eXtreme Federated Learning (XFL): a layer-wise approach

Federated learning (FL) is a machine learning technique that builds models by using distributed data across devices. FL aggregates parameter updates from locally trained models, avoiding the need for user data exchange. However, for resource-constrained devices (e.g., IoT and Mobile), optimization in FL becomes crucial, particularly with large deep learning models. In this work, we introduce XFL (eXtreme Federated Learning), a new approach that aims to drastically reduce the amount of exchanged data by transmitting only a single layer of each client’s model in each round. Our main contribution lies in the development and evaluation of this layer-wise model aggregation strategy, which demonstrates its potential in significantly reducing communication costs. Validation experiments demonstrate up to 88.9% data reduction, with a small impact on the global model’s performance compared to the baseline algorithm. By effectively addressing the challenge of the communications, XFL enables more efficient and practical federated learning on resource-constrained devices.