A hierarchical federated learning model with adaptive model parameter aggregation

Abstract

With the proposed Federated Learning (FL) paradigm based on the idea of ?data available but invisible?, participating nodes which create or hold data can perform local model training in a distributed manner, then a global model can be trained only by continuously aggregating model parameters or inter mediate results from different nodes, thereby achieving a balance between data privacy protection and data sharing. However, there are some challenges when deploying a FL model. First, there may be hierarchical associations between participating nodes, so that the datasets held by each node are no longer independent of each other. Secondly, due to the possible abnormal delay of data transmission, it can seriously influence the aggregation of model parameters. In response to the above challenges, this paper proposes a newly designed FL framework for the participating nodes with hierarchical associations. In this framework, we design an adaptive model parameter aggregation algorithm, which can dynamically decide the aggregation strategy according to the state of network connection between nodes in different layers. Additionally, we conduct a theoretical analysis of the convergence of the proposed FL frame work based on a non-convex objective function. Finally, the experimental results show that the proposed framework can be well applied to applications in different network connections, and can achieve faster model convergence efficiency while ensuring the accuracy of the model prediction.