FCS: A Fault-Tolerance Asynchronous Federated Learning System with Model Checkpoint Storage

Federated Learning (FL) is a decentralized machine learning framework that enables large-scale training across multiple devices while maintaining strong privacy guarantees. However, as the number of participating devices increases, communication heterogeneity creates significant challenges for the aggregation server, especially in maintaining reliable communication and guaranteeing training progress for all devices. To ensure FL training stability and relax the strict synchronous communication requirements, we propose the Federated Check-point Storage (FCS) protocol. FCS is a message-free coordination protocol that supports fully asynchronous FL model aggregation by leveraging model checkpointing storage to decouple the strict communication dependency of the device and server during the FL training process. Specifically, FCS isolates the communication instructions required by the server to manage the availability of the device's local training and weight submission from the server's global model aggregation process. We show that the FCS protocol better handles intermittent device availability and discuss its robustness property in avoiding communication deadlock. Experiment results demonstrate that the FCS protocol improves both the round-to-accuracy convergence rates and the time-to-target accuracy of the FL training tasks on FEMNIST and CIFAR-10 datasets. Compared to state-of-the-art synchronous FL protocols under intermittent device availability, we show that FCS could be a robust FL training system for heterogeneous device networks of any scale.