EVENT DOSSIER
To address the deviation issues caused by intermittent client unavailability due to resource constraints or uncertainty in federated learning, researchers proposed the FedSWE algorithm. This algorithm achieves provable robustness towards heterogeneous and non-stationary random client availability through a hybrid local update mechanism that combines missing computation compensation, stable global updates, and implicit gossiping. Compared with the standard FedAvg, FedSWE introduces only minor memory and computational overhead, enabling linear acceleration in specific situations and converging to the stationary point of non-convex objective functions. The research team verified the analytical results and performance advantages of this algorithm through numerical experiments simulating diverse client unavailability dynamics on real datasets.
本文提出 FedSWE 算法，旨在解决联邦学习中因资源约束或不确定性导致客户端间歇性不可用引发的偏差问题。FedSWE 通过补偿缺失计算、稳定扩散全局更新及隐式 gossiping 混合本地更新，实现了对异构和非平稳随机客户端可用性的可证明鲁棒性。与标准 FedAvg 相比，该算法仅引入轻微的记忆和计算开销，能在特定情况下实现线性加速并收敛于非凸目标函数的驻点。研究者在真实数据集上通过多样化客户端不可用动态的数值实验验证了分析结果。