StoCFL: A Stochastically Clustered Federated Learning Framework for Non-IID Data with Dynamic Client Participation
StoCFL is a new method for federated learning that groups clients with similar data distributions to handle Non-IID data better. It uses stochastic client clustering to dynamically decide how many clusters exist and bi-level optimization where cluster models share knowledge via a global model. This allows the system to work even when the number of participating clients changes during training.