DADM

The implementation of "DADM: Dual Alignment of Domain and Modality for Face Anti-spoofing".

Motivation of dual alignment of domain and modality:

An overview of the proposed DADM architecture:

Congifuration Environment

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

Data

Dataset.

Download the WMCA, CASIA-SURF, CASIA-CeFA, and PADISI-USC datasets.

Data Pre-processing.

Please refer to Data Preprocess.

Training and Testing

Run like this:

CUDA_VISIBLE_DEVICES=0 
python train_cross_lx.py --lr 5e-5 --batchsize 16 --modality RGBDIR --model dadm --train SURF CeFA USC --test WMCA

Citation

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

@article{yang2025dadm,
  title={Dadm: Dual alignment of domain and modality for face anti-spoofing},
  author={Yang, Jingyi and Lin, Xun and Yu, Zitong and Zhang, Liepiao and Liu, Xin and Li, Hui and Yuan, Xiaochen and Cao, Xiaochun},
  journal={arXiv preprint arXiv:2503.00429},
  year={2025}
}

@inproceedings{yang2025dadm,
  title={Dadm: Dual alignment of domain and modality for face anti-spoofing},
  author={Yang, Jingyi and Lin, Xun and Yu, Zitong and Zhang, Liepiao and Liu, Xin and Li, Hui and Yuan, Xiaochen and Cao, Xiaochun},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={12045--12056},
  year={2025}
}