DLIF

The implementation of Generalized Face Anti-spoofing via Finer Domain Partition and Disentangling Liveness-irrelevant Factors (DLIF).

The motivation of the proposed DLIF method:

An overview of the proposed DLIF architecture:

Congifuration Environment

• python 3.10
• torch 1.12.1
• torchvision 0.13.1
• cuda 11.4

Data

Dataset.

Download the OULU-NPU, CASIA-FASD, Idiap Replay-Attack, MSU-MFSD and CASIA-Spoof datasets.

Data Pre-processing.

MTCNN algotithm is utilized for face detection and face alignment. All the detected faces are normlaize to 256x256x3, where only RGB channels are utilized for training.

Training

Move to the folder ./protocol/I_C_M_to_O/ and just run like this:

python -m torch.distributed.launch --nproc_per_node=ngpus train.py

The file config.py contains all the hype-parameters used during training.

Testing

Run like this:

python test.py

Citation

Please cite our paper if the code is helpful to your research.

@article{yang2024generalized,
  title={Generalized Face Anti-spoofing via Finer Domain Partition and Disentangling Liveness-irrelevant Factors},
  author={Yang, Jingyi and Yu, Zitong and Ni, Xiuming and He, Jia and Li, Hui},
  journal={arXiv preprint arXiv:2407.08243},
  year={2024}
}