A decentralized asynchronous federated learning framework for edge devices

Abstract

The traditional synchronous federated learning framework ensures global model consistency and accuracy. However, it is limited by the computational power differences between devices and the influence of non-IID data, which leads to inefficient training and insufficient model generalization performance. In this paper, we propose a decentralized asynchronous federated learning framework. The framework uses smart contracts deployed on the blockchain to manage edge devices for enhanced flexibility. At first, the framework performs model aggregation and validation through the use of consensus groups. It eliminates the potential single point of failure associated with centralized parameter servers. In addition, we propose a Federated Learning with Dynamically Growing Cache (FedDgc) method in a non-IID environment. The method reduces redundant gradient information exchange during initial feature extraction while maintaining the learning capability of the global model. Finally, the experimental results show that our framework has better test performance and guarantees the convergence speed of the model during training.