3

In federated learning (FL), clients typically access data from just one distribution. Ideally, the learned models would generalize to out-of-distribution (OOD), i.e., domain generalization (DG). However, centralized DG methods cannot easily be …