Research on Edge oriented Federated Learning Training Optimization

Abstract

Federated learning, as a new distributed computing paradigm, reduces communication bandwidth consumption while ensuring data privacy and security, and can also utilize data from other terminal devices for collaborative training. However, in real edge computing scenarios, the data collected by edge devices usually have certain heterogeneity, which will lead to weight divergence, catastrophic forgetting of knowledge and other phenomena in the model training process of federated learning. Many existing federated learning methods improve from one direction of local client updates and global model updates, inevitably overlooking the impact of the other. Therefore, this article proposes a group based federated continuous learning method (FCL). Firstly, clients with similar data distributions are grouped together to reduce the weight differences between different clients within the same group; Then, during the local training process of the model, intelligent synaptic algorithm terms are introduced, and the learning of each group is modeled as a continuous learning task, in order to integrate knowledge between different local models and improve the model's ability to recognize and analyze old learning tasks. Experiments on the MNIST and CIFAR-10 standard datasets have shown that the model testing accuracy of FCL has improved by approximately 0.31% to 2.17% compared to FedProx, Scaffold, and FedCurve algorithms. Effectively enhancing the anti forgetting ability of the training model, further improving the convergence speed and accuracy of the training model.