Research map / Federated learning and privacy / Personalized and heterogeneous FL
Combating Data Imbalances in Federated Semi-supervised Learning with Dual Regulators
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 FedDure, a federated semi-supervised learning framework with coarse-grained and fine-grained regulators trained via bi-level optimization to handle both cross-client and within-client label/data distribution imbalance.
Problem
Federated semi-supervised learning where labeled and unlabeled data differ in distribution both across clients (external imbalance) and within a client between labeled and unlabeled data (internal imbalance), which causes existing FSSL methods to underperform even labeled-only training.
Approach
Two adaptive regulators: a C-reg that uses the cross-entropy change on labeled data as a reward to scale local updates on unlabeled data, and an F-reg that produces per-instance soft weights for unlabeled samples; both are updated alternately in a meta-learning style bi-level optimization, with the C-reg initialized from the server model to counter client drift.
Key results
- On CIFAR-10 under (DIR, DIR) setting FedDure reaches 57.73% accuracy vs. 45.56% for FedMatch, outperforming baselines by at least 11% on CIFAR-10 and CINIC-10.
- On CINIC-10 (IID, IID) FedDure achieves 56.36% vs. 43.73% for FedMatch.
- Ablation on CIFAR-10 (DIR, DIR): Baseline (FedAvg+FixMatch) 39.95%, Baseline+MAML 47.69%, FedDure 57.73%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FedDure's dual regulators instead of FedMatch | FSSL · CIFAR-10 (IID, DIR) | accuracy: 51.59 -> 66.85 | Helped |
| Add C-reg to the variant without C-reg (ablation) | FedDure · CIFAR-10 (DIR, DIR) | accuracy: 54.79 -> 57.73 | Helped |
| Add C-reg to the variant without C-reg (ablation) | FedDure · CIFAR-10 (IID, DIR) | accuracy: 64.98 -> 66.85 | Helped |
| Add C-reg to the variant without C-reg (ablation) | FedDure · CIFAR-10 (IID, IID) | accuracy: 65.41 -> 67.69 | Helped |
| Use client-specific F-reg rather than averaging F-reg across rounds (ablation) | FedDure · CIFAR-10 (DIR, DIR) | accuracy: 50.08 -> 57.73 | Helped |
| Use client-specific F-reg rather than averaging F-reg across rounds (ablation) | FedDure · CIFAR-10 (IID, DIR) | accuracy: 58.26 -> 66.85 | Helped |
| Use client-specific F-reg rather than averaging F-reg across rounds (ablation) | FedDure · CIFAR-10 (IID, IID) | accuracy: 59.13 -> 67.69 | Helped |
| Use dynamic F-reg weights instead of a fixed pseudo-label threshold (ablation) | FedDure · CIFAR-10 (DIR, DIR) | accuracy: 56.46 -> 57.73 | Helped |
| Use dynamic F-reg weights instead of a fixed pseudo-label threshold (ablation) | FedDure · CIFAR-10 (IID, DIR) | accuracy: 65.92 -> 66.85 | Helped |
| Use dynamic F-reg weights instead of a fixed pseudo-label threshold (ablation) | FedDure · CIFAR-10 (IID, IID) | accuracy: 66.54 -> 67.69 | Helped |
| Use client-specific C-reg rather than averaging C-reg across rounds (ablation) | FedDure · CIFAR-10 (DIR, DIR) | accuracy: 52.32 -> 57.73 | Helped |
| Use client-specific C-reg rather than averaging C-reg across rounds (ablation) | FedDure · CIFAR-10 (IID, DIR) | accuracy: 60.07 -> 66.85 | Helped |
Surprising findings
- Existing FSSL methods, and some semi-supervised baselines (e.g., on CIFAR-10 and Fashion-MNIST under DIR-DIR), perform worse than FedAvg trained only on labeled data, indicating unlabeled data can hurt under distribution mismatch.
Limitations
- Gains plateau as the number of selected clients grows beyond a threshold.
- Performance under weak class mismatch (IID, DIR) is slightly affected relative to the IID setting.
Experimental setup
Datasets: CIFAR-10 Fashion-MNIST CINIC-10
Benchmarks: CIFAR-10 Fashion-MNIST CINIC-10
Metrics: accuracy
Baselines: FedAvg* FedAvg-SL FedProx-SL FedAvg+UDA FedProx+UDA FedAvg+Fixmatch FedProx+Fixmatch FedMatch
Builds on: FedAvg FedMatch FixMatch MAML
Closest papers by meaning
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.92)
- SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning (AAAI 2025 · similarity 0.90)
- Progressive Distribution Matching for Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.90)
- 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)
- Fed-QSSL: A Framework for Personalized Federated Learning under Bitwidth and Data Heterogeneity (AAAI 2024 · 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.