MemWAFL: Efficient Model Aggregation for Wireless Ad Hoc Federated Learning in Sparse Dynamic Networks
Koshi Eguchi, Hideya Ochiai, Hiroshi Esaki · 2023
In recent years, privacy-sensitive data has become increasingly prevalent in autonomous vehicles, smart devices, and sensor nodes that can move around and make opportunistic contact with each other. The federation of these nodes has been discussed in the context of federated learning with a centralized mechanism. However, due to multi-vendor issues, relying on a specific server operated by a third party is not desirable in some cases. To address this challenge, wireless ad hoc federated learning (WAFL) has been proposed to realize a fully distributed collaborative machine learning with the nodes physically encountered. WAFL can develop generalized models from non-lID datasets stored in distributed nodes by exchanging and aggregating them with each other over opportunistic node-to-node contacts. Despite these advancements, WAFL's learning speed and accuracy remain relatively slow in sparsely and dynamically connected situations because the model update frequency has to be lower than that in static and frequently connected situations. In this study, we propose MemWAFL, which improves learning accuracy and speed in sparsely and dynamically connected situations. We incorporated a mechanism to store the models of the nodes contacted in the WAFL learning algorithm. As a result, we successfully improved the average learning speed by up to 1209 epochs and the learning accuracy by up to approximately 0.42% compared to the original WAFL in sparsely and dynamically connected situations.