Heart ID: Biometric Identification Using Wearable MIMO RF Heart Sensors

• Published in: IEEE journal of electromagnetics, RF and microwaves in medicine and biology Vol. 7; no. 1; pp. 3 - 14
• Main Authors: Conroy, Thomas B., Hui, Xiaonan, Sharma, Pragya, Kan, Edwin C.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.03.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Antennas; Biomedical monitoring; Biometric identification; Biometric recognition systems; Biometrics; Electrocardiography; Face recognition; Feature extraction; Heart; Human performance; MIMO; MIMO communication; Monitoring; neural networks; Radio frequency; Unsupervised learning; wearable sensors
• ISSN: 2469-7249; 2469-7257
• DOI: 10.1109/JERM.2022.3223034

Biometric identification (ID) has become increasingly prevalent in the digital era. Static biometric methods, such as fingerprint and facial recognition are widely accepted, yet generally vulnerable to targeted presentation attacks. Current development has expanded to dynamic biometrics, such as gait and electrocardiogram, that enable continuous authentication and are significantly more resistant to presentation attacks. However, dynamic biometrics often involve cumbersome acquisition which restricts their widespread use. Here, we introduce Heart ID, a novel dynamic biometric system that uses near-field coherent sensing (NCS) with a multiple-in multiple-out (MIMO) radio-frequency (RF) antenna setup to non-invasively acquire detailed recordings of internal cardiac dielectric boundary motion over clothing. NCS couples localized energy to the heart to derive interpersonal structural differences, while MIMO significantly increases the biometric entropy compared to single-point observation. We performed a human study of 20 subjects as well as 2 longitudinal evaluations, and employed an unsupervised feature extraction method to explore the ID performance of this new biometric. We found an ensemble classification approach using features derived from unsupervised learning can achieve accuracy exceeding 99% at a 40-second epoch.