FedSynHAR: a framework based on feature-enhanced adaptive pruning-mutual distillation for federated human activity...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Deep learning for human activity recognition (HAR) raises privacy concerns.
• Federated learning (FL) offers a privacy-preserving alternative by training models collaboratively without raw data sharing.
• Existing FL for HAR faces challenges from non-IID data, heterogeneous models, and edge hardware limitations.

Purpose of the Study:

• To develop a lightweight federated learning framework, FedSynHAR, for human activity recognition.
• To address privacy concerns while overcoming challenges in FL for HAR.
• To improve the efficiency, robustness, and accuracy of federated HAR systems.

Main Methods:

• FedSynHAR combines Gradient-Importance-based Adaptive Pruning (GIAP) and Channel-guided Feature-level Mutual Distillation (CFMD).
• GIAP reduces computational and communication overhead by pruning networks based on gradient importance.
• CFMD uses channel importance for mutual distillation, mitigating performance degradation and enhancing robustness under non-IID conditions.

Main Results:

• FedSynHAR demonstrated significant improvements on UCI-HAR and PAMAP2 datasets.
• Achieved approximately 2x faster convergence than FedAvg on UCI-HAR.
• Reached 94.91% accuracy under non-IID settings with up to two orders of magnitude overhead reduction.
• Showcased robustness under stronger heterogeneity on the PAMAP2 dataset.

Conclusions:

• FedSynHAR effectively addresses privacy concerns in HAR through federated learning.
• The proposed framework significantly enhances convergence speed, accuracy, and efficiency.
• FedSynHAR offers a robust and lightweight solution for federated human activity recognition, even with heterogeneous data and hardware constraints.