Debias with Backdoor (DwB)

Official Implementation for Paper "Backdoor for Debias: Mitigating Model Bias with Backdoor Attack-based Artificial Bias"

Authors

• Shangxi Wu (Beijing Institute of Technology)
• Qiuyang He (Beijing Institute of Technology)
• Jian Yu (Beijing Institute of Technology)
• Jitao Sang (Beijing Institute of Technology)

Citation

@article{wu2024backdoor,
  title={Backdoor for Debias: Mitigating Model Bias with Backdoor Attack-based Artificial Bias},
  author={Wu, Shangxi and He, Qiuyang and Yu, Jian and Sang, Jitao},
  journal={arXiv preprint arXiv:2303.01504},
  year={2024}
}

Requirements

Environment

• Python 3.7+
• TensorFlow 2.10+
• TensorFlow Datasets
• NumPy
• tqdm
• pickle

Installation

bash
pip install tensorflow tensorflow-datasets numpy tqdm

Usage

Dataset Preparation

1. Automatic download via TensorFlow Datasets:

python
train_data = tfds.load('celeb_a', split='train', batch_size=256)
test_data = tfds.load('celeb_a', split='test', batch_size=256)

2. Preprocessing includes:

• Image normalization: (img / 255 - 0.5) * 2
• Bias matrix calculation
• Trigger patch generation

Customization Options

Parameter	Description	Example Values
BIAS	Protected attribute	'Male', 'Young'
KEY_WORDS	Target classification attribute	'Attractive', 'Gray_Hair'
EPOCHS	Training iterations	5-20

Outputs

• Model checkpoints: result/{KEY_WORDS}/model_[epoch].ckpt
• Metrics recording: result/{KEY_WORDS}/pic_result.pic
• Evaluation logs: result/{KEY_WORDS}/log.txt

License

Distributed under MIT License.