset

To train adversarial filter on universal dataset, simply run:

python train.py --experiment CelebA_filter --name NAME --CelebA_train_mode CelebA_FFHQ \
    --mi 50.0 --gc 50.0 --dc 50.0 --gr 100.0 \
    --shortcut_layers 1 --inject_layers 1 --enc_layers 5 --dec_layers 5 --dis_layers 5

Adult Income dataset

To train adversarial filter on universal dataset, simply run:

python train.py --experiment Adult_filter --name NAME --Adult_train_mode all \
    --mi 5.0 --gc 5.0 --dc 5.0 --gr 10.0 --epochs 100

IMDB Face dataset

To train adversarial filter on universal dataset, simply run:

python train.py --experiment IMDB_filter --name NAME --IMDB_train_mode all \
    --mi 50.0 --gc 50.0 --dc 50.0 --gr 100.0 \
    --shortcut_layers 1 --inject_layers 1 --enc_layers 5 --dec_layers 5 --dis_layers 5

Evaluation

To evaluate adversarial filter, please use eval.py. To see all command-line accessible arguments run:

python eval.py --help

To load the trained filter for evaluation, you can:

either (1) use --filter_name FILTER_NAME --filter_hp FILTER_HP --filter_idx FILTER_IDX. For example, by indicating --filter_name reproduce --filter_hp mi10.0_gc100.0_dc100.0_gr100.0 --filter_idx 19, the filter which is saved as result/reproduce/mi10.0_gc100.0_dc100.0_gr100.0/weights.19.pth will be loaded.

or (2) use --filter_path FILTER_PATH to explicitly indicate the absolute path of filter.

Colored MNIST dataset

To reproduce the results of adversarial filter in the extreme bias point, simply run:

python eval.py --experiment CMNIST_downstream_our --name NAME --biased_var 0 \
    --filter_train_mode universal 

To reproduce the results of baseline model in the extreme bias point, simply run:

python eval.py --experiment CMNIST_downstream_baseline --name NAME --biased_var 0

CelebA dataset

To reproduce the results of adversarial filter in the extreme bias point, simply run:

python eval.py --experiment CelebA_downstream_our --name NAME --CelebA_train_mode CelebA_train_ex \
    --attributes Blond_Hair --CelebA_test_mode unbiased_ex --filter_train_mode universal
python eval.py --experiment CelebA_downstream_our --name NAME --CelebA_train_mode CelebA_train_ex \
    --attributes Blond_Hair --CelebA_test_mode conflict_ex --filter_train_mode universal

To reproduce the results of baseline model in the extreme bias point, simply run:

python eval.py --experiment CelebA_downstream_baseline --name NAME \
    --CelebA_test_mode unbiased_ex --attributes Blond_Hair
python eval.py --experiment CelebA_downstream_baseline --name NAME \
    --CelebA_test_mode conflict_ex --attributes Blond_Hair

Adult Income dataset

To reproduce the results of adversarial filter in the extreme bias point, simply run:

python eval.py --experiment Adult_downstream_our --name NAME \
--Adult_train_mode eb1_balanced --Adult_test_mode eb2_balanced --filter_train_mode universal
python eval.py --experiment Adult_downstream_our --name NAME \
--Adult_train_mode eb1_balanced --Adult_test_mode balanced --filter_train_mode universal

To reproduce the results of baseline model in the extreme bias point, simply run:

python eval.py --experiment Adult_downstream_baseline --name NAME \ 
--Adult_train_mode eb1_balanced --Adult_test_mode eb2_balanced 
python eval.py --experiment Adult_downstream_baseline --name NAME \ 
--Adult_train_mode eb1_balanced --Adult_test_mode balanced 

IMDB Face dataset

To reproduce the results of adversarial filter in the extreme bias point, simply run:

python eval.py --experiment IMDB_downstream_our --name NAME \ 
--IMDB_train_mode eb1 --IMDB_test_mode eb2 --filter_train_mode universal
python eval.py --experiment IMDB_downstream_our --name NAME \ 
--IMDB_train_mode eb1 --IMDB_test_mode unbiased --filter_train_mode universal
python eval.py --experiment IMDB_downstream_our --name NAME \ 
--IMDB_train_mode eb2 --IMDB_test_mode eb1 --filter_train_mode universal
python eval.py --experiment IMDB_downstream_our --name NAME \ 
--IMDB_train_mode eb2 --IMDB_test_mode unbiased --filter_train_mode universal

To reproduce the results of baseline model in the extreme bias point, simply run:

python eval.py --experiment IMDB_downstream_baseline --name NAME \ 
--IMDB_train_mode eb1 --IMDB_test_mode eb2
python eval.py --experiment IMDB_downstream_baseline --name NAME \ 
--IMDB_train_mode eb1 --IMDB_test_mode unbiased
python eval.py --experiment IMDB_downstream_baseline --name NAME \ 
--IMDB_train_mode eb2 --IMDB_test_mode eb1
python eval.py --experiment IMDB_downstream_baseline --name NAME \ 
--IMDB_train_mode eb2 --IMDB_test_mode unbiased