Taming Self-Supervised Learning for Presentation Attack Detection: De-Folding and De-Mixing

This is an official pytorch implementation of 'Taming Self-Supervised Learning for Presentation Attack Detection: De-Folding and De-Mixing', which is accepted by IEEE Transactions on Neural Networks and learning systems. [Paper] [arxiv]

Requirements

• python 3.6
• pytorch 1.1.0
• torchvision 0.3.0
• numpy 1.19.5
• pandas 0.25.3

Pre-processing

Face

• Dataset

  Download the OULU-NPU(O), CASIA-FASD(C), Idiap Replay-Attack(I) and MSU-MFSD(M) datasets.

• Face alignment

  MTCNN algotithm is utilized for face detection/alignment.

• Data Label Generation

  Move to the $root/face and generate the data label list:

  python data_find.py --data_path dataPath
  

  dataPath is the path of data.

Fingerprint

• Dataset

  Download the LivDet 2017 datasets.

• Data Label Generation

  Move to the $root/fingerprint and generate the label:

    python data_find.py --data_path dataPath
  

  dataPath is the path of data.

Usage

• Move to the $root/face or $root/fingerprint and run pretext task:

  python ifom_pretext.py --savedir pretextPath
  

  pretextPath is the filename to save model, which is in $root/face/pretext_model.

• And then move to the $root/face and run downstream task:

  python train_downstream.py --loadModelPath ./pretext_model/pretextPath/xxxx.pth --savedir downstreamPath
  

  ./pretext_model/pretextPath/xxxx.pth is the path of model trained from IF-OM and downstreamPath is the path to save the downstream task model. Besides, log file can be found at ./downstream_log/downstreamPath.

Note that, for each case, the configuration/hyper-parameter is different. To making this project easy-to-use, the corresponding details will be publicly available soon.

Trained Weights

The trained model weights are available in current directory.

• Face

  The model trained in OULU-NPU and CASIA-FASD, which has a 18.96% EER, 89.48% AUC and 30.18% TDR@FDR=1.0% tested on CASIA-FASD and Idiap Replay-Attack

  ./trained_weights/Downstream_train_om_test_ci.pth
  

• Fingerprint

  The model trained with Orcathus sensor on LivDet2017, which has a 1.40% EER, 99.75% AUC and 1.38% TDR@FDR=1.0% under the cross-material setting.

  ./trained_weights/Downstream_train_Orcathus_test_ Orcathus.pth
  

Example

Move to the $root/face or $root/fingerprint and run like this:

python example.py

Such shell can log the prediction and ground truth of test samples using our trained model.

BibTeX

@article{kong2023taming,
  title={Taming self-supervised learning for presentation attack detection: De-folding and de-mixing},
  author={Kong, Zhe and Zhang, Wentian and Liu, Feng and Luo, Wenhan and Liu, Haozhe and Shen, Linlin and Ramachandra, Raghavendra},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}