Biometric Authentication based on Enhanced Remote Photoplethysmography Signal Morphology

Welcome to use the code from our paper "Biometric Authentication based on Enhanced Remote Photoplethysmography Signal Morphology". The training weights, data, and results can be downloaded here. Please unzip rppg_biometrics_artifacts.zip to the main folder.

Requirement

Please see requirement.txt for the required python environment.

Dataset

The data structures are as follows.

data_example
├── h5_obf # this folder contain data examples from OBF dataset
    ├── 001_xx_1.h5 # 001 means the subject ID number. 1 means the first session (pre-exercise)
        ├── imgs # shape (T, 6, 6, 3), the downsampled videos frames,permutation will be performaned during loading data. T is the number of frames, 6 is the spatial dimension, 3 is the RGB channels.
        ├── bvp # shape (T,), the ground truth cPPG signal. NOT USED DURING TRAINING. only used for morphology evaluation.
    ├── 001_xx_2.h5 # 001 means the subject ID number. 2 means the second session (post-exercise)
    .
    ├── 100_xx_1.h5
    ├── 100_xx_2.h5

├── external_cppg.h5 # A combination of cPPG biometrics datasets including Biosec2, BIDMC, and PRRB for rPPG-cPPG hybrid training. 

Training

The 1st training stage: rPPG Unsupervised Pre-training

if ./data_example/h5_obf contains the complete OBF data, one can run python rppg_model_pretraining.py to start the 1st training stage. More details can be found in the comments of the .py file.

Weights and irrelevant power ratios (IPR) are saved during training at ./results (The training records are already in the folder.). The best weight is chosen at the lowest IPR by ./notebooks/sacred_train.ipynb. The chosen weight are stored at ./rppg_model_pretrained_weights.pt, which will be used in the 2nd training stage.

The 2nd training stage: rPPG-cPPG Hybrid Training

if ./data_example/h5_obf contains the complete OBF data and ./data_example/external_cppg.h5 exists, one can run python joint_rppg_cppg_hybrid_training.py to start the 2nd training stage. More details can be found in the comments of the .py file. Weights are saved during training. joint_inference.py is used to get the ID prediction results and the rPPG signals at different epochs as shown below.

for epoch in {0..29}
do
    python joint_inference.py with train_exp_num=1 e=$(awk "BEGIN {print(${epoch}*100)}") -i $(awk "BEGIN {print(${epoch}*100)}")
done

The training records after the 2nd training stage are in ./joint_results/default/1. One can run python eval_eer_auc.py to get the EER and AUC metrics. One can run python eval_morph.py to get the Pearson correlation results for the morphology evaluation. More details can be found in the comments of the .py files.

Citation

@article{sun2024biometrics,
  title={Biometric Authentication Based on Enhanced Remote Photoplethysmography Signal Morphology},
  author={Sun, Zhaodong and Li, Xiaobai and Komulainen, Jukka and Zhao, Guoying},
  booktitle={International Joint Conference on Biometrics (IJCB)},
  year={2024},
}