Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedAvP: Augment Local Data via Shared Policy 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 FedAvP, a federated data augmentation method that shares only augmentation policies (not data or data statistics) among clients, using a first-order federated meta-policy loss and joint model-policy training.
Problem
Federated learning suffers from limited and heterogeneous local data, but existing federated data augmentation methods share input-level or feature-level information that can leak private data via gradient matching reconstruction attacks, and standard automated augmentation policy searches assume ce
Approach
Share only augmentation policy parameters across clients: a Federated Meta-Policy Loss treats the policy loss as a meta-update loss over FedAvg rounds, approximated to first order so gradients stay within each client, and a Reptile-style meta-learned policy network yields personalized policies per client, trained jointly with the model.
Key results
- On CIFAR-100 with alpha=0.1, FedAvP reaches 45.96% test accuracy versus 41.21% for the best default-augmentation FedFA baseline (FedFA + Default 41.21%).
- On SVHN with alpha=0.1, FedAvP achieves 89.81% versus 87.60% for FedDyn + Default.
- FedAvP (Fast Update) adds only about 0.38 MB per round over FedAvg (15.73 vs 15.35 MB) while sharing no data statistics; reconstruction PSNR for FedAvP policy gradients (8.77) is comparable to FedAvg-level leakage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use shared policy aggregation instead of training each client with only its own local policy (ablation) | FedAvP · CIFAR-100 (α = 0.1) | test accuracy: 43.86 -> 45.96 | Helped |
| Use shared policy aggregation instead of training each client with only its own local policy (ablation) | FedAvP · SVHN (α = 0.1) | test accuracy: 87.05 -> 89.81 | Helped |
| Update policies periodically rather than at every local step (ablation) | FedAvP · CIFAR-100 (α = 5.0) | test accuracy: 50.47 -> 49.97 | Hurt |
| Update policies periodically rather than at every local step (ablation) | FedAvP · FEMNIST | test accuracy: 84.27 -> 84.47 | No clear effect |
| Set policy exploration regularization ϵ to 0.4 (ablation) | FedAvP · CIFAR-10 (α = 5.0) | test accuracy: 83.78 -> 83.98 | No clear effect |
| Use a 3-layer augmentation search space rather than 2 layers (ablation) | FedAvP (Fast Update) · SVHN (α = 0.1), round 500 | test accuracy: 87.86 -> 92.07 | Helped |
| Use a 3-layer augmentation search space rather than 2 layers (ablation) | FedAvP (Fast Update) · SVHN (α = 5.0), round 500 | test accuracy: 95.14 -> 95.01 | Hurt |
| Use FedAvP instead of FedFA + RandAugment | standard CNN · CIFAR-100 (α = 5.0) | test accuracy: 48.86 -> 50.47 | Helped |
| Use FedAvP instead of FedFA + TrivialAugment | standard CNN · CIFAR-100 (α = 0.1) | test accuracy: 43.45 -> 45.96 | Helped |
| Use FedAvP instead of FedMix + RandAugment | standard CNN · CIFAR-10 (α = 5.0) | test accuracy: 83.08 -> 83.78 | Helped |
| Use FedAvP instead of FedFA + Default | standard CNN · SVHN (α = 0.1) | test accuracy: 87.33 -> 89.81 | Helped |
| Use FedAvP instead of FedGen + Default | standard CNN · FEMNIST | test accuracy: 81.86 -> 84.27 | Helped |
Surprising findings
- Applying RandAugment or TrivialAugment to federated baselines does not consistently improve accuracy and sometimes hurts it (e.g., SCAFFOLD + TrivialAugment drops to 14.70% on SVHN with alpha=0.1).
- FedAvP (Fast Update) with 2-layer small policy network sometimes reaches higher accuracy than full FedAvP in some settings, and FedAvP's full version costs substantially more wall-clock time than FedAvg.
Limitations
- Joint training requires backpropagation through policy gradients at every local step, adding client compute overhead.
- Assumes FedAvg as the underlying algorithm and does not account for model personalization in the FMPL computation.
- Experiments use small CNNs and a small ViT; reconstruction defense is not designed as a privacy defense and policy gradients are not shown to reduce leakage.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 SVHN FEMNIST
Benchmarks: CIFAR-10 CIFAR-100 SVHN FEMNIST
Metrics: test accuracy PSNR (reconstruction attack) computation time to target accuracy communication cost (MB)
Baselines: FedAvg FedProx FedDyn FedExP FedGen FedMix FedFA RandAugment TrivialAugment ATSPrivacy FedNova SCAFFOLD
Builds on: FedAvg AutoAugment Reptile RandAugment TrivialAugment
Closest papers by meaning
- A Bayesian Approach for Personalized Federated Learning in Heterogeneous Settings (NeurIPS 2024 · similarity 0.88)
- A Simple Data Augmentation for Feature Distribution Skewed Federated Learning (CVPR 2025 · similarity 0.87)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.87)
- FedP3: Federated Personalized and Privacy-friendly Network Pruning under Model Heterogeneity (ICLR 2024 · similarity 0.87)
- Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning (AAAI 2025 · similarity 0.86)
- Low Precision Local Training is Enough for Federated Learning (NeurIPS 2024 · similarity 0.86)
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.