Federated Learning for IoMT-Enhanced Human Activity Recognition with Hybrid LSTM-GRU Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Wearable sensors and mobile devices enable human activity recognition (HAR).
• Centralized data processing for HAR raises significant privacy concerns.
• Balancing accuracy and privacy is crucial for real-world HAR applications.

Purpose of the Study:

• To propose a federated learning framework for privacy-preserving human activity recognition.
• To enhance feature extraction and classification accuracy in decentralized environments.
• To address privacy risks inherent in traditional centralized HAR data processing.

Main Methods:

• Federated Averaging for local model training on decentralized data.
• A hybrid Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) model for temporal dependency detection.
• Dual-feature extraction using convolutional blocks and an attention mechanism for pattern and relationship identification.

Main Results:

• The proposed system achieved superior performance on UCI-HAR, HARTH, and HAR7+ datasets.
• Demonstrated high classification accuracy and F1-scores compared to existing methods.
• Validated the framework's effectiveness in privacy preservation within decentralized environments.

Conclusions:

• The federated learning framework offers a scalable and reliable solution for privacy-conscious HAR.
• Combines advanced feature extraction with privacy-preserving techniques for robust real-time activity classification.
• Highlights the transformative potential of federated learning in advancing HAR applications.