Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Partial Label Learning with Local-Adaptive Augmentation and 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.
Formalizes federated partial label learning (FedPLL) and proposes FedPLL-LAAR, combining moving-average pseudo-label disambiguation, adaptive MixUp augmentation, and local-global gradient alignment regularization, evaluated on CIFAR-10/100 and SVHN under non-IID partial-label settings.
Problem
Federated learning with imprecise partial (candidate) labels on distributed clients, where local data is non-IID, scarce, and noisily annotated, causing client drift and label disambiguation difficulties.
Approach
Extends PLL into federated learning by maintaining per-instance pseudo-label vectors updated by moving average of local predictions, plus MixUp augmentation weighted by global-model confidence and a gradient-alignment regularizer between local and global gradients to reduce client drift.
Key results
- On CIFAR-10 with noise level 0.3 (non-IID, federated), FedPLL-LAAR reaches 90.35% accuracy vs. 73.56% for FedCorr, the best baseline.
- On CIFAR-100 with noise level 0.05, FedPLL-LAAR reaches 68.53% vs. 52.37% for RoFL, the best baseline.
- On SVHN with noise level 0.5, FedPLL-LAAR reaches 93.65% vs. 80.16% for FedCorr; the gap to centralized learning is relatively small.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedPLL-LAAR instead of FedCorr | FedPLL-LAAR · CIFAR-10, ρ = 0.1, 0.3, 0.5 | accuracy: 85.20±0.25% -> 93.09±0.20%; 73.56±0.84% -> 90.35±0.42%; 60.76±0.65% -> 77.66±0.52% | Helped |
| Use FedPLL-LAAR instead of FedCorr | FedPLL-LAAR · SVHN, ρ = 0.1, 0.3, 0.5 | accuracy: 95.29±0.13% -> 96.13±0.72%; 92.92±0.25% -> 95.81±0.56%; 80.16±0.28% -> 93.65±0.36% | Helped |
| Use FedPLL-LAAR instead of RoFL | FedPLL-LAAR · CIFAR-100, ρ = 0.01, 0.05, 0.1 | accuracy: 60.41±0.21% -> 73.27±0.15%; 52.37±0.31% -> 68.53±0.19%; 39.66±0.42% -> 60.66±0.10% | Helped |
| Add candidate label disambiguation (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 53.86% -> 90.35%; 73.77% -> 95.81%; 26.62% -> 68.53% | Helped |
| Add adaptive MixUp loss (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 82.61% -> 90.35%; 92.92% -> 95.81%; 50.89% -> 68.53% | Helped |
| Add gradient alignment loss (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 88.67% -> 90.35%; 94.34% -> 95.81%; 65.17% -> 68.53% | Helped |
| Add adaptive MixUp and gradient alignment losses (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 78.78% -> 90.35%; 91.36% -> 95.81%; 48.79% -> 68.53% | Helped |
| Add MixUp weight w1 (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 87.83% -> 90.35%; 94.78% -> 95.81%; 58.99% -> 68.53% | Helped |
| Add MixUp weight w2 (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 89.04% -> 90.35%; 94.80% -> 95.81%; 60.92% -> 68.53% | Helped |
| Add both MixUp weights w1 and w2 (ablation) | FedPLL-LAAR · CIFAR-10, SVHN, CIFAR-100 | accuracy: 86.61% -> 90.35%; 93.53% -> 95.81%; 54.56% -> 68.53% | Helped |
Surprising findings
- Ablation shows dropping candidate label disambiguation causes the largest drop (CIFAR-10 accuracy 90.35% to 53.86%).
Limitations
- Higher per-round communication cost due to additional gradient interchange.
- Evaluated only on three image datasets with simulated partial labels and non-IID partitions.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 SVHN
Benchmarks: CIFAR-10 CIFAR-100 SVHN
Metrics: test accuracy
Baselines: FedCorr RoFL FedX FedAvg
Builds on: FedAvg MixUp partial label generation procedure (Lv et al. 2020) FedCorr non-IID partition procedure
Closest papers by meaning
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.89)
- Realistic Evaluation of Deep Partial-Label Learning Algorithms (ICLR 2025 · similarity 0.88)
- Controller-Guided Partial Label Consistency Regularization with Unlabeled Data (AAAI 2024 · similarity 0.87)
- Language-Guided Transformer for Federated Multi-Label Classification (AAAI 2024 · similarity 0.87)
- Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.86)
- Exploiting Label Skews in Federated Learning with Model Concatenation (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.