Know Me by My Pulse : Toward Practical Continuous Authentication on Wearable Devices via Wrist-Worn PPG

Wei Shao, Zequan Liang, Ruoyu Zhang, Ruijie Fang, Ning Miao, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun, Chongzhou FangEmail: {wayshao, zqliang, ryuzhang, rjfang, nmiao, ekay, srafatirad, hhomayoun}@ucdavis.edu, [email protected] · 2025

Read Know Me by My Pulse : Toward Practical Continuous Authentication on Wearable Devices via Wrist-Worn PPG by Wei Shao, Zequan Liang, Ruoyu Zhang, Ruijie Fang, Ning Miao, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun, Chongzhou FangEmail: {wayshao, zqliang, ryuzhang, rjfang, nmiao, ekay, srafatirad, hhomayoun}@ucdavis.edu, [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-13690.txt.

Introduction

Wei Shao and co-authors present a way in 2025 to unlock devices by pulse, using a smartwatch. The watch reads the pulse at the wrist with light, a method called PPG. Unlike heart-tracing methods that need chest contact, this one runs all the time without bothering the user.

The team built the system on a real watch and tested it at a low sampling rate to save power. Short four-second pulse clips were enough to tell users apart in most cases, with an average accuracy near eighty-eight percent on their data set. Training on clips from active times, not just rest, made it work better on the move. The low rate cut sensor power use by about half versus high-rate setups.

Catalog record

CREATORWei Shao, Zequan Liang, Ruoyu Zhang, Ruijie Fang, Ning Miao, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun, Chongzhou FangEmail: {wayshao, zqliang, ryuzhang, rjfang, nmiao, ekay, srafatirad, hhomayoun}@ucdavis.edu, [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-13690

