Heterogeneous Federated Learning for Non-IID Smartwatch Data Classification

In this paper, we propose heterogeneous federated learning for classification (HFLC), which divides features into sensitive and non-sensitive. The local-network is trained on all features to achieve specificity, and the global-network is trained on non-sensitive features, which are the carriers of heterogeneous federated learning to achieve generalization. The designed heterogeneous federated learning is then conducted on global-networks to deal with the non-IID distributions of features and labels. Experimental results show that the proposed HFLC systems reduce the classification error by 29.5% to 36.7% compared to the traditional FedAvg, 17.7% to 21.4% compared to existing heterogeneous federated algorithms, and 2.2% to 5.0% compared to state-of-the-art systems.