FedSIR: Spectral Client Identification and Relabeling for Federated Learning with Noisy Labels

The newly introduced multi-stage framework, FedSIR, tackles the issue of noisy labels in federated learning. In contrast to current techniques that depend on noise-resistant loss functions or dynamics, FedSIR leverages the spectral characteristics of client feature representations to detect and reduce label noise. Initially, it discerns clean and noisy clients by examining the spectral consistency of class-specific feature subspaces, all while maintaining low communication overhead. Subsequently, clean clients supply spectral references, enabling noisy clients to accurately relabel their corrupted samples using dominant class directions and residual subspaces. Additionally, a noise-aware aggregation mechanism bolsters overall robustness. This method aims to enhance model performance in distributed environments with unreliable data labels.