Semi-Asynchronous DAG-Based Federated Learning for Internet of Things

Federated learning (FL) enables clients to participate in training, keeping sensitive data locally. FL for Internet of Things (IoT) faces significant challenges due to the heterogeneity of devices, frequent client dropouts, and the straggler effect inherent in traditional synchronous paradigms, where the aggregation occurs only when all local models are collected. Directed Acyclic Graph (DAG)-based FL records the global models and local models, enabling the aggregation to occur asynchronously. This paper proposes a Semi-Asynchronous DAG-based Federated Learning (SADAG-FL) framework where IoT devices can choose to train the model independently or employ edge servers for assistance. Dropout of IoT devices intra- or inter-groups is supported, and a novel straggler handing method is proposed. For a device reconnecting after a dropout, its local model is trained based on its local dataset and further updated based on the current global model and the series of expired global models recorded in the DAG during its disconnection period. The experimental results in CIFAR-10 demonstrate that our approach accelerates convergence and improves the accuracy of the model under dynamic dropout conditions, outperforming conventional synchronous and asynchronous FL baselines.