Flow: Per-Instance Personalized Federated Learning Through Dynamic Routing

Personalization in Federated Learning (FL) aims to modify a collaboratively trained global model according to each client. Current approaches to personalization in FL are at a coarse granularity, i.e. all the input instances of a client use the same personalized model. This ignores the fact that some instances are more accurately handled by the global model due to better generalizability. To address this challenge, this work proposes Flow, a fine-grained stateless personalized FL approach. Flow creates dynamic personalized models by …

client cs.lg current dynamic federated learning flow global instance instances per personalization personalized routing through