Extension of physical activity recognition with 3D CNN using encrypted multiple sensory data to federated learning...

Area of Science:

• Computer Science
• Biomedical Engineering
• Cybersecurity

Background:

• The Internet of Medical Things (IoMT) enables smart healthcare via wearable sensors, but data transmission poses privacy risks.
• Federated learning trains models locally, aggregating them centrally without sharing raw data, enhancing privacy.
• Data encryption further secures sensitive biomedical information during transmission.

Purpose of the Study:

• To propose a novel method for preserving client privacy and protecting biomedical data during internet transmission.
• To integrate data encryption with federated learning for enhanced security in IoMT applications.
• To evaluate the effectiveness of 3D convolutional neural networks (CNNs) with encryption and federated learning for human activity recognition.

Main Methods:

• Utilized 3-dimensional convolutional neural networks (3D CNNs) for human activity recognition using multi-sensory data.
• Applied bitwise XOR operator for data encryption before transmission.
• Extended 3D CNNs to traditional federated learning and multi-key homomorphic encryption-based federated learning.

Main Results:

• The 3D CNN method achieved high accuracy (94.6%-94.9%) without encryption or federated learning on benchmark datasets.
• Data encryption slightly reduced accuracy to 89.5% but remained competitive with state-of-the-art methods.
• The proposed federated learning scheme prevented inference of private results by unauthorized parties accessing trained models.

Conclusions:

• A novel sensory data representation translates bio-signals into 3D activity images.
• Proposed 3D CNN methods demonstrate superior performance in human activity recognition compared to other deep learning approaches.
• The data-encrypted federated learning approach is feasible and efficient for privacy preservation in smart healthcare.