synthaspoof's Introduction

SynthASpoof

Note

This is the official repository of the work: SynthASpoof: Developing Face Presentation Attack Detection Based on Privacy-friendly Synthetic Data (link)

*** Accepted at CVPR workshops 2023 ***

SynthASpoof dataset

The SynthASpoof is the synthetic-based face presentation attack detection datasets, including synthetic-generated 25,000 bona fide images and 78,800 corresponding attacks collected by presenting the printed/replayed images to capture cameras (one mobile phone, two different tablets, and one webcam).

The dataset can be downloaded via Google Drive (please share your name, affiliation, and official email in the request form).

The image samples in SynthASpoof are shown:

NOTE: Please note that the SynthASpoof data aims at being a development (training) data for face PAD. All the evalution are conducted on the authentic data to real-world scenarios.

Data preparation

SynthASpoof dataset: the images should be preprocessed including face detection and crop via data_preprocess/crop_face.py before training.

Test real datasets: for testing, we use four real face PAD dataset, MSU-MFSD, CASIA-FAS, Idiap ReplayAttack, and OULU-NPU. The data format in the real face PAD datasets are videos. We sampled 25 frames in the average time interval of each video.

Then, CSV files are generated for each dataset for further training and evaluation. The format of the dataset CSV file is:

image_path,label
/image_dir/image_file_1.png, bonafide
/image_dir/image_file_2.png, bonafide
/image_dir/image_file_3.png, attack
/image_dir/image_file_4.png, attack

Training

Example of training:

python train.py \
  --csv_1 'SynthASpoof_data.csv' \
  --csv_2  'RealPAD_data.csv'

The loss is computed based on csv_1. csv_2 is used for style transfer.

Testing

Example of testing:

python train.py \
  --test_csv 'test_data.csv' \
  --model_path 'model.pth'

where test_data.csv contains image path and the corresponding label (bonafide or attack).

Models

The models trained on SynthASpoof with MixStyle can be downloaded via owncloud.

if you use SynthASpoof dataset in this repository, please cite the following paper:

@misc{synthaspoof,
  doi = {10.48550/ARXIV.2303.02660},
  url = {https://arxiv.org/abs/2303.02660},
  author = {Fang, Meiling and Huber, Marco and Damer, Naser},
  title = {SynthASpoof: Developing Face Presentation Attack Detection Based on Privacy-friendly Synthetic Data},
  publisher = {arXiv},
  year = {2023},
  copyright = {Creative Commons Attribution Non Commercial Share Alike 4.0 International}
}

License

The dataset, the implementation, or trained models, use is restricted to research purpuses. The use of the dataset or the implementation/trained models for product development or product competetions (incl. NIST FRVT MORPH) is not allowed. This project is licensed under the terms of the Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license. Copyright (c) 2020 Fraunhofer Institute for Computer Graphics Research IGD Darmstadt.

synthaspoof's People

Contributors

Stargazers

Watchers

synthaspoof's Issues

ValueError: All-NaN slice encountered

On running the test script as per the format provided in the README:
python test.py \ --test_csv 'test_data.csv' \ --model_path 'model.pth'

I'm getting the error message:
np.abs(differ_tpr_fpr_1) [nan nan nan nan nan nan nan nan nan] Traceback (most recent call last): File "C:\Users\ssais\Documents\AntiSpoofing_Githubrepos\SynthASpoof\test.py", line 94, in <module> main(test_csv=args.test_csv, args=args) File "C:\Users\ssais\Documents\AntiSpoofing_Githubrepos\SynthASpoof\test.py", line 29, in main AUC, HTER = test_model(model, test_loader) File "C:\Users\ssais\Documents\AntiSpoofing_Githubrepos\SynthASpoof\test.py", line 62, in test_model AUC, _, _, HTER = performances_cross_db(raw_test_scores, gt_labels) File "C:\Users\ssais\Documents\AntiSpoofing_Githubrepos\SynthASpoof\utils.py", line 75, in performances_cross_db et_err_threhold_cross_db right_index = np.nanargmin(np.abs(differ_tpr_fpr_1)) File "C:\Users\ssais\Documents\AntiSpoofing_Githubrepos\SynthASpoof\env\lib\site-packages\numpy\lib\nanfunctions.py", line 552, in nanargmin raise ValueError("All-NaN slice encountered") ValueError: All-NaN slice encountered
The issue seems to be with the fact that either fpr or tpr is a list with all-Nan values, due to which the values of np.abs(differ_tpr_fpr_1) is [nan nan nan nan nan nan nan nan nan] (I printed it out as a debugging measure). Please do provide the solution to this issue.

meilfang / synthaspoof Goto Github PK