Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FedFixer, a federated learning method with a dual (global + personalized) model structure that alternately selects clean samples for each other under heterogeneous label noise, stabilized by confidence and distance regularizers.
Problem
Federated learning suffers from heterogeneous, client-specific label noise, where high-loss client-specific clean samples are hard to distinguish from noisy labels, so centralized noisy-label methods and existing federated noisy-label methods underperform.
Approach
A personalized model and a global model alternately select low-loss samples for each other (Co-teaching style), with a confidence regularizer (CR) against unconfident predictions and a distance regularizer (DR) tying the personalized model to the global model to prevent overfitting on small local noisy data.
Key results
- On CIFAR-10 with IID, rho=1, tau=0.5, FedFixer reaches 62.87% vs 54.22% for Global+CORES^2 and 52.30% for FedCorr.
- On CIFAR-10 non-IID rho=1, tau=0.5, FedFixer reaches 59.01% vs 49.49% for RFL.
- On Clothing1M non-IID, FedFixer achieves 70.52% vs 67.50% for FedCorr.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Distance Regularizer (ablation) | FedFixer · CIFAR-10 IID ρ=0.5, τ=0.3 | accuracy: 84.45 -> 87.06 | Helped |
| add the Confidence Regularizer (ablation) | FedFixer · CIFAR-10 IID ρ=0.5, τ=0.3 | accuracy: 75.73 -> 87.06 | Helped |
| use Alternate Updates (ablation) | FedFixer · CIFAR-10 non-IID ρ=1.0, τ=0.5 | accuracy: 54.84 -> 59.01 | Helped |
| include the Personalized Model (ablation) | FedFixer · CIFAR-10 non-IID ρ=1.0, τ=0.5 | accuracy: 34.44 -> 59.01 | Helped |
| use FedFixer instead of FedCorr | federated learning · Clothing1M non-IID | accuracy: 67.50 -> 70.52 | Helped |
| use FedFixer instead of FedCorr | federated learning · CIFAR-10 IID ρ=1, τ=0.5 | accuracy: 52.30 ± 0.91 -> 62.87 ± 0.17 | Helped |
Surprising findings
- Local + CORES^2 is nearly ineffective and can collapse when applied directly on local clients.
- FedCorr, which relies on having clean clients, becomes ineffective when all clients have noisy labels (rho=1).
Limitations
- The method does not include label correction, so it trails FedCorr by about 1% in some IID settings with lower noise.
- Performance on the most heterogeneous settings (e.g., CIFAR-10 rho=1, tau=0.5) remains modest at about 60% accuracy.
Experimental setup
Datasets: MNIST CIFAR-10 Clothing1M
Benchmarks: MNIST CIFAR-10 Clothing1M
Metrics: test accuracy F-score for label error detection
Baselines: Local + CORES^2 Global + CORES^2 FedAvg FedProx FedCorr RFL MR
Builds on: Co-teaching FedAvg FedCorr CORES^2
Closest papers by meaning
- FedDiv: Collaborative Noise Filtering for Federated Learning with Noisy Labels (AAAI 2024 · similarity 0.91)
- FedClean: A General Robust Label Noise Correction for Federated Learning (ICML 2025 · similarity 0.91)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.89)
- Federated Label-Noise Learning with Local Diversity Product Regularization (AAAI 2024 · similarity 0.89)
- Federated Learning with Extremely Noisy Clients via Negative Distillation (AAAI 2024 · similarity 0.89)
- Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.