FedLoop: Heterogeneity Mitigation in Federated Learning

Federated Learning (FL), as a new privacy computing technology, has been researched and applied in many fields, such as the industrial internet of things (IIoT), intelligent media, and crowd sensing. However, clients' data easily be heterogeneous since they distribute in different environments, which causes performance degradation. To mitigate the impact of statistical heterogeneity, we proposed a novel federated learning framework FedLoop. Unlike conventional FL algorithms that aggregate client model parameters in every communication round, in FedLoop, the coordinate servers do not aggregate client models and just forward client models to others in each communication round. FedLoop only aggregates client model parameters to get a final model at the last iteration. Our framework can train client models with different client datasets without compromising user privacy. As the servers do not need to aggregate client models in each round, FedLoop has an advantage in computing costs. Experimental results demonstrate that FedLoop is competitive in mitigating the effects of statistical heterogeneity with a lower computing time cost.