Answers are grounded in this paper and cite their sources.
Federated learning (FL) is a burgeoning distributed learning architecture aiming at privacy protection, where multiple clients collaborate to complete the model training task without sharing their private local data. However, clients may exhibit unpredictable dropout behaviors in the practical context of edge intelligence networks, which seriously deteriorates the convergence performance of the FL system. Additionally, the behavior of client dropout may aggravate the heterogeneity of local data distribution in FL, giving rise to a bias in the direction of global model updates and reducing the model performance. To counteract this issue, we propose an innovative FL algorithm named FedCFM, which adopts a model contrastive fusion scheme and a global aggregation mode with momentum to alleviate the negative impact of the dropouts. Specifically, FedCFM adds a modified contrastive fusion regularizer to maximize the consistency between local model and global model, which reducing the drift the global objective towards the local objective. FedCFM also adds a momentum term during global aggregation, using both accumulated gradients and current gradients to regularize the direction of global model updates and enhance system robustness. Experiments demonstrate that FedCFM significantly outperforms other baseline methods with regard to convergence performance and model accuracy.
No takes yet. Share an insight, caveat, or question.
Ren et al. (2025) studied this question.
Synapse has enriched 7 closely related papers on similar clinical questions. Consider them for comparative context: