Fault-Tolerant Offloading of Federated Learning Training Tasks in IoT-Edge Systems
Bipul Bikram Thapa, Jianbo Zhang, Lena Mashayekhy · 2025
The rapid proliferation of Internet of Things (IoT) devices has led to an exponential increase in data generation at the edge, driving unprecedented demands for efficient and scalable distributed learning frameworks. Federated Learning (FL) offers a promising solution by enabling decentralized model training across devices while preserving data privacy. However, the computational limitations of resource-constrained IoT devices present challenges for timely local training. Offloading FL training tasks to edge servers provides a practical solution. However, in dynamic and failure-prone environments, ensuring fault tolerance during offloading becomes critical. This paper introduces a novel fault-tolerant offloading framework, called ResilientEdgeFL, designed to address these challenges by offloading FL client-side training tasks to edge servers, while incorporating fault tolerance to ensure efficient and reliable operations in dynamic IoT-edge environments. ResilientEdgeFL constructs a flow graph based on offloading costs and server constraints, and solves it using a mincost max-flow approach based on the network simplex method. Additionally, it incorporates an adaptive fault-tolerance solution that assigns backup servers and triggers fallback execution when a failure occurs, ensuring minimal disruptions to the training process. Through extensive simulations, we demonstrate that ResilientEdgeFL significantly improves offloading success rates and reduces FL training latency. Our findings highlight the importance of resilient fault-aware offloading for scalable FL deployments in real-world IoT-edge environments.