A Review on Federated Learning on Sensor-Based Human Activity Recognition

Deep learning has demonstrated exceptional human activity recognition (HAR) performance by extracting complex features from inertial data. However, this centralized training approach aggregates data from multiple user devices into a central server and raises significant privacy concerns. Federated learning (FL) is proposed as an alternative. It provides a privacy-preserving scheme by training data analytics models on local users’ devices rather than transferring raw data to a central server for data processing. Although FL is widely applied to various pattern recognition applications, its use in sensor-based HAR is limited, and reviews of the HAR application are even scarcer. Therefore, this paper provides a comprehensive review of FL in HAR. This paper analyzes FL’s architectural design, data model training strategies, and model aggregation techniques. A comparative analysis between FL-based and machine learning methods is presented. The challenges, including data heterogeneity, data privacy, and communication costs, are identified through the findings, while the potential research direction of FL in HAR is underscored. This paper provides insights into the current state of FL for HAR, pinpoints research gaps, and outlines encountered challenges and potential research directions.