Research map / Federated learning and privacy / Personalized and heterogeneous FL
Progressive Distribution Matching for Federated Semi-Supervised 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 FedPDM, a federated semi-supervised learning method that progressively matches pseudo-label distributions to an EMA-estimated class prior via Sinkhorn-Knopp optimal transport, plus logit adjustment for local skew.
Problem
Federated semi-supervised learning suffers from pseudo-label distribution mismatch caused by skewed, heterogeneous unlabeled client data, which existing FSSL methods do not explicitly address.
Approach
Formulate pseudo-label refinement as an entropy-regularized optimal transport problem that matches EMA-model predictions on unlabeled data to an estimated class prior, solved with Sinkhorn-Knopp, and combine it with logit adjustment for debiased local training, all without extra client-server communication.
Key results
- On CIFAR-100 under the Mixed scenario, FedPDM surpasses the best competitor by 4.7% (72.14 vs 67.57 for CBAFed).
- Under aggressive heterogeneity (alpha=0.1), FedPDM reaches 78.96% on CIFAR-10 versus 72.61% for CBAFed.
- On ISIC2018 FedPDM drops only 1.96% when heterogeneity increases, while FedIRM drops 5.11%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedPDM instead of CBAFed | FSSL · Fashion-MNIST, Mixed | accuracy: 83.62 -> 85.19 | Helped |
| Use FedPDM instead of RSCFed | FSSL · SVHN, Mixed | accuracy: 84.25 -> 87.69 | Helped |
| Use FedPDM instead of CBAFed | FSSL · CIFAR-10, Mixed | accuracy: 80.20 -> 83.74 | Helped |
| Use FedPDM instead of CBAFed | FSSL · CIFAR-100, Mixed | accuracy: 67.57 -> 72.14 | Helped |
| Use FedPDM instead of RSCFed | FSSL · ISIC2018, Mixed | accuracy: 68.59 -> 69.50 | Helped |
| Use FedPDM instead of CBAFed | FSSL · Fashion-MNIST, Pure | accuracy: 81.49 -> 85.76 | Helped |
| Use FedPDM instead of CBAFed | FSSL · SVHN, Pure | accuracy: 89.90 -> 94.21 | Helped |
| Use FedPDM instead of CBAFed | FSSL · CIFAR-10, Pure | accuracy: 69.65 -> 74.02 | Helped |
| Use FedPDM instead of CBAFed | FSSL · CIFAR-100, Pure | accuracy: 43.66 -> 47.71 | Helped |
| Use FedPDM instead of CBAFed | FSSL · ISIC2018, Pure | accuracy: 72.05 -> 73.89 | Helped |
| Use FedPDM instead of CBAFed | FSSL · CIFAR-10, Mixed, α = 0.1 | accuracy: 72.61 -> 78.96 | Helped |
| Use FedPDM instead of CBAFed | FSSL · CIFAR-100, Mixed, α = 0.1 | accuracy: 63.60 -> 69.40 | Helped |
Surprising findings
- Existing FSSL methods are upper-bounded by FedAvg on fully-labeled data according to the authors' generalization analysis.
- FSSL pseudo-labeling with FixMatch shows predictions strongly biased toward incorrect classes even early in training, with bias worsening over rounds.
Limitations
- Relies on an estimated class prior of unlabeled data, which can be inaccurate early in training (mitigated by linear rampup).
- Experiments limited to image classification datasets with a single ResNet18 main backbone and 10 clients.
Experimental setup
Datasets: Fashion-MNIST SVHN CIFAR-10 CIFAR-100 ISIC2018
Benchmarks: Fashion-MNIST SVHN CIFAR-10 CIFAR-100 ISIC2018
Metrics: accuracy multi-class F1-score
Baselines: FedIRM RSCFed CBAFed FedAvg
Builds on: FixMatch FedAvg Cuturi 2013 Sinkhorn distances logit adjustment RSCFed
Closest papers by meaning
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.93)
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.91)
- Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models (AAAI 2024 · similarity 0.91)
- Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch (CVPR 2025 · similarity 0.90)
- Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.90)
- BSemiFL: Semi-supervised Federated Learning via a Bayesian Approach (ICML 2025 · similarity 0.90)
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.