Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedDiv: Collaborative Noise Filtering for Federated Learning with Noisy Labels
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 FedDiv, a one-stage federated learning with noisy labels framework that learns a global Gaussian-mixture noise filter from aggregated client loss statistics plus a predictive-consistency sampler for credible local training samples.
Problem
Federated learning with noisy labels (F-LNL) where clients hold noisy or clean labels; prior local noise filters use only private data and are unstable and sub-optimal.
Approach
Build a federated noise filter by aggregating per-client two-component GMM parameters fit to per-sample losses (federated EM), so each client filters noise using knowledge from all clients, and select training samples via prediction consistency between global and de-biased local models.
Key results
- On CIFAR-100 IID with (rho, tau)=(0.8, 0.5), FedDiv outperforms FedCorr by 6.39% per the paper's text.
- On CIFAR-10 non-IID with (p, alpha)=(0.3, 10), FedDiv reaches 85.31% vs FedCorr 81.57%, a 3.74% gain.
- On Clothing1M, FedDiv reaches 72.96% vs FedCorr 72.55%, a 0.41% improvement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace FedCorr with FedDiv | CIFAR-10 IID · CIFAR-10 IID, ρ=0.8, τ=0.5 | test accuracy: 90.59±0.70 -> 91.44±0.25 | Helped |
| replace FedCorr with FedDiv | CIFAR-10 non-IID · CIFAR-10 non-IID, p=0.3, αDir=10, ρ=0.6, τ=0.5 | test accuracy: 81.57±3.68 -> 85.31±2.28 | Helped |
| replace FedCorr with FedDiv | CIFAR-100 IID · CIFAR-100 IID, ρ=0.8, τ=0.5 | test accuracy: 59.10±5.12 -> 65.49±2.20 | Helped |
| replace FedCorr with FedDiv | CIFAR-100 non-IID · CIFAR-100 non-IID, p=0.7, αDir=10, ρ=0.4, τ=0.0 | test accuracy: 72.73±1.02 -> 74.47±0.34 | Helped |
| replace FedCorr with FedDiv | Clothing1M non-IID · Clothing1M non-IID | test accuracy: 72.55 -> 72.96±0.43 | Helped |
| use federated filtering instead of current-round-only filter aggregation (ablation) | FedDiv · CIFAR-10 non-IID, p=0.3, αDir=10, ρ=0.6, τ=0.5 | test accuracy: 83.22±2.61 -> 85.31±2.28 | Helped |
| use federated filtering instead of current-round-only filter aggregation (ablation) | FedDiv · CIFAR-100 non-IID, p=0.7, αDir=10, ρ=0.4, τ=0.0 | test accuracy: 73.06±0.93 -> 74.47±0.34 | Helped |
| replace separate local filters with the federated noise filter (ablation) | FedDiv · CIFAR-10 non-IID, p=0.3, αDir=10, ρ=0.6, τ=0.5 | test accuracy: 81.34±3.65 -> 85.31±2.28 | Helped |
| replace separate local filters with the federated noise filter (ablation) | FedDiv · CIFAR-100 non-IID, p=0.7, αDir=10, ρ=0.4, τ=0.0 | test accuracy: 71.37±0.76 -> 74.47±0.34 | Helped |
| add noisy-sample relabeling and PCS (ablation) | FedDiv · CIFAR-10 non-IID, p=0.3, αDir=10, ρ=0.6, τ=0.5 | test accuracy: 82.17±3.06 -> 85.31±2.28 | Helped |
| add noisy-sample relabeling and PCS (ablation) | FedDiv · CIFAR-100 non-IID, p=0.7, αDir=10, ρ=0.4, τ=0.0 | test accuracy: 73.09±1.89 -> 74.47±0.34 | Helped |
| add predictive consistency-based sample reselection (ablation) | FedDiv · CIFAR-10 non-IID, p=0.3, αDir=10, ρ=0.6, τ=0.5 | test accuracy: 82.83±2.59 -> 85.31±2.28 | Helped |
Limitations
- Gains over FedCorr are small on Clothing1M and some settings have large standard deviations.
- Baseline results are borrowed from FedCorr's reported numbers rather than rerun.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Clothing1M
Benchmarks: CIFAR-10 CIFAR-100 Clothing1M
Metrics: best test accuracy standard deviation over five trials
Baselines: FedAvg RoFL ARFL JointOpt DivideMix FedCorr
Builds on: FedAvg FedCorr DivideMix MixUp
Closest papers by meaning
- FedClean: A General Robust Label Noise Correction for Federated Learning (ICML 2025 · 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 0.89)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.89)
- Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 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.