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

2026-04-22 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

FedSIR is a multi-stage framework designed to enhance robust federated learning (FL) performance in the presence of noisy labels across distributed clients. Unlike methods that focus on noise-tolerant loss functions or loss dynamics, FedSIR utilizes the spectral structure of client feature representations to identify and mitigate label noise. The framework operates through three main components: identifying clean and noisy clients via spectral consistency analysis of class-wise feature subspaces with minimal communication, enabling noisy clients to relabel corrupted samples using spectral references from clean clients, and employing a noise-aware training strategy. This strategy integrates logit-adjusted loss, knowledge distillation, and distance-aware aggregation to stabilize federated optimization. Experiments on standard FL benchmarks show FedSIR consistently outperforms existing methods for FL with noisy labels.

Key takeaway

For research scientists developing robust federated learning systems, FedSIR offers a novel approach to address label noise by leveraging spectral analysis. You should consider integrating spectral consistency checks and clean client spectral references into your noise mitigation strategies. This method provides a strong alternative to purely loss-function-based or loss-dynamics-based techniques, potentially leading to more stable and accurate model training.

Key insights

FedSIR uses spectral analysis of client features to identify and relabel noisy data in federated learning.

Principles

• Spectral consistency identifies client noise.
• Clean clients provide relabeling references.
• Noise-aware training stabilizes optimization.

Method

FedSIR identifies noisy clients via spectral consistency, relabels samples using clean client spectral references, then trains with logit-adjusted loss, knowledge distillation, and distance-aware aggregation.

In practice

• Analyze feature subspaces for noise.
• Use clean client data as spectral references.
• Integrate logit-adjusted loss for robustness.

Topics

• Federated Learning
• Noisy Labels
• Spectral Analysis
• Client Identification
• Label Relabeling

Code references

• sinagh72/FedSIR

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Federated Medical Image Segmentation under Real-World Label Noise: A Benchmark Suite for Noisy Label Learning Method Selection — The new benchmark suite enables realistic evaluation and selection of Federated Noisy Label Learning methods for medical image segmentation.
• FedIDM: Achieving Fast and Stable Convergence in Byzantine Federated Learning through Iterative Distribution Matching — FedIDM uses distribution matching and condensed data to achieve fast, stable, and robust federated learning convergence.
• Limits of spectral learning under noise — Spectral learning exhibits a fundamental noise threshold that limits functional structure recovery from noisy data.
• Frequency-Domain Latent Attention Gating for Cross-Domain Token Aggregation — FLaG enhances token aggregation by processing representations in the frequency domain, improving model performance across diverse tasks.
• VARS-FL: Validation-Aligned Client Selection for Non-IID Federated Learning in IoT Systems — Validation-aligned, history-aware client selection significantly improves federated learning convergence and stability in non-IID IoT environments.
• Beyond Weights and Gradients: A Taxonomy of Federated Learning Messages — Federated learning messages extend beyond weights, requiring a taxonomy for utility and privacy trade-offs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.