Two-stage model fusion scheme based on knowledge distillation for stragglers in federated learning

Federated learning, as an emerging distributed learning paradigm, enables devices (also called clients) storing local data to collaboratively participate in a training task without the data leaving the devices, aiming to achieve the effect of integrating multiparty data while meeting privacy protection requirements. However, the participating clients are autonomous entities in a real-world environment, with heterogeneity and network instability, which leads to FL being plagued by stragglers when intermediate training results are synchronously interacted. To this end, this paper proposes a new FL scheme with a two-stage fusion process based on knowledge distillation, which transfers knowledge of straggler models to the global model without delaying the training speed, thus balancing efficiency and model performance. We have evaluated the proposed algorithm on three popular datasets. The experimental results show that FedTd improves training efficiency and maintains good model accuracy compared to baseline methods under heterogeneous conditions, exhibiting strong robustness against stragglers. By our approach, the running time can be accelerated by 1.97– \(3.32\times\) $3.32\times$ under scenarios with higher level of data heterogeneity.