Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedClean: A General Robust Label Noise Correction for 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 FedClean, a label-noise correction framework for federated learning that works even when all clients are noisy, combining local centralized noisy-label learning, a two-stage label correction, and adaptive sample-size-weighted aggregation (ASSA).
Problem
Label noise in client-side datasets degrades federated learning global models, and existing federated label-noise correction methods assume clean clients exist and often assume specific noise distributions.
Approach
Each client runs local centralized noisy-label learning (Co-teaching or Joint Optim) to infer labels and select clean samples; a two-stage correction (inferred-label-dominated using a collaborative per-sample loss, then global-model-dominated) relabels noisy samples, and ASSA weights clients by their number of clean/corrected samples, without distinguishing clean from noisy clients.
Key results
- On CIFAR-10 IID with all clients noisy (rho=1, tau=0.5), FedClean2 reaches 83.75% accuracy versus 47.49% for FedCorr and 62.87% for FedFixer.
- On CIFAR-100 IID with rho=1, tau=0.5, FedClean2 achieves 66.54% versus 40.92% for FedCorr.
- On Clothing1M (non-IID, all clients noisy), FedClean2 attains 72.39% accuracy, the highest among compared methods (FedFixer 70.52%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedClean2 instead of FedCorr | Federated learning · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 47.49 ± 1.98 -> 83.75 ± 2.03 | Helped |
| Use FedClean2 instead of FedFixer | Federated learning · CIFAR-10 non-IID (ρ=1, τ=0.5) | accuracy: 59.01 ± 0.55 -> 80.55 ± 1.82 | Helped |
| Use FedClean2 instead of FedFixer | Federated learning · Clothing1M non-IID | accuracy: 70.52 -> 72.39 | Helped |
| Add local CNLL (ablation) | FedClean2 · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 47.74 -> 83.75 | Helped |
| Add both label-correction stages (ablation) | FedClean2 · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 76.45 -> 83.75 | Helped |
| Add correction stage I (ablation) | FedClean2 · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 82.58 -> 83.75 | Helped |
| Add correction stage II (ablation) | FedClean2 · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 79.85 -> 83.75 | Helped |
| Add ASSA (ablation) | FedClean2 · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 82.91 -> 83.75 | Helped |
| Add Mixup (ablation) | FedClean2 · CIFAR-10 IID (ρ=1, τ=0.5) | accuracy: 83.55 -> 83.75 | Helped |
| Add local CNLL (ablation) | FedClean2 · CIFAR-10 non-IID (ρ=1, τ=0.5) | accuracy: 40.11 -> 80.55 | Helped |
| Add both label-correction stages (ablation) | FedClean2 · CIFAR-10 non-IID (ρ=1, τ=0.5) | accuracy: 73.94 -> 80.55 | Helped |
| Add correction stage I (ablation) | FedClean2 · CIFAR-10 non-IID (ρ=1, τ=0.5) | accuracy: 79.19 -> 80.55 | Helped |
Surprising findings
- FedCorr, the strongest baseline when only some clients are noisy, collapses when all clients are noisy (e.g., 47.49% on CIFAR-10 IID at rho=1), while FedClean stays robust.
Limitations
- With no synthetic noise on some settings FedClean is not best; the authors attribute the gap to relatively basic CNLL methods used.
- The privacy extension zkCor* is described but not empirically evaluated; the theoretical propositions (Theorems 3.1-3.2) are informal with approximations in the proofs.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Clothing1M
Benchmarks: CIFAR-10 CIFAR-100 Clothing1M
Metrics: accuracy
Baselines: FedAvg FedProx FedCorr FedNoRo FedBeat FedELC FedFixer
Builds on: FedCorr Co-teaching Joint Optim Mixup FedAvg
Closest papers by meaning
- FedDiv: Collaborative Noise Filtering for Federated Learning with Noisy Labels (AAAI 2024 · similarity 0.93)
- FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning (AAAI 2024 · similarity 0.91)
- Federated Learning with Extremely Noisy Clients via Negative Distillation (AAAI 2024 · similarity 0.91)
- Federated Label-Noise Learning with Local Diversity Product Regularization (AAAI 2024 · similarity 0.88)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.87)
- Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch (CVPR 2025 · similarity 0.86)
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.