Research map / Federated learning and privacy / Personalized and heterogeneous FL
Learn How to Query from Unlabeled Data Streams in 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 LeaDQ, a QMIX-based multi-agent RL method that learns decentralized client-side data query policies for streaming unlabeled data in federated learning, guided by global model state and reward.
Problem
Selecting which streaming unlabeled samples to annotate on decentralized clients in federated learning, where local selection ignores the global training objective and data cannot be shared.
Approach
Frames federated data querying as a Dec-POMDP solved with QMIX under centralized training and decentralized execution: each client has a local Q-network over its arriving samples, mixed into a joint Q-value conditioned on a global state (confidence of the global model on a server held-out set), with reward equal to held-out accuracy gain.
Key results
- On SVHN with N_q/N_u=1/10, LeaDQ reaches 65.02% accuracy versus 64.67% for the best baseline KAFAL and 59.28% for Uncertainty.
- On CIFAR-100, LeaDQ surpasses the best baseline at later training stages by around 6%.
- With 50 clients on SVHN, LeaDQ attains 74.09% versus 73.57% for Coreset; KAFAL drops to 44.77%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LeaDQ instead of the strongest listed baseline | Federated querying on SVHN · SVHN, tested Nq/Nu ratios | accuracy (%) | Helped |
| use LeaDQ instead of the strongest listed baseline | Federated querying under distribution-based skew · SVHN, α=0.1, α=0.5, and α=1.0 | accuracy (%) | Helped |
| use LeaDQ instead of the strongest listed baseline | Federated querying under quantity skew · SVHN, C=2, C=3, and C=4 | accuracy (%) | Helped |
| use LeaDQ instead of the strongest listed baseline | Federated querying · Tiny-ImageNet, α=0.5 and α=0.1 | accuracy (%) | Helped |
| use LeaDQ instead of the strongest listed baseline | Federated querying with 50 clients · SVHN | accuracy (%) | Helped |
| vary episode length J across 5, 20, and 30 (ablation) | LeaDQ · SVHN | accuracy: 61.00% ; 59.79% ; 59.79% | No clear effect |
Surprising findings
- KAFAL collapses to about 20% accuracy under quantity-skew non-IID partitioning on SVHN, indicating existing federated active learning baselines can fail under heterogeneity.
- Episode length J has minimal impact on final accuracy in the reported range.
Limitations
- Global state requires a server-held held-out dataset assumed to follow the target distribution.
- Evaluated only with FedAvg and QMIX; more advanced FL and MARL algorithms are not explored.
- Simulation with 10 clients and small query budgets (N_q=1 per round) in main experiments.
Experimental setup
Datasets: SVHN CIFAR-100 20Newsgroup MNIST Tiny-ImageNet
Benchmarks: SVHN CIFAR-100 20Newsgroup Tiny-ImageNet
Metrics: model accuracy KL distribution divergence
Baselines: Uncertainty Coreset LoGo KAFAL
Builds on: QMIX FedAvg Dec-POMDP LoGo KAFAL
Closest papers by meaning
- FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning (NeurIPS 2025 · similarity 0.86)
- FedDiv: Collaborative Noise Filtering for Federated Learning with Noisy Labels (AAAI 2024 · similarity 0.86)
- (FL)$^2$: Overcoming Few Labels in Federated Semi-Supervised Learning (NeurIPS 2024 · similarity 0.85)
- Ranking-based Client Imitation Selection for Efficient Federated Learning (ICML 2024 · similarity 0.84)
- Subgraph Federated Learning for Local Generalization (ICLR 2025 · similarity 0.84)
- Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.84)
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.