Cross-silo federated learning trains a shared model across geo-distributed data centers over wide-area networks whose heterogeneous, fluctuating bandwidth throttles the repeated model-exchange rounds. FEDCOD accelerates this communication at the application layer, decoupled from the learning algorithm so training accuracy is preserved: it applies a coding mechanism that exploits idle bandwidth via silo-to-silo (client-to-client) transfers and dynamically tunes coding redundancy to route around bottlenecks and absorb fluctuations, cutting average communication time by up to 62% in real-world experiments.
This paper introduces FEDCOD, an application-layer communication protocol for cross-silo federated learning, in which geo-distributed data centers ('silos') collaboratively train a shared model over …