Search papers, labs, and topics across Lattice.
FedSIR is introduced, a federated learning framework robust to noisy labels that leverages the spectral structure of client feature representations to identify and relabel noisy data. Clean and noisy clients are distinguished by analyzing the spectral consistency of class-wise feature subspaces, with clean clients providing spectral references for noisy clients to relabel corrupted samples. A noise-aware training strategy, integrating logit-adjusted loss, knowledge distillation, and distance-aware aggregation, further stabilizes federated optimization, achieving SOTA results on standard FL benchmarks.
Spectral analysis of client feature representations can identify and relabel noisy data in federated learning, outperforming existing noise-tolerant loss and loss-dynamic approaches.
Federated learning (FL) enables collaborative model training without sharing raw data; however, the presence of noisy labels across distributed clients can severely degrade the learning performance. In this paper, we propose FedSIR, a multi-stage framework for robust FL under noisy labels. Different from existing approaches that mainly rely on designing noise-tolerant loss functions or exploiting loss dynamics during training, our method leverages the spectral structure of client feature representations to identify and mitigate label noise. Our framework consists of three key components. First, we identify clean and noisy clients by analyzing the spectral consistency of class-wise feature subspaces with minimal communication overhead. Second, clean clients provide spectral references that enable noisy clients to relabel potentially corrupted samples using both dominant class directions and residual subspaces. Third, we employ a noise-aware training strategy that integrates logit-adjusted loss, knowledge distillation, and distance-aware aggregation to further stabilize federated optimization. Extensive experiments on standard FL benchmarks demonstrate that FedSIR consistently outperforms state-of-the-art methods for FL with noisy labels. The code is available at https://github.com/sinagh72/FedSIR.
