By analyzing the spectral structure of feature representations, you can identify noisy labels in federated learning and use clean clients to help relabel corrupted data—without needing to share raw data or redesign loss functions.
FedSIR tackles a major challenge in federated learning: when training data across distributed devices contains mislabeled examples. The method identifies which devices have clean vs. noisy labels by analyzing the mathematical structure of their learned features, then uses clean devices to help noisy devices fix their labels.