Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedAS: Bridging Inconsistency in Personalized 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 FedAS, a personalized federated learning framework that aligns the shared backbone with each client's previous local backbone (client-side Parameter Alignment) and weights server aggregation by trace-of-Fisher-Information (server-side Client Synchronization) to handle intra- and inter-client inconsistency.
Problem
Personalized federated learning suffers from intra-client inconsistency between asynchronously updated personalized and shared parameters, and inter-client inconsistency caused by under-trained straggler clients that corrupt aggregation under non-iid data.
Approach
Align the received global backbone's feature outputs with the client's own previous-round backbone features before local training, and weight server aggregation by each client's trace of the Fisher Information Matrix so that under-trained stragglers contribute less.
Key results
- On CIFAR-100 (beta=0.1, P=1), FedAS reaches 56.37% vs 52.03% for the strongest baseline FedPAC, a gain of 4.34 points.
- On CIFAR-10 (beta=0.1, P=1), FedAS reaches 91.03% vs 90.06% for FedALA.
- Ablation on CIFAR-10 (P=0.6): Parameter Alignment alone and Client Synchronization alone each improve over baseline FedPer, and combining both gives 90.81%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedAS rather than the strongest reported baseline | PFL with a 6-layer CNN · CIFAR100, β=0.1 and β=0.3; P=0.2, 0.4, 0.6, 1 | accuracy: ↑6.89; ↑5.41; ↑4.89; ↑4.34; ↑4.14; ↑3.10; ↑3.27; ↑2.83 | Helped |
| Use FedAS rather than the strongest reported baseline | PFL with a 6-layer CNN · CIFAR10, β=0.1 and β=0.3; P=0.2, 0.4, 0.6, 1 | accuracy: ↑0.31; ↑0.71; ↑0.81; ↑0.97; ↑1.20; ↑0.66; ↑0.46; ↑0.08 | Helped |
| Add Parameter Alignment and Client Synchronization (ablation) | FedAS · CIFAR10 (P=0.2, 0.4, 0.6, 1) | accuracy: ↑0.58; ↑1.10; ↑0.91; ↑1.09 | Helped |
| Add Client Synchronization to Parameter Alignment (ablation) | FedAS · CIFAR10 (P=0.2, 0.4, 0.6, 1) | accuracy: 89.34 -> 89.92; 89.50 -> 90.60; 89.72 -> 90.81; 89.91 -> 91.03 | Helped |
| Add Parameter Alignment to Client Synchronization (ablation) | FedