Biosignal Authentication Considered Harmful Today

Codebase to accompany Biosignal Authentication Considered Harmful Today:

Veena Krish, Nicola Paoletti, Milad Kazemi, Scott Smolka, Amir Rahmati (2024). Biosignal Authentication Considered Harmful Today. In USENIX Security Symposium (USENIX Sec).

Overview

This codebase is divided into three parts:
1. custom_datasets: loading and processing raw data, used across the codebase
2. model_training: training the cyclegan network used to generate spoofed biosignals
3. authentication_systems, an implementation of the systems tested in the paper

1. custom_datasets

Data handlers for loading, processing and splitting data are packaged as a python module that can be imported in other sections of the overall codebase. This can be

Packaging:

First build the custom_datasets package:

cd custom_datasets
python -m build

Then install the package in the main venv as:

pip install --editable .

Notes:

• Need to upgrade pip beyond 21.3 in order to build and install editable
• PyPA User Guide if you encounter any issues

Main usage:

The file datasets.py contains generators for yielding a data sample for a given configuration, defined in properties.py``. Arguments datasets.get_data`:

• dataset_name: as defined in properties.py. E.g. 'dalia', 'capno'
• data_type: biosignal modality. E.g. 'ECG', 'PPG'
• spoof: bool for requesting generated/spoofed data or original. False is overriden if spoof_name supplied
• spoof_name: name of the spoof generation method (e.g. 'cardiogan_contrastive', 'video_ecg_contrastive')
• split: oneof: ['train', 'test', 'all'], denotes the split on subjects (note: not on time), as defined in properties.py. Generally used for training/testing the cyclegan's generalization ability
• fraction_time: fraction of total signal time to yield, typically used for training v testing the authentication systems.

from custom_datasets import datasets
generator = dataasets.get_data(dataset_name='dalia', data_type='PPG', spoof=False, spoof_name=None, split='train', fraction_time=(0, 0.5), session=None)
subject_ix, subject_name, session_name, ppg_npy = next(generator)

# For paired data (for evaluating model training):
generator = dataasets.get_data('dalia', ['PPG', 'ECG'] spoof=False, split='train')
_, _, _, paired_data = next(generator)
ppg_npy = paried_data['PPG']
ecg_npy = paried_data['ECG']

model_training

Contains pytorch-based model training scripts. View example usage with python train_contrastive_cardiogan.py --help. DataLoaders pull paried raw data from custom_datasets.get_paired_data and shuffle to train unpaired translation. Each DataLoader specifies the source and target datatype for the given dataset.

Example usage:

# Train the main ppg -> ecg spoofing model
python train_contrastive_cardiogan.py --dataset ecgppg_cardiogan

# Train an example video -> ecg spoofing model
python train_contrastive_cardiogan.py --dataset rppgecg_hcitagging

authentication_systems

The following authentication systems are implemented (using public codebases if available). Evaluation scripts are included in each directory for testing the false acceptance rate of spoofed data.

ECG

ECGXtractor

Published as: ECG Biometric Recognition: Review, System Proposal, and Benchmark Evaluation. P Melzi, R Tolosana, R Vera-Rodriguez - IEEE Access, 2023.

Implementation is obtained directly from authors' provided codebase and modified to work with our datasets.

Usage:

# Prepare data for <dataset_name>
python src/prepare_dataset.py --dataset <dataset_name>  # note: this creates train/train.json and train/val.json
python src/prepare_dataset.py --dataset <dataset_name> --eval  # note: this creates eval/train.json that has nothing and eval/val.json
python src/prepare_dataset.py --dataset bidmc --eval --spoof_name cardiogan_contrastive

# Make config files for autoencoder and siamese network training and testing
python src/make_configs.py --dataset <dataset_name> --autoencoder
python src/make_configs.py --dataset <dataset_name>
python src/make_configs.py --dataset <dataset_name> --eval
python src/make_configs.py --dataset <dataset_name> --spoof_name cardiogan_contrastive --eval

# Train
python src/train.py --config_file configs/autoencoder/<dataset_name>/config_autoencoder.json --rename latest
python src/train.py --config_file configs/siamese/<dataset_name>/config_train.json --rename latest

# Then predict and save the working EER, EER_threshold for a withheld validation section
python src/predict.py --dataset <dataset_name> --config_file configs/siamese/<dataset_name>/config_train.json --model_name latest --save_stats

# Eval using the predicted EER/thresh on a final withheld section, over 10 attempts
python src/eval.py --dataset <dataset_name> --model_name latest --spoof_name original
python src/eval.py --dataset <dataset_name> --model_name latest --spoof_name cardiogan_contrastive

DeepECG

Published as: Deep-ECG: Convolutional neural networks for ECG biometric recognition. RD Labati, E Muñoz, V Piuri, R Sassi, F Scotti - Pattern Recognition Letters, 2019

Usage:

