FedPrime: An Adaptive Critical Learning Periods Control Framework for Efficient Federated Learning in Heterogeneity Scenarios

Haizhou Du, Zhiyuan Yang

Abstract: Federated learning (FL) is an emerging distributed optimization paradigm that learns from data samples distributed across many clients with data privacy protection in the artificial intelligence of things (AIoT). Adaptive client selection can improve FL efficiency in clients’ training progress. Still, it is not yet well understood, especially with clients’ data and client heterogeneity in the real world. Most existing FL methods assume that all clients in training phases are equally essential and have the same learning abilities with random selection in each round. However, this assumption has been proven invalid due to ignoring system heterogeneity and different critical learning periods (CLP) for each client. In this paper, we propose an adaptive critical learning periods control framework, FedPrime, to augment client selection in FL based on the fine-grained clients’ utility. We first adopt a fine-grained CLP detection for each heterogeneous client. Thus, we design an adaptive CLP control mechanism to fully leverage client selection in the training phase of FL. Extensive experiments based on various models and datasets validate that FedPrime framework achieves improved model accuracy up to 69.28% than the state-of-the-art methods. Moreover, FedPrime also keeps good generalization performance.