Enhancing Privacy-Preserving Personal Identification Through Federated Learning With Multimodal Vital Signs Data

Abstract

Personal identification (PI) can be verified using multimodal vital sign measurement methodologies in human physiology, such as electrocardiography (ECG) and radar signals. However, these processes are inevitably associated with concerns over privacy during the data collection and analysis stages, as well as discomfort during the physical measurement stages. To mitigate these issues, we explored the utilization of federated learning (FL) and noncontact sensors to ensure privacy protection and alleviate contact-related discomfort, respectively. Our objective was to establish the viability of privacy-secured PI models using FL. Furthermore, we examined the performance of FL-based PI models that incorporate non-contact radar signals, comparing the performance levels of five conventional machine learning (ML) models with those of five FL-based models using ECG and radar signals. Our experimental results indicate that although the ECG-based models exhibited superior overall accuracy, their radar-based counterparts demonstrated only slightly lower accuracy. These results confirm the effectiveness of FL-based PI models when compared with standard ML models. Thus, our study augments the evolution of privacy-guarded PI processes and lays a robust groundwork for future research in this field.

Keywords

• Bidirectional long short-term memory
• ensemble model
• electrocardiogram
• federated learning
• long short-term memory
• personal identification