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

Sina Gholami; Abdulmoneam Ali; Tania Haghighi; Ahmed Arafa; Minhaj Nur Alam

✨ TL;DR

FedSIR is a federated learning framework that identifies noisy labels and clients by analyzing spectral properties of feature representations, then uses clean clients to help relabel corrupted samples. The method combines spectral analysis with noise-aware training strategies to achieve robust collaborative learning despite label noise.

01 · Problem

Federated learning enables distributed model training without sharing raw data, but label noise across clients significantly degrades performance. Existing approaches typically rely on noise-tolerant loss functions or analyzing loss dynamics during training, which may not fully address the heterogeneous nature of label noise in federated settings where different clients may have different noise characteristics.

02 · Approach

FedSIR uses a three-stage framework: (1) identifies clean and noisy clients by examining spectral consistency of class-wise feature subspaces with minimal communication overhead, (2) leverages clean clients as spectral references to enable noisy clients to relabel corrupted samples using dominant class directions and residual subspaces, and (3) applies noise-aware training combining logit-adjusted loss, knowledge distillation, and distance-aware aggregation to stabilize federated optimization.

03 · Key insights

What the paper shows.

01Spectral structure of client feature representations provides a reliable signal for identifying label noise without requiring explicit noise rate estimation

02Clean clients can serve as spectral references to guide relabeling in noisy clients, enabling knowledge transfer across the federation

03Combining spectral-based relabeling with noise-aware training strategies (logit adjustment, distillation, distance-aware aggregation) provides complementary benefits for robust learning

04The approach achieves communication efficiency by analyzing spectral properties locally while only aggregating necessary information

04 · Results

FedSIR consistently outperforms state-of-the-art methods for federated learning with noisy labels on standard FL benchmarks. The framework effectively identifies clean and noisy clients, successfully relabels corrupted samples, and maintains training stability under various noise conditions and heterogeneous noise distributions across clients.

05 · Limitations

The paper does not explicitly discuss computational overhead of spectral analysis or scalability to very large numbers of clients. The approach assumes that some clients have clean labels to serve as spectral references, which may not hold in severely corrupted scenarios. The reliance on spectral properties may be sensitive to the dimensionality of feature representations and the number of classes, which are not thoroughly analyzed.

✨ Generated by Claude · Apr 25, 2026 · Read the PDF for authoritative content.

What the paper shows.

↘ Related papers