Research map / Federated learning and privacy / Personalized and heterogeneous FL
PA3Fed: Period-Aware Adaptive Aggregation for Improved 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 PA3Fed, a plug-in federated aggregation module that reweights client models using intra-client score changes during Critical Learning Periods identified via federated gradient norm.
Problem
Statistical heterogeneity in federated learning causes unstable per-client loss fluctuations and treats all training periods as equally important, while most methods focus only on inter-client optimization.
Approach
Computes a per-client aggregation coefficient from the change in local performance score between rounds, and applies this adaptive reweighting only during detected Critical Learning Periods (detected by slope of Federated Gradient Norm); it is orthogonal to inter-client methods and can be plugged into them.
Key results
- Up to ~15% accuracy improvement on CIFAR-10 under high non-IID (h=0.1) when PA3Fed is added to FedAvg-family methods.
- Loss standard deviation reduced versus FedAvg (e.g., standard deviation 0.089 to 0.066 in the stability table).
- Optimal beta found at 0.3 in ablation over {0,...,0.5}.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with AlexNet · CIFAR-10, h=0.1 | test accuracy: 38.66 -> 53.99 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with AlexNet · CIFAR-10, h=0.2 | test accuracy: 44.25 -> 58.85 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with AlexNet · CIFAR-10, h=0.3 | test accuracy: 48.14 -> 58.94 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with ResNet-18 · CIFAR-100, h=0.1 | test accuracy: 28.33 -> 36.97 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with ResNet-18 · CIFAR-100, h=0.2 | test accuracy: 31.84 -> 41.01 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with ResNet-18 · CIFAR-100, h=0.3 | test accuracy: 35.99 -> 42.23 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with AlexNet · FMNIST, h=0.1 | test accuracy: 74.40 -> 78.88 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with AlexNet · FMNIST, h=0.2 | test accuracy: 76.94 -> 81.60 | Helped |
| Add PA3Fed adaptive aggregation to FedAvg | FedAvg with AlexNet · FMNIST, h=0.3 | test accuracy: 77.30 -> 81.86 | Helped |
| Adapt aggregation throughout the entire training period instead of only during CLP (ablation) | PA3Fed + FedAvg · CIFAR-10 and FMNIST | test accuracy | Hurt |
| Set adaptation degree β to 0.3 rather than other tested values (ablation) | PA3Fed · CIFAR-10 and FMNIST | test accuracy | Helped |
| Vary the non-IID degree across h values (ablation) | PA3Fed · CIFAR-10 and FMNIST | test accuracy | Mixed |
Limitations
- Evaluated only on image classification with small CNN/ResNet models and 50 simulated clients.
- Relies on hyperparameters (beta, delta, CLP threshold) that are tuned on the experiments.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Fashion-MNIST
Benchmarks: CIFAR-10 CIFAR-100 Fashion-MNIST
Metrics: final test accuracy loss standard deviation convergence speed
Baselines: FedAvg FedProx FedNova VRL-SGD FedMut
Builds on: FedAvg CriticalFL FedProx SCAFFOLD FedMut
Closest papers by meaning
- FedAWA: Adaptive Optimization of Aggregation Weights in Federated Learning Using Client Vectors (CVPR 2025 · similarity 0.89)
- An Aggregation-Free Federated Learning for Tackling Data Heterogeneity (CVPR 2024 · similarity 0.88)
- FedAA: A Reinforcement Learning Perspective on Adaptive Aggregation for Fair and Robust Federated Learning (AAAI 2025 · similarity 0.87)
- pFedMxF: Personalized Federated Class-Incremental Learning with Mixture of Frequency Aggregation (CVPR 2025 · similarity 0.86)
- Bridging Model Heterogeneity in Federated Learning via Uncertainty-based Asymmetrical Reciprocity Learning (ICML 2024 · similarity 0.86)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · 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.