FAS-Aug

The code repo for the paper "Towards Data-Centric Face Anti-Spoofing: Improving Cross-domain Generalization via Physics-based Data Synthesis".

Setup

Code downloading

git clone https://github.com/CeceliaSoh/FAS-Aug.git

Build Environment

cd FAS-AUG
conda create --name fas-aug python
conda activate fas-aug
pip install -r requirements.txt

Data Preparation

1. Please download the datasets you need. Due to the license, I cannot redistribute the datasets. You have to download the datasets by submitting your requests. The links for downloading the datasets can be found in https://github.com/RizhaoCai/Awesome-FAS.
2. Configure directories
   2.1 Go to data/extract_frames_mtcnn.py
   2.2 Configure the MTCCN path (check line 14 of data/extract_frames_mtcnn.py)
   2.3 Check lines 22 and 23. Configure ROOT_DIR and SAVE_DIR
3. Run scripts to extract data frames

   cd data
   # Change the DATASET_NAME you need; Check the keys of the dict 'DATASET_DIR' in data/extract_frames_mtcnn.py
   DATASET_NAME='CASIA-FASD'; python extract_frames_mtcnn.py ${DATASET_NAME}
   cd ..

Please configure _C.DATA.ROOT_DIR in config/defaults.py, which is the root directory where you save the preprocessed data.

Training

# Train models for the I&C&M to O protocol
python train.py --trainer vit_convpass DATA.EXTRA_DOMAIN 3 DATA.TRAIN_LIST "['data/data_list/CASIA-ALL.csv', 'data/data_list/MSU-MFSD-ALL.csv', 'data/data_list/REPLAY-ALL.csv']" DATA.TEST "data/data_list/OULU-NPU-ALL.csv" TRAIN.AUG.BAG 'FAS_Augmentations'

Some important hyperparameters:

• DATA.TRAIN_LIST: A list of training domain datasets, where each domain is stored in a separate CSV file. More datasets' CSV file can be found in data/data_list.
• DATA.TEST: The test domain datasets.
• TRAIN.AUG.BAG: Defines the augmentation strategy applied to the training data. When set to FAS_Augmentations, it applies the FAS augmentation technique proposed in this work.
• DATA.EXTRA_DOMAIN: The number of additional domains for applying the FAS augmentation. The default value is 3.

More parameters can be found in config/defaults.py
This file provide many basic configs for data, network, and training. Some model-specific/custom configs are in each sub-folders of [models] (e.g. models/vit_convpass/custom_config.py). Custom configs will be merged with the deafult config.

Citation

If you use this repo, please kindly cite the below papers

@article{cai2024towards,
  title={Towards Data-Centric Face Anti-Spoofing: Improving Cross-domain Generalization via Physics-based Data Synthesis},
  author={Cai, Rizhao and Soh, Cecelia and Yu, Zitong and Li, Haoliang and Yang, Wenhan and Kot, Alex},
  journal={arXiv preprint arXiv:2409.03501},
  year={2024}
}