Mitigating Label Noise in Federated Learning with Regularized Features and Robust Loss

Federated Learning enables decentralized model training across distributed devices while preserving data privacy and reducing communication overhead. However, the presence of heterogeneous client data and noisy labels poses significant challenges, often leading to degraded model performance, especially in communication-constrained environments. In this work, we propose FedRFRL, a robust federated learning framework designed to mitigate the adverse effects of label noise by combining feature regularization and a robust loss function. Specifically, FedRFRL introduces a representation regularization strategy that leverages a high-dimensional projection layer to promote the learning of diverse and discriminative features across clients. Additionally, we incorporate adaptive sharpness-aware minimization (ASAM) to encourage flatter minima for improved generalization under non-IID and noisy settings. Extensive experiments on benchmark datasets, including CIFAR-10, CIFAR-100, SVHN, Fashion-MNIST, and the real-world noisy datasets CIFAR-N and Clothing1M, demonstrate that FedRFRL consistently outperforms existing methods, achieving strong robustness against both symmetric and asymmetric label noise. These results highlight the effectiveness and scalability of FedRFRL in real-world federated learning scenarios with imperfect supervision.