/RTK-PAD

This is an official pytorch implementation of 'Fingerprint Presentation Attack Detector Using Global-Local Model'. This paper is accepted by IEEE trans on Cybernetics.

Primary LanguagePython

RTK-PAD

This is an official pytorch implementation of 'Fingerprint Presentation Attack Detector Using Global-Local Model', which is accepted by IEEE Transactions on Cybernetics

Fingerprint Presentation Attack Detector Using Global-Local Model (IEEE TCYB)

Requirements

  • numpy>=1.19.2
  • Pillow>=8.3.2
  • pytorch>=1.6.0
  • torchvision>=0.7.0
  • tqdm>=4.62.2
  • scikit-image>=0.18.3
  • scikit-learn>= 0.24.2
  • matplotlib>=3.4.3
  • opencv-python>= 4.5.3

Datasets

The proposed method is evaluated on a publicly-available benchmark, i.e. LivDet 2017, and you can download such dataset through link

Results

Usage

The RTK-PAD method is trained through three steps:

  • Data Preparation

    Generate the image list:

    python datafind.py \
--data_path {Your path to save LivDet2017}

    For example, python train_local_shuffling.py --data_path /data/fingerprint/2017 And then you can get data_path.txt to establish a Dataset Class() provided by pytorch.

  • Pre-trained Model Preparation

    RTK-PAD consists of Global Classifier and Local Classifier and we use two different initializations for them.

    For Global Classifier, the pre-trained model is carried on ImageNet, and you can download the weights from Link

    When it comes to Local Classifier, we propose a self-supervised learning based method to drive the model to learn local patterns. And you can obtain such initialization by

    python train_local_shuffling.py \
--sensor [D/G] \

    D refers to DigitalPersona and G is GreenBit. Since Orcanthus is with the different sizes of the images, we have a specific implementation for such case, which is hard to merge into this code.

  • Training models

    python train_main.py \
--train_sensor [D/G] \
--mode [Patch/Whole] \
--savedir {Your path to save the trained model} \

Evaluation

For evaluation, we can obtain RTK-PAD inference by

python evaluation.py \
--test_sensor [D/G]
--global_model_path {Your path to save the global classifier})
--patch_model_path {Your path to save the local classifier}
--patch_num 2 \

Citation

Please cite our work if it's useful for your research.

  • BibTex:
@article{liu2021fingerprint,
  title={Fingerprint Presentation Attack Detector Using Global-Local Model},
  author={Liu, Haozhe and Zhang, Wentian and Liu, Feng and Wu, Haoqian and Shen, Linlin},
  journal={IEEE Transactions on Cybernetics},
  year={2021},
  publisher={IEEE}
}