ProtoHAR: Prototype Guided Personalized Federated Learning for Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Federated Learning (FL) is gaining traction for sensor-based Human Activity Recognition (HAR).
• Real-world sensor data is often Non-Independent and Identically Distributed (Non-IID), leading to sparse data and inconsistent distributions across devices.
• Traditional FL methods struggle with Non-IID data, causing model drift, slow convergence, and communication burdens in heterogeneous environments.

Purpose of the Study:

• To address the challenges of Non-IID sensor data in cross-device FL for HAR.
• To propose an efficient framework that decouples representation and classification in heterogeneous FL settings.
• To improve the performance and convergence speed of FL models for HAR.

Main Methods:

• Introduced ProtoHAR, a prototype-guided FL framework for HAR.
• Leveraged global prototypes to correct activity feature representation, enabling knowledge flow without privacy leakage.
• Developed a method to train a better classifier, mitigating local model drift during personalized training.

Main Results:

• ProtoHAR demonstrated superior performance compared to state-of-the-art FL algorithms on four diverse HAR datasets (USC-HAD, UNIMIB-SHAR, PAMAP2, HARBOX).
• The framework achieved faster convergence speeds in HAR tasks.
• Experiments were conducted in both controlled and real-world scenarios, validating the robustness of ProtoHAR.

Conclusions:

• ProtoHAR effectively tackles the Non-IID challenge in sensor-based HAR within a federated learning context.
• The proposed prototype-guided approach enhances feature representation and classifier training for improved personalized models.
• ProtoHAR offers a promising solution for efficient and accurate cross-device HAR using federated learning.