Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Label-Noise Learning with Local Diversity Product Regularization
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 FedLNL, a federated label-noise learning scheme that estimates a local noise transition matrix via Bayesian Dirichlet updates alternated with classifier training, plus a decomposable local diversity product regularizer, and reports gains under limited local data.
Problem
Label noise in federated learning degrades models, and existing FL label-noise methods perform poorly when clients have only limited local training samples; accurately estimating the noise transition matrix locally is the key difficulty.
Approach
Replace gradient-based local NTM learning with a Bayesian Dirichlet-categorical update of concentration parameters (using confusion counts estimated from the local classifier), alternated with classifier training, and make the VolMinNet-style volume regularizer decomposable across clients via a local diversity product regularizer that provably preserves the optimum.
Key results
- On CIFAR-100 (flip-0.4), FedLNL reaches 68.14% test accuracy versus 48.69% for the second-best FedCorr, an improvement of up to 25.98% across settings.
- On CIFAR-10 with limited local data (100 clients), FedLNL achieves 68.14%-88.63% across flip-0.4 and flip-0.45 settings, improving flip-0.45 by 19.51% over FedLSR.
- On Clothing1M, FedLNL reaches 73.36% (i.i.d.) and 72.95% (non-i.i.d.) versus 69.94% and 68.96% for FedCorr.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the LDP regularizer instead of the regularizers in S-adaptation-FL and VolMinNet-FL (ablation) | FedLNL-G · CIFAR-10 and CIFAR-100 with limited local training samples | test accuracy: 19.16%; 21.30% | Helped |
| Use Bayesian alternating updates for the local NTM instead of gradient-based updates (ablation) | FedLNL · CIFAR-100 with limited local training samples | test accuracy | Helped |
| Use Bayesian alternating updates for the local NTM instead of gradient-based updates (ablation) | FedLNL · CIFAR-10, sufficient and limited local training samples | test accuracy | No clear effect |
| Use the LDP regularizer to enable federated optimization (ablation) | FedLNL · CIFAR-10 and CIFAR-100 with limited local training samples | test accuracy | Helped |
| Use FedLNL instead of FedCorr | CIFAR-100 · pair-flipping noise rate 0.45 | test accuracy: 25.98% | Helped |
| Use FedLNL instead of FedLSR | CIFAR-10 · pair-flipping noise rate 0.45 | test accuracy: 19.51% | Helped |
| Use FedLNL instead of FedLSR | CIFAR-10 · asymmetric noise rate 0.4 | test accuracy: 8.78% | Helped |
| Use FedLNL instead of FedCorr | CIFAR-10 · pair-flipping noise rate 0.4, non-i.i.d. data | test accuracy: 17.74% | Helped |
| Use FedLNL instead of FedCorr | Clothing1M · i.i.d. and non-i.i.d. data partitions | test accuracy | Helped |
Surprising findings
- Gradient-based local NTM learning (S-adaptation, VolMinNet) degrades sharply with limited local samples, with accuracy drops of 19.16% and 21.30% reported on CIFAR-10.
Limitations
- Assumes class-dependent, instance-independent label noise (T(x)=T).
- Relies on the local classifier being reasonably well trained to estimate confusion matrices.
- Non-i.i.d. evaluation only on CIFAR-10 and Clothing1M; FedLSR underperforms on Clothing1M i.i.d.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Clothing1M
Benchmarks: CIFAR-10 CIFAR-100 Clothing1M
Metrics: test accuracy
Baselines: FedAvg S-adaptation-FL VolMinNet-FL DivideMix-FL RoFL FedLSR FedCorr
Builds on: VolMinNet FedAvg Dirichlet-categorical conjugacy forward correction loss
Closest papers by meaning
- FedDiv: Collaborative Noise Filtering for Federated Learning with Noisy Labels (AAAI 2024 · similarity 0.89)
- FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning (AAAI 2024 · similarity 0.89)
- FedClean: A General Robust Label Noise Correction for Federated Learning (ICML 2025 · similarity 0.88)
- Federated Learning with Extremely Noisy Clients via Negative Distillation (AAAI 2024 · similarity 0.86)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.86)
- Federated Partial Label Learning with Local-Adaptive Augmentation and Regularization (AAAI 2024 · 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.