FedOpenHAR: Federated Multitask Transfer Learning for Sensor-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Wearable Technology

Background:

• Wearable and mobile devices generate valuable user behavior data via motion sensors.
• Current analysis often focuses on single tasks like human activity recognition (HAR) using centralized deep learning.
• Distributed machine learning offers an alternative, avoiding data transmission to central servers.

Purpose of the Study:

• To introduce FedOpenHAR, a federated transfer learning framework for multitask learning.
• To enable simultaneous human activity recognition (HAR) and device position identification.
• To explore efficient model adaptation for new tasks and classes in a federated setting.

Main Methods:

• Utilized the OpenHAR framework with ten datasets for model training.
• Employed federated transfer learning with task-specific and personalized layers.
• Implemented the DeepConvLSTM architecture within the Flower federated learning environment.

Main Results:

• FedOpenHAR achieved 72.4% accuracy, outperforming centralized training (64.5%).
• Federated transfer learning performance was comparable to isolated individual training (72.6%).
• Demonstrated efficient adaptation for new clients with novel tasks or classes.

Conclusions:

• Federated transfer learning in FedOpenHAR enhances multitask analysis of wearable sensor data.
• The framework provides a robust and adaptable solution for distributed machine learning on sensor data.
• FedOpenHAR offers significant advantages for scalability and incorporating new functionalities in federated learning systems.