Standard federated learning, e.g., [FedAvg](http://proceedings.mlr.press/v54/mcmahan17a.html), assumes that all the participating clients build their local models with the same architecture, which limits its utility in real-world scenarios. In practice, clients can build their models with ***heterogeneous model architectures*** for specific local tasks. When faced with **data heterogeneity**, **model heterogeneity**, **communication overhead**, and **intellectual property (IP) protection**, Heterogeneous Federated Learning (HtFL) emerges.
0 commit comments