AUTHOR=Chen Aiguo , Fu Yang , Sha Zexin , Lu Guoming TITLE=An EMD-Based Adaptive Client Selection Algorithm for Federated Learning in Heterogeneous Data Scenarios JOURNAL=Frontiers in Plant Science VOLUME=13 YEAR=2022 URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2022.908814 DOI=10.3389/fpls.2022.908814 ISSN=1664-462X ABSTRACT=

Federated learning is a distributed machine learning framework that enables distributed nodes with computation and storage capabilities to train a global model while keeping distributed-stored data locally. This process can promote the efficiency of modeling while preserving data privacy. Therefore, federated learning can be widely applied in distributed conjoint analysis scenarios, such as smart plant protection systems, in which widely networked IoT devices are used to monitor the critical data of plant production to improve crop production. However, the data collected by different IoT devices can be dependent and identically distributed (non-IID), causing the challenge of statistical heterogeneity. Studies have also shown that statistical heterogeneity can lead to a marked decline in the efficiency of federated learning, making it challenging to apply in practice. To promote the efficiency of federated learning in statistical heterogeneity scenarios, an adaptive client selection algorithm for federated learning in statistical heterogeneous scenarios called ACSFed is proposed in this paper. ACSFed can dynamically calculate the possibility of clients being selected to train the model for each communication round based on their local statistical heterogeneity and previous training performance instead of randomly selected clients, and clients with heavier statistical heterogeneity or bad training performance would be more likely selected to participate in the later training. This client selection strategy can enable the federated model to learn the global statistical knowledge faster and thereby promote the convergence of the federated model. Multiple experiments on public benchmark datasets demonstrate these improvements in the efficiency of the models in heterogeneous settings.