HFSDN: Hierarchical Federated Learning Based SDN Architecture for Fault-Tolerance

Abstract

Federated learning (FL) offers an appealing way to train models across distributed devices without centralizing raw data, but deploying it in large, real-world networks quickly reveals serious limitations. Communication bottlenecks, unreliable connections, and the risk of a single aggregation server becoming a point of failure all undermine its practical viability. This paper proposes a Hierarchical Federated Learning-based SoftwareDefined Networking (HFSDN) architecture that addresses these problems by introducing a three-tier structure, in which SDN controllers serve as intermediate aggregation nodes rather than routing everything through a single central server. The key idea is straightforward: by distributing aggregation across multiple tiers, the system reduces the amount of data that travels across the network, contains the damage when individual nodes fail, and adapts more gracefully to changing network conditions. Sensitive data never leaves the edge devices - only model updates move through the network, routed via SDN-managed pathways. We evaluated HFSDN through extensive simulations and found meaningful improvements over standard Federated Averaging (FedAvg). On MNIST and FashionMNIST, HFSDN achieved 6.2% and 5.4% higher test accuracy, respectively, and reached 60% accuracy in 40 training rounds compared to 60 for FedAvg. Importantly, HFSDN retained 99% of its accuracy when half the clients dropped out, while FedAvg degraded by 5 percentage points under the same conditions.