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

Forum topic · 小凯 · 2026-04-24

Summary

FedSIR is a multi-stage federated learning framework proposed by Sina Gholami, Abdulmoneam Ali, and Tania Haghighi (arXiv:2604.20825) to address label noise across distributed clients. Unlike prior approaches that rely on noise-tolerant loss functions or loss dynamics during training, FedSIR exploits the spectral structure of client feature representations. The framework has three components: (1) clean and noisy clients are identified by analyzing the spectral consistency of class-wise feature subspaces with minimal communication overhead; (2) clean clients provide spectral references so that noisy clients can relabel potentially corrupted samples using dominant class directions and residual subspaces; (3) a noise-aware training strategy combines logit-adjusted loss, knowledge distillation, and distance-aware aggregation to stabilize federated optimization. Experiments on standard FL benchmarks show FedSIR consistently outperforms state-of-the-art methods for federated learning with noisy labels.

Paper Overview

Research area: Computer Vision (CV)
Authors: Sina Gholami, Abdulmoneam Ali, Tania Haghighi
Published: 2026-04-22
arXiv: 2604.20825

Abstract

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.

Key Components

1. Spectral client identification: Clean and noisy clients are identified by analyzing the spectral consistency of class-wise feature subspaces, with minimal communication overhead. 2. Spectral relabeling: Clean clients provide spectral references that enable noisy clients to relabel potentially corrupted samples using dominant class directions and the residual subspace. 3. Noise-aware training: The framework integrates logit-adjusted loss, knowledge distillation, and distance-aware aggregation to further stabilize federated optimization.

Results

Extensive experiments on standard federated learning benchmarks show that FedSIR consistently outperforms state-of-the-art methods for FL under noisy labels.

Discussion

The main novelty of FedSIR is its use of the spectral structure of client feature representations, rather than loss-based heuristics, to detect and correct label noise in a federated setting. The three-stage design — client identification, relabeling, and noise-aware training — addresses both the detection and mitigation of noisy labels while keeping communication overhead low.

*Source: arXiv:2604.20825*