FedSigmoid: a sigmoid-based dynamic federated learning algorithm
Jiayi Du, Xucheng Li, Taiyin Zhao · 2025
Federated learning is commonly used in security-sensitive areas, but it faces system heterogeneity problems. Numerous methods have been proposed to address this issue; however, existing solutions still exhibit limitations. Drawing inspiration from FedNova, in this research, we introduce the Sigmoid function to the aggregation weight calculation. Combining this innovative aggregation weight mechanism with an adaptive client-training strategy, we propose FedSigmoid. Finally, through comparative experiments on the EMNIST dataset, we demonstrate the effectiveness of our proposed FedSigmoid algorithm in addressing computational heterogeneity.