AI 中文总结
研究无线去中心化联邦学习收敛慢问题，提出预算感知、以簇为中心的DFL框架，通过划分簇和分配回程预算，在簇内快速并行聚合模型、簇头间不频繁交换，实现O(1/t)收敛速率，经实验验证相比基线能加速收敛。
AI 中文摘要
去中心化联邦学习（DFL）通过边缘设备间的对等模型交换摒弃了传统联邦学习的中央服务器。这种无服务器架构在大型设备到设备（D2D）网络中实现了即席、灵活的分布式学习。但无线DFL收敛缓慢，因对等模型聚合存在高延迟和错误。我们通过在掉队节点提供可靠回程链路增强网络连接来解决这些聚合瓶颈。基于此，我们的预算感知、以簇为中心的DFL框架先将网络划分为密集连接的簇，再将有限回程预算分配给选定簇头。两层协议在簇内快速并行模型聚合，簇头间进行不频繁的簇间交换，在t次迭代中收敛速率为O(1/t)。图像分类任务的数值实验证实了该方法相比现有DFL基线仅需少量策略性放置的回程链路就能加速收敛。
英文摘要
Decentralized federated learning (DFL) dispenses with the central server of classical FL by utilizing peer-to-peer model exchanges among edge devices. This server-free architecture enables ad-hoc, flexible distributed learning in large device-to-device (D2D) networks. However, wireless DFL converges slowly because peer-to-peer model aggregation incurs high delays and errors. Each DFL training round involves many-to-many gradient sharing over wireless channels, resulting in uncoordinated channel access, large communication errors from stragglers, and slow model consensus, especially in large-scale D2D networks with pronounced clustering structures. We address these aggregation bottlenecks by provisioning a few reliable backhaul links at straggling nodes to enhance network connectivity. Building on this idea, our budget-aware, cluster-centric DFL framework first partitions the network into densely connected clusters, and then allocates the limited backhaul budget to selected cluster heads. The resulting two-tier protocol executes fast, parallel model aggregation within clusters and infrequent inter-cluster exchanges among the heads, yielding an O(1/t) convergence rate in t iterations. Numerical experiments on image-classification tasks confirm that our approach accelerates convergence compared to state-of-the-art DFL baselines with only a few strategically placed backhaul links.
Comments6 pages, 3 figures, presented at 2025 IEEE Globecom Workshops
