Research map / Federated learning and privacy / Personalized and heterogeneous FL
(FL)$^2$: Overcoming Few Labels in 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 (FL)^2, a federated semi-supervised learning method for labels-at-server settings that combines client-specific adaptive thresholding, sharpness-aware consistency regularization on high-confidence samples, and learning status-aware aggregation to reduce confirmation bias with very few labels.
Problem
Federated semi-supervised learning in the labels-at-server setting has a large performance gap to centralized SSL when labeled data is scarce, mainly due to confirmation bias amplified by multiple local epochs and separation of labeled and unlabeled data.
Approach
Combines a client-specific adaptive pseudo-labeling threshold (driven by each client's global-model confidence), a SAM-style adversarial weight perturbation applied to a high-confidence subset with a consistency loss between original and perturbed outputs, and aggregation weights that favor clients with lower learning status (1 - threshold).
Key results
- Improves classification accuracy by up to 23.0% over existing FSSL methods, per the paper's claim.
- On Fashion-MNIST with 40 labels under non-IID Dir(0.3), (FL)^2 reaches 63.2% vs 12.8% for SemiFL.
- On AGNews with 20 labels under balanced IID, (FL)^2 reaches 87.0% vs 47.4% for SemiFL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use (FL)^2 instead of SemiFL | Federated semi-supervised learning · Fashion-MNIST, Unbalanced Non-IID Dir(0.3), 40 labeled samples | accuracy: 12.8(4.8) -> 63.2(0.5) | Helped |
| Use (FL)^2 instead of SemiFL | Federated semi-supervised learning · AGNews, Unbalanced Non-IID Dir(0.3), 20 labeled samples | accuracy: 59.1(13.7) -> 73.6(3.7) | Helped |
| Use (FL)^2 instead of SemiFL | Federated semi-supervised learning · Fashion-MNIST, Balanced IID, 40 labeled samples | accuracy: 10.0(0.0) -> 49.8(34.5) | Helped |
| Use (FL)^2 instead of SemiFL | Federated semi-supervised learning · AGNews, Balanced IID, 20 labeled samples | accuracy: 47.4(14.3) -> 87.0(0.6) | Helped |
| Apply naive SAM to client training | FSSL · FSSL benchmarks | classification accuracy | Hurt |
| Use the original SAM objective rather than sharpness-aware consistency regularization | FSSL · FSSL benchmarks | classification accuracy | No clear effect |
Surprising findings
- Naively applying SAM degrades FSSL performance because it also sharpens generalization on incorrectly pseudo-labeled samples.
- Minimizing the sharpness of incorrectly pseudo-labeled samples reduces model performance.
Limitations
- SemiFL training became unstable with the unlabeled batch size of 32, requiring a batch size of 10 for that baseline.
- Some runs fail under very few labels (e.g., one balanced IID Fashion-MNIST run collapsed to around 10%), indicating residual instability.
Experimental setup
Datasets: CIFAR-10 SVHN Fashion-MNIST AGNews
Benchmarks: CIFAR-10 SVHN Fashion-MNIST AGNews
Metrics: classification accuracy (%) standard deviation across three seeds
Baselines: SemiFL FedCon FedMatch FreeMatch SoftMatch
Builds on: FreeMatch SemiFL FixMatch Sharpness-Aware Minimization FedAvg
Closest papers by meaning
- Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.93)
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.92)
- Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch (CVPR 2025 · similarity 0.92)
- Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators (AAAI 2024 · similarity 0.92)
- Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models (AAAI 2024 · similarity 0.91)
- BSemiFL: Semi-supervised Federated Learning via a Bayesian Approach (ICML 2025 · similarity 0.91)
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.