FedUIG: User Importance-Guided Federated Learning with Noisy and Heterogeneous Clients

Heterogeneous Federated Learning (HFL) is a challenging task that allows clients with different model structures to engage in collaborative training. In real-world scenarios, due to the difficulty of accurate annotation, clients may have different proportions of noisy labels. Incorrect labels can lead to performance degradation of models in FL. In this paper, we formulate a more realistic federated learning problem with noisy labels where client data, client model and the noisy labels are heterogeneous. Furthermore, we propose a multi-stage framework, FedUIG, designed to identify clean clients and quantify the importance of clients. In particular, it contains three modules: (1) Local Robust Learning Module: we apply a noise-tolerant loss function based on semi-supervised training. (2) User Importance Guidance Module: FedUIG identifies clean clients based on the independently measured training loss reduction rate for each client, and then FedUIG re-weights noisy clients by measuring prediction difference between noisy and clean clients on public dataset. (3) Robust Knowledge Distillation Module: heterogeneous clients engage in collaborative training through knowledge distillation based on public dataset. Extensive experiments conducted on the widely-used CIFAR-IO and ANIMAL-ION datasets, with varying noise levels, demonstrate that FedUIG is robust to noisy datasets and outperforms state-of-the-art methods across multiple noise levels.