Answers are grounded in this paper and cite their sources.
Federated learning (FL) is a prominent machine learning (ML) framework that enables training a global model across distributed client devices without exposing their private data. However, traditional FL methods often face challenges in achieving fast convergence, particularly in heterogeneous or non-independent and identically distributed (non-IID) data environments. These issues arise primarily from inaccurate gradient estimation during the optimization process. To address these limitations, this article introduces an adaptive optimizer, federated fractional gradient descent with adaptive parameter updates (FedFGD-APUs). In particular, the proposed FedFGD-APU optimizer leverages fractional-order derivatives for efficient gradient computation. It employs an adaptive strategy to dynamically adjust fractional order, learning rate, and momentum to improve learning convergence and stability. We also develop a theoretical computational framework linking federated optimization with adaptive fractional derivatives, establishing the convergence of FedFGD-APU for nonconvex optimization settings with a convergence rate of O(1/T) . We conduct comprehensive experiments on MNIST, CIFAR-10, and NSL-KDD datasets to validate FedFGD-APU's effectiveness under both IID and non-IID data distributions. The experimental results demonstrate that FedFGD-APU significantly improves communication speedups and testing accuracy across heterogeneous data scenarios.
No takes yet. Share an insight, caveat, or question.
Uddin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: