Clustered DFL framework accelerates wireless model training with limited backhaul
Peer-to-peer learning gains speed via strategic clustering and just a few backhaul links.
Decentralized federated learning (DFL) removes the central server bottleneck by enabling peer-to-peer model exchanges among edge devices. However, in large-scale device-to-device (D2D) networks, uncoordinated channel access and stragglers cause high delays and slow consensus. Zheng, Liu, and Zhang tackle this by first partitioning the network into densely connected clusters based on device proximity. They then provision a limited budget of reliable backhaul links only at selected cluster heads, avoiding the cost of fully connected infrastructure.
The resulting two-tier protocol runs fast, parallel model aggregation inside each cluster over D2D links and infrequent inter-cluster exchanges among heads via backhaul. The authors mathematically prove an O(1/t) convergence rate, meaning training error drops inversely with iterations. Numerical experiments on image classification confirm that this cluster-centric DFL significantly accelerates convergence compared to existing DFL baselines, using just a handful of strategically placed backhaul links. The work shows that a small infrastructure investment can dramatically improve the speed and reliability of distributed AI training in wireless networks.
- Network is partitioned into dense clusters for fast intra-cluster aggregation over D2D links.
- Limited backhaul budget is allocated only to cluster heads for inter-cluster model exchange.
- Achieves O(1/t) convergence rate and outperforms state-of-the-art DFL baselines in image classification experiments.
Why It Matters
Practical DFL at scale becomes feasible with minimal backhaul investment, accelerating edge AI training.