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

Paper: 2604.19729 Authors: Abdulmoneam Ali, Ahmed Arafa Categories: cs.LG, cs.AI

Problem

Personalized Federated Learning (PFL) aims to learn multiple task-specific models across heterogeneous data distributions. Existing approaches rely on iterative optimization (model update trajectories) to cluster users, but these are:

• Inherently vulnerable to low-quality data and noisy labels
• Corrupted updates distort clustering decisions
• Degrade personalization performance

Key Insight

Learning-dynamics-based methods fail with noisy data because corrupted updates misrepresent true user relationships.

FB-NLL Framework

A feature-centric framework that decouples user clustering from iterative training dynamics:

1. Geometry-aware clustering: Uses spectral structure of feature representation covariances
2. Subspace similarity: Leverages intrinsic heterogeneity of local feature spaces
3. One-shot clustering: Performed prior to training, reducing communication overhead

Key Innovation: Feature-Consistency-Based Correction

Within clusters, detects and corrects noisy labels by:

• Leveraging directional alignment in learned feature space
• Assigning labels based on class-specific feature subspaces
• Label-agnostic and robust to noise

Results

Across diverse datasets and noise regimes:

• Consistently outperforms state-of-the-art baselines
• Better average accuracy and performance stability
• Model-independent and integrates with existing noise-robust techniques

Takeaways

• Feature-based clustering is more robust to noise than dynamics-based approaches
• One-shot clustering reduces communication and computation costs
• Geometry-aware methods enable effective PFL in noisy environments