Novel Robust Photoplethysmogram-Based Authentication

• Published in: IEEE sensors journal Vol. 22; no. 5; pp. 4675 - 4686
• Main Authors: Pu, Limeng, Chacon, Pedro J., Wu, Hsiao-Chun, Choi, Jin-Woo
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.03.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Authentication; autoencoder; Biometrics; Blood circulation; Blood flow; continuous authentication; Data collection; Datasets; Electronic devices; Feature extraction; Heart rate; multiwavelets; neural networks; Peripheral circulation; Photoplethysmogram (PPG); Real time; Robustness; Sensors; Skin; Training; Wearable devices
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2022.3146291

As the tremendous advances in the biosensing technology, the biometric authentication approach has been adopted to replace the conventional password-based systems. Due to their harmless nature, photoplethysmogram (PPG) related applications have been quite intriguing in recent years. PPG can measure volumetric blood-flow changes in the peripheral circulation. Because PPG signals are very easy to acquire, compared to other biometrics (fingerprints, iris/retina, etc.), especially by low-cost wearable electronic devices, they are widely adopted to measure heart rates. In this paper, we devise a novel robust PPG based authentication system, which is capable of continuously authenticating the user instead of one-time authentication as carried out by the conventional techniques. This new system consists of preprocessing and filtering, motion artifact (MA) removal, template and feature extraction, and training. The associated multiwavelet-based feature extraction mechanism facilitates more reliable features than the conventional scalar-wavelet schemes so that it allows the learning model to better distinguish users. The associated autoencoder perfectly transforms the input PPG signal into the latent space. Finally, any distance measure can be adopted to classify or authenticate users. In order to build a sufficiently large dataset, we have combined three public datasets and our own dataset collected by ourselves locally together to create a new dataset consisting of 120 subjects totally. We have already built the hardware prototype successfully for real-time training data collection and real-time authentication. The equal-error-rate at the authentication stage reaches 5.5% and the identification accuracy reaches 98% at the identification stage.