# Train
python train.py --dataset <dataset_name> --save latest 

# Test and get the EER ("--save_stats" will save the eer threshold to the model file)
python test.py --dataset <dataset_name> --model_path saved_models/<dataset_name>/latest.pt --save_stats
# Eval using the saved EER over 10 attempts on a separate section of data, on original and spoofed data

python eval.py --dataset <dataset_name> --model_path saved_models/<dataset_name>/latest.pt --spoof_name original
python eval.py --dataset <dataset_name> --model_path saved_models/<dataset_name>/latest.pt --spoof_name cardiogan_contrastive

EDITH

Published as: [Ibtehaz, Nabil, et al. [EDITH: ECG biometrics aided by deep learning for reliable individual authentication. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021](Ibtehaz, Nabil, et al. "EDITH: ECG biometrics aided by deep learning for reliable individual authentication." IEEE Transactions on Emerging Topics in Computational Intelligence 6.4 (2021): 928-940. )

Usage:

# Train feature extractor and siamese models (n.b. train_siamese also gets EER on a withheld set and saves to model)
python train_baseclassifier.py --dataset <dataset_name> --save base
python train_siamese.py --dataset <dataset_name> --saved_model saved_models/<dataset_name>/base.pt --save siamese

# Eval on original/spoofed datasets
python eval.py --dataset <dataset_name> --saved_base_model saved_models/<dataset_name>/base.pt --saved_siamese_model saved_models/<dataset_name>/siamese.pt --spoof_name cardiogan_contrastive

KeyToYourHeart

Published as: A Key to Your Heart: Biometric Authentication Based on ECG Signals. N Samarin - School of Informatics, University of Edinburgh, 2018

Usage:

python generate_data.py --dataset <dataset_name> --spoof_name original --split train
python generate_data.py --dataset <dataset_name> --spoof_name original --split test
python generate_data.py --dataset <dataset_name> --spoof_name cardiogan_contrastive --split test

2. Train, get test EER and save
python train.py --dataset <dataset_name> --save

3. Eval on 10 trials
python eval.py --dataset <dataset_name> --model_dir saved_models/<dataset_name> --spoof_name original
python eval.py --dataset <dataset_name> --model_dir saved_models/<dataset_name> --spoof_name cardiogan_contrastive

PPG

CorNET

Published as: CorNET: Deep learning framework for PPG-based heart rate estimation and biometric identification in ambulant environment. D Biswas et al. IEEE Transactions on Biomedical Circuits and Systems, 2019

Usage:

# Train and save test-split EER: 
`python train.py --dataset <dataset_name> --save latest
# Eval on orig and spoofed datasets
python eval.py --dataset dalia --models_dir saved_models/<dataset_name>/latest/
python eval.py --dataset dalia --models_dir saved_models/<dataset_name>/latest/ --spoof_name cardiogan_contrastive

Hwang2020

Published as: Evaluation of the time stability and uniqueness in PPG-based biometric system DY Hwang et al. IEEE Transactions on Information Forensics and Security, 2020

Generally based off unofficial implementation (written by an author of the paper but not linked within)

Usage:

# Train and get test-split EER: 
python train.py --dataset <dataset_name> --lstm --save latest
# Eval on original and spoofed sets :
python eval.py --dataset <dataset_name> --models_dir saved_models/<dataset_name>/latest/
python eval.py --dataset <dataset_name> --spoof_name cardiogan_contrastive --models_dir saved_models/<dataset_name>/latest

SCG

WaveletTransform

Published as: Exploring seismocardiogram biometrics with wavelet transform.PY Hsu, PH Hsu, HL Liu. 2020 25th International Conference on Pattern Recognition (ICPR), 2021

Usage:

python train_and_eval.py --wavelet morse --spoof_name original`
python train_and_eval.py --wavelet morse --spoof_name cardiogan_contrastive`

MotionArtifactResilient

Published as: Motion artifact resilient SCG-based biometric authentication using machine learning PY Hsu, PH Hsu, TH Lee, HL Liu, 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) 2021

Usage:

python train.py --save latest
python eval.py --model_path saved_models/latest --spoof_name cardiogan_contrastive

BCG

Herbert2018

Published as: Ballistocardiogram-based authentication using convolutional neural networks. Hebert, Joshua, et al. Worcester Polytechnic Institute, 2018

python train.py --dataset <dataset_name> --save latest
python eval.py --dataset dataset --models_dir saved_models/<dataset_name>/latest
python eval.py --dataset dataset --models_dir saved_models/<dataset_name>/latest --spoof_name cardiogan_contrastive

ZhangRNN

Published as: Ballistocardiogram based person identification and authentication using recurrent neural networks. Zhang, Xianwen, et al. 2018 11th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI), 2018

python train.py --dataset <dataset_name> --save latest
python eval.py --dataset dataset --models_dir saved_models/<dataset_name>/latest
python eval.py --dataset dataset --models_dir saved_models/<dataset_name>/latest --spoof_name cardiogan_contrastive