Asynchronous Online Federated Learning with Limited Storage
Ιοannis Schoinas, Anna Triantafyllou, Anastasios Drosou, Dimitrios Tzovaras, Panagiotis G. Sarigiannidis · 2024
The Internet of Things (IoT) generates vast amounts of data through sensors, enabling AI to train accurate models and develop Industry 4.0 (I4.0) systems that adapt to real-time changes. Federated Learning (FL), a distributed machine-learning paradigm, allows training across multiple devices while preserving data privacy by keeping the data local. However, classical FL’s synchronous aggregation strategy is inefficient for heterogeneous IoT devices. Additionally, Iot edge devices face storage limitations and generate continuous data streams, requiring adaptive online learning. We propose ASynchronous Online Limited Storage Federated Learning (ASOLS-Fed), where edge devices perform online learning on limited local data streams. ASOLS-Fed updates the global model asynchronously, mitigating straggler effects caused by slow or dropped devices. Experiments on benchmark datasets show that ASOLS-Fed converges well, maintaining stronger predictive performance even with limited training samples.