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Know Me by My Pulse : Toward Practical Continuous Authentication on Wearable Devices via Wrist-Worn PPG PubID: pubid: This is the camera-ready version accepted at NDSS 2026. The official publication will be available at: https://www.ndss-symposium.org Wei Shao, Zequan Liang, Ruoyu Zhang, Ruijie Fang, Ning Miao, Ehsan Kourkchi, Setareh Rafatirad, Houman Homayoun, Chongzhou FangEmail: {wayshao, zqliang, ryuzhang, rjfang, nmiao, ekay, srafatirad, hhomayoun}@ucdavis.edu, [email protected] Abstract Biometric authentication using physiological signals offers a promising path toward secure and user-friendly access control in wearable devices. While electrocardiogram (ECG) signals have shown high discriminability, their intrusive sensing requirements and discontinuous acquisition limit practicality. Photoplethysmography (PPG), on the other hand, enables continuous, non-intrusive authentication with seamless integration into wrist-worn wearable devices. However, most prior work relies on high-frequency PPG (e.g., 75–500 Hz) and complex deep models, which incur significant energy and computational overhead—impeding deployment in power-constrained real-world systems. In this paper, we present the first real-world implementation and evaluation of a continuous authentication system on a smartwatch, We-Be Band, using low-frequency (25 Hz) multi-channel PPG signals. Our method employs a Bi-LSTM with attention mechanism to extract identity-specific features from short (4 s) windows of 4-channel PPG. Through extensive evaluations on both public datasets (PTTPPG) and our We-Be Dataset (26 subjects), we demonstrate strong classification performance with an average test accuracy of 88.11%, macro F1-score of 0.88, False Acceptance Rate (FAR) of 0.48%, False Rejection Rate (FRR) of 11.77%, and Equal Error Rate (EER) of 2.76%. Our 25 Hz system reduces sensor power consumption by 53% compared to 512 Hz and 19% compared to 128 Hz setups without compromising performance. We find that sampling at 25 Hz preserves authentication accuracy, whereas performance drops sharply at 20 Hz while offering only trivial additional power savings, underscoring 25 Hz as the practical lower bound. Additionally, we find that models trained exclusively on resting data fail under motion, while activity-diverse training improves robustness across physiological states. I Introduction Authentication is a critical part of modern digital systems, ensuring that only authorized users can access sensitive information and resources [ 20 , 7 , 25 ] . Traditional authentication methods, e.g. passwords, and PINs, have been proven to be vulnerable to breaches and often inconvenient [ 11 , 8 ] . In the world of wearable IoT devices, where users’ sensitive real-time data are being continuously collected and processed on the device, it is important to provide seamless authentication protection mechanisms beyond traditional authentication approaches. Biometric authentication is a compelling approach to address this challenge, relying on the distinct physiological or behavioral patterns of different individuals, which can be considered reliable sources of authentication [ 30 , 2 , 21 ] . Cardiac signal-based approaches, e.g. electrocardiogram (ECG)-based and photoplethysmography (PPG)-based authentication, are especially attractive because they can be integrated into real-time systems to enable continuous authentication [ 28 , 22 , 36 ] . ECG-based biometrics have shown high accuracy in research; however, ECG signal collection is often hampered by its intrusive nature and discontinuous acquisition (typically requiring contact electrodes or active user engagement, such as placing a finger on a sensor). Fig. 1 : Authentication System Overview PPG, by contrast, is easily captured by the optical sensors already embedded in most wearables, making it non-intrusive, ubiquitous, and power-efficient. Unlike ECG, which demands low‑impedance skin–electrode contact and thus imposes more hardware complexity and user burden, PPG can be sensed passively with minimal user involvement. Although ECG yields cleaner waveforms and zero phase lag because it measures the heart’s electrical impulse directly, these benefits come at the cost of comfort and form‑factor flexibility. PPG’s main drawbacks are a higher susceptibility to motion artifacts, ambient‑light interference, and a physiological lag behind the cardiac electrical event, all of which can degrade signal quality in everyday use. Nevertheless, early studies that used high‑frequency PPG have reported excellent authentication accuracy under controlled conditions [ 33 , 23 , 31 ] . Despite PPG’s potential, gaps remain in the current literature on continuous PPG-based authentication as well as in a practical way. These solutions frequently fall short in enabling continuous, real-world operation due to practical constraints like power consumption and comfort, and often rely on high-quality PPG signals collected at high sampling rates (typically 75–500 Hz) using professional or medical-grade sensors [ 10 , 4 , 18 ] . Such high-fidelity data capture rich heartbeat morphology but at the cost of increased power consumption and data volume. This is impractical for power-constraint wearable devices. Moreover, much of the past work has been evaluated on pre-collected datasets or in laboratory scenarios with stationary subjects. Practical real-world validation of PPG authentication has been limited. Few studies attempted on-device deployment of PPG authentication algorithms, and they often do not address power and energy impacts. In this paper, we address these challenges by proposing a continuous authentication system using low-frequency multi-channel PPG signals and evaluate the system through practical real-world demonstration, which to the best of our knowledge, is the first work to fully implement and evaluate a continuous biometric authentication system on a smartwatch. Figure 1 shows the high-level overview of the authentication system. We chose to rely solely on the PPG signal. While incorporating additional sensors like Galvanic Skin Response (GSR), accelerometer, or even ECG could help, our aim is to demonstrate strong results using just PPG. Our approach leverages the We-Be Band—a research-oriented smartwatch equipped with PPG and other physiological sensors [ 34 , 12 ] —to collect photoplethysmogram data from users. Uniquely, we sample the PPG at native 25 Hz, substantially lower than in most prior studies, to enable ultra-low-power operation. By using multi-channel PPG (multiple wavelengths/sensors), we capture complementary cardiovascular features even at this lower sampling rate. These signals are fed into a machine learning model that learns each user’s heartbeat pattern. The model is lightweight and is suitable for real-time inference on the edge device or wearable’s embedded processor. Our system thus emphasizes an energy-aware design: it maintains robust authentication accuracy while minimizing sampling frequency and computation, thereby optimizing for battery life of wearable devices. To evaluate our approach, we aim to answer the following key Research Questions (RQs): • RQ1: How does PPG sampling frequency affect authentication accuracy and generalization in real-world wearable use cases? • RQ2: To what extent does multi-channel PPG improve authentication resilience against impersonation and signal noise compared to single-channel input? • RQ3: What is the impact of activity-aware training (e.g., walking, typing) on the robustness of PPG-based biometric models in dynamic environments? • RQ4: Can a low-power authentication model operating at 25 Hz PPG sampling achieve accuracy and robustness comparable to high-frequency counterparts on real-world wearable devices? • RQ5: How feasible is real-time, continuous biometric authentication using low-frequency PPG on embedded devices in terms of power and usability? We answer these questions through extensive experimentation with our prototype system. In terms of authentication performance, our results show that the proposed 25 Hz multi-channel PPG approach can achieve strong accuracy—for instance, about 88% classification accuracy in identifying users, with an equal error rate (EER) below 3%. This is comparable to state-of-the-art methods that use significantly higher sampling rates, indicating that a higher rate is not strictly necessary for effective biometric recognition. Importantly, the system maintains this accuracy in realistic

Read the complete text