Enhancing Privacy-Preserving Personal Identification Through Federated Learning With Multimodal Vital Signs Data
Tae-Ho Hwang, Jingyao Shi, Kang Yoon Lee
Abstract
Personal identification can be achieved through various multimodal vital sign measurement methodologies in human physiology, such as electrocardiogram (ECG) and radar signals. However, this process inevitably brings privacy protection issues into light during data collection and analysis. Additionally, potential discomfort may be associated with the physical contact during these measurements. To mitigate these issues, our study explores the utilization of federated learning (FL) for privacy protection and non-contact sensors, such as radar, to alleviate contact-related discomfort. Our objective was to establish the viability of privacy-secured personal identification (PI) models that utilize FL. Furthermore, we scrutinized the performance of FL-based PI models that incorporate non-contact radar signals. We compared the performance levels of five conventional machine learning (ML) models with that of five FL-based models by utilizing ECG and radar signals. The experimental outcomes indicated that ECG-based models had superior accuracy overall, with the FL-CNN-BLSTM model yielding an accuracy of 97.3%. Models based on radar signals displayed slightly lower accuracy, with the FL-CNN-BLSTM model delivering an accuracy of 81.9%. These results confirmed the effectiveness of FL-based PI models employing radar signals. Our research therefore augments the evolution of privacy-guarded PI processes utilizing radar signals and FL and lays a robust groundwork for upcoming studies in this field.