Deep learning-based photoplethysmography biometric authentication for continuous user verification

Abstract

Biometric authentication methods have gained prominence as secure and convenient alternatives to traditional passwords and PINs. In this paper, we propose a novel approach for biometric authentication using photoplethysmography (PPG) signals and deep learning techniques. PPG is a non-invasive method that measures variations in blood volume within microvascular tissue beds, and it is typically used for monitoring heart rate and oxygen saturation. Our research leverages the unique characteristics of PPG signals to develop a robust and continuous user verification system. The primary goal of our study is to explore the feasibility and effectiveness of PPG-based biometric authentication, enabling a seamless and secure means of confirming the identity of individuals. We use a diverse dataset of PPG signals from various individuals, ensuring that it encompasses differences in skin tone, age, and other variables that can influence PPG signal characteristics. The collected data undergoes careful preprocessing, including noise removal, baseline correction, and heartbeat segmentation. For the core of our authentication system, we design and train a multiscale feature fusion deep learning (MFFD) model. This model, utilizing a Convolutional Neural Network (CNN) architecture, takes as input the relevant features extracted from PPG signals and learns to differentiate between individuals based on their unique PPG patterns. In this study, the input is constructed by gradually incorporating various features, beginning with a single PPG signal. In this study, the CNN model was trained independently, followed by the implementation of score fusion techniques. Our evaluation demonstrates the effectiveness of the PPG-based biometric authentication system, achieving high accuracy while addressing key security concerns. We consider false acceptance rate (FAR) and false rejection rate (FRR) to assess the system's performance. The model achieves the Accuracy of 99.5 % on BIDMC, 98.6 % on MIMIC, 99.2 % on CapnoBase dataset.