A Privacy and Energy-Aware Federated Framework for Human Activity Recognition.

Area of Science:

• Biomedical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Centralized deep learning for human activity recognition (HAR) with wearable sensors faces privacy, communication, and efficiency challenges.
• Existing models struggle with the demands of continuous healthcare monitoring.

Purpose of the Study:

• To propose a federated learning framework for privacy-preserving and energy-efficient HAR.
• To develop a hybrid Spiking Neural Network (SNN) and Long Short-Term Memory (LSTM) model (S-LSTM) for enhanced HAR performance.

Main Methods:

• Developed a hybrid S-LSTM model integrating SNNs and LSTMs for synergistic event-driven and sequence modeling capabilities.
• Employed surrogate gradient learning and backpropagation through time for fully supervised end-to-end training.
• Evaluated the model on two public datasets within a federated learning framework.

Main Results:

• The S-LSTM model outperformed LSTM, CNN, and S-CNN models in accuracy and energy efficiency.
• Achieved high accuracy rates: 97.36% for indoor and 89.69% for outdoor HAR scenarios.
• Demonstrated a 32.30% improvement in energy efficiency compared to standard LSTM models.
• Personalization through local data fine-tuning boosted individual user accuracy by up to 9%.

Conclusions:

• The proposed federated learning framework with the S-LSTM model offers a viable solution for energy-efficient and privacy-preserving HAR.
• The hybrid S-LSTM architecture effectively combines the strengths of SNNs and LSTMs for superior performance.
• Personalization significantly enhances HAR model accuracy for individual users, underscoring its importance in healthcare applications.