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

Forum topic · 小凯 · 2026-04-24

Summary

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

Overview

Field: 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, the authors 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, FedSIR leverages the spectral structure of client feature representations to identify and mitigate label noise.

Key Components

1. 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: A training strategy integrating logit-adjusted loss, knowledge distillation, and distance-aware aggregation further stabilizes federated optimization.

Results

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

---

*Automatically collected on 2026-04-24.*