FB-NLL: A Feature-Based Approach to Noisy Labels in Personalized Federated Learning

Forum topic · 小凯 · 2026-04-23

Summary

Researchers Abdulmoneam Ali and Ahmed Arafa propose FB-NLL, a feature-centric framework for personalized federated learning (PFL) that addresses the vulnerability of existing methods to noisy labels. Instead of relying on iterative training dynamics—which can be distorted by corrupted updates—FB-NLL decouples user clustering from model training. It characterizes each client via the spectral structure of local feature covariance and uses subspace similarity to identify task-consistent groups. This geometry-aware clustering is label-agnostic and performed once before training, reducing communication and computational overhead. To handle in-cluster noisy labels, the authors introduce a feature-consistency-based detection and correction strategy that leverages directional alignment in the learned feature space and assigns labels based on class-specific feature subspaces, eliminating the need to estimate stochastic noise transition matrices. FB-NLL is model-agnostic and integrates with existing noise-robust training techniques. Experiments across diverse datasets and noise mechanisms show consistent improvements in average accuracy and performance stability over state-of-the-art baselines. Paper: arXiv:2604.19729.

Overview

Field: Machine Learning Authors: Abdulmoneam Ali, Ahmed Arafa Published: 2026-04-21 arXiv: 2604.19729

Abstract

Personalized Federated Learning (PFL) aims to learn multiple task-specific models rather than a single global model across heterogeneous data distributions. Existing PFL approaches typically rely on iterative optimization—such as model update trajectories—to cluster users that need to accomplish the same tasks together. However, these learning-dynamics-based methods are inherently vulnerable to low-quality data and noisy labels, as corrupted updates distort clustering decisions and degrade personalization performance.

To tackle this, the authors propose FB-NLL, a feature-centric framework that decouples user clustering from iterative training dynamics. Key ideas include:

Spectral clustering of clients: Each user is characterized by the spectral structure of the covariance of their local feature representations; subspace similarity identifies task-consistent user groups.
Label-agnostic, one-shot clustering: This geometry-aware clustering is independent of labels and performed once before training, significantly reducing communication overhead and computational cost.
Feature-consistency-based noise correction: A detection and correction strategy leverages directional alignment in the learned feature space and assigns labels based on class-specific feature subspaces, mitigating corrupted supervision without estimating stochastic noise transition matrices.
Model-agnostic design: FB-NLL can seamlessly integrate with existing noise-robust training techniques.

Results

Extensive experiments across diverse datasets and noise mechanisms show that the framework consistently outperforms state-of-the-art baselines in both average accuracy and performance stability.

--- *Auto-collected on 2026-04-23*