Federated Learning with Heterogeneous Feature Adaptation for Human Activity Recognition

Federated learning promotes knowledge sharing in data-sensitive domains, such as Human Activity Recognition (HAR). However, data heterogeneity, namely, non-iid feature, can degrade the performance by causing client drift. We propose an effective knowledge distillation method incorporating a novel batch normalization setup within the federated learning aggregation process. This approach enables the global model to align closely with client models in parameter and feature spaces. Our method demonstrates superior performance and warm start capability compared to other approaches across various non-iid HAR datasets.