Research map / Federated learning and privacy / Personalized and heterogeneous FL
SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated 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 SemiDFL, a semi-supervised decentralized federated learning method combining neighborhood pseudo-labeling, consensus MixUp with a consensus-trained diffusion model, and accuracy-based adaptive aggregation, evaluated on MNIST, Fashion-MNIST and CIFAR-10.
Problem
Decentralized federated learning where clients have labeled, unlabeled, or mixed data under highly non-IID distributions, with no central server to coordinate semi-supervised learning.
Approach
Establishes consensus in both model and data spaces: pseudo-labels are estimated with neighborhood classifiers and a neighborhood-adaptive threshold, a diffusion model is trained and consensus-aggregated across clients to synthesize shared data for MixUp, and aggregation weights are set adaptively by each classifier's accuracy on generated data.
Key results
- On MNIST with alpha=100 and r=0.1% labeled, SemiDFL reaches 90.88% vs 71.84% for FlexMatch and 68.98% for DFL-LB.
- On CIFAR-10 with alpha=100 and r=5% labeled, SemiDFL reaches 69.40% vs 65.54% for MixMatch and 55.96% for DFL-LB.
- Ablation on CIFAR-10 shows neighborhood pseudo-labeling (NPL) improves over vanilla pseudo-labeling, e.g., 49.79% vs 32.13% at alpha=100, r=1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use neighborhood classifiers and neighborhood-qualified pseudo-label counts for filtering, instead of Vanilla PL or APL (ablation) | SemiDFL · CIFAR-10; α=100 and 0.1, r=5% and 1% | accuracy | Helped |
| Use consensus diffusion-generated data for MixUp instead of local-only MixUp (ablation) | SemiDFL · MNIST; α=100 and 0.1, r=0.5% and 0.1% | accuracy | Helped |
| Use adaptive weights based on generated-data accuracy instead of constant aggregation weights (ablation) | SemiDFL · Fashion-MNIST; α=100 and 0.1, r=0.5% and 0.1% or 1% | accuracy | Mixed |
Surprising findings
- Adaptive aggregation based on accuracy on generated data (AdaGen) performs slightly below adaptive aggregation on a shared test dataset (AdaTest) in some settings, e.g., Fashion-MNIST alpha=0.1, r=1%.
Limitations
- Experiments limited to small image benchmarks (MNIST, Fashion-MNIST, CIFAR-10) and simple CNN/ResNet-18 models.
- Relies on a shared-distribution generated validation set for aggregation weights, whose quality depends on the diffusion model.
Experimental setup
Datasets: MNIST Fashion-MNIST CIFAR-10
Benchmarks: MNIST Fashion-MNIST CIFAR-10
Metrics: averaged accuracy across clients standard deviation
Baselines: DFL-UB DFL-LB MixMatch FlexMatch CBAFed
Builds on: FixMatch MixMatch FlexMatch MixUp classifier-free guidance
Closest papers by meaning
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.92)
- Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models (AAAI 2024 · similarity 0.91)
- Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.91)
- Decentralized Sporadic Federated Learning: A Unified Algorithmic Framework with Convergence Guarantees (ICLR 2025 · similarity 0.91)
- BSemiFL: Semi-supervised Federated Learning via a Bayesian Approach (ICML 2025 · similarity 0.91)
- NTK-DFL: Enhancing Decentralized Federated Learning in Heterogeneous Settings via Neural Tangent Kernel (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.