Research map / Federated learning and privacy / Personalized and heterogeneous FL
BSemiFL: Semi-supervised Federated Learning via a Bayesian Approach
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 BSemiFL, which re-labels local data in semi-supervised federated learning by a Bayesian-weighted ensemble of global and local model pseudo-labels, with per-sample weights from estimated closeness to each model's training distribution.
Problem
Semi-supervised federated learning (SSFL) where clients hold unlabeled data and the server holds a few labels; the re-labeling step is error-prone when using only the global model (adaptation/distribution gap) or only the local model (local knowledge solidification and small data).
Approach
Combine global and local model predictions per sample, weighting each by a variational-Bayesian estimate of how close the sample is to each model's training distribution (the sampling probability), rather than using a fixed average or a single model; theory shows the ensemble loss is below the max of the two individual losses and below naive averaging under a stated condition.
Key results
- On CIFAR-10 with Shards (100 clients, 2 shards), BSemiFL reaches 78.31% test accuracy vs. 68.57% for Orchestra (about 9.8 points higher).
- On SVHN Shards (2 shards), BSemiFL reaches 94.52% vs. SemiFL's 88.36%.
- On CIFAR-100 with Shards (2 shards), Orchestra (39.93%) outperforms BSemiFL (39.09%), the one setting where BSemiFL is not best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Bayesian-weighted ensemble instead of simple averaging to combine global and local pseudo-label predictions (ablation) | BSemiFL · CIFAR-10, Shards (100, 2) | test accuracy | Helped |
| Increase warm-up period from 20 to 100 rounds (ablation) | BSemiFL · CIFAR-10, Shards (100, 2) | test accuracy: 78% -> about 79.5% | Helped |
| Vary pseudo-label threshold among 0.7, 0.8, and 0.9 (ablation) | BSemiFL · CIFAR-10, Shards (100, 2) | test accuracy: all around 78% | No clear effect |
| Increase server labeled dataset size from 250 to 500 samples (ablation) | BSemiFL · CIFAR-10, Shards (100, 2) | test accuracy: 68% to 78% (a 10% gain) | Helped |
| Increase server labeled dataset size from 500 to 1000 samples (ablation) | BSemiFL · CIFAR-10, Shards (100, 2) | test accuracy: 78% to 81% (a 3% gain) | Helped |
| Use BSemiFL rather than the strongest baseline | BSemiFL · CIFAR-10, Shards (100, 2) | test accuracy: 78.31±1.32 vs. 68.57±2.55; 9.8% | Helped |
| Use BSemiFL rather than SemiFL | BSemiFL · CIFAR-10, Dirichlet α=10 | test accuracy: 88.19% vs. 87.27% | Hurt |
| Use BSemiFL rather than Orchestra | BSemiFL · CIFAR-100, Shards (100, 2) | test accuracy: 39.09±3.12 vs. 39.93±2.17 | Hurt |
Surprising findings
- Local-only re-labeling yields a global model about 15% less accurate than global-only re-labeling, despite the local model being more accurate on local data.
- On CIFAR-10 with Dirichlet alpha=10, SemiFL slightly outperforms BSemiFL (88.19% vs. 87.27%).
Limitations
- Theoretical guarantee of the Bayesian ensemble (Theorem 6.2) holds only under an explicit conditional assumption that the authors argue is plausible via Bayes' theorem rather than proven generally.
- BSemiFL does not outperform Orchestra on CIFAR-100 with 2 shards, which the authors attribute to Orchestra being a self-supervised method suited to concentrated class distributions.
- Evaluation is limited to simulated image classification with three datasets and a fixed set of non-IID partitions.
Experimental setup
Datasets: SVHN CIFAR-10 CIFAR-100
Benchmarks: SVHN CIFAR-10 CIFAR-100
Metrics: test accuracy
Baselines: SemiFL FedMatch FedU FedEMA Orchestra (FL)^2 FedFAME pFedKnow
Builds on: SemiFL FixMatch FedMatch variational Bayesian inference
Closest papers by meaning
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.91)
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.91)
- Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.90)
- Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models (AAAI 2024 · similarity 0.89)
- Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch (CVPR 2025 · similarity 0.89)
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.88)
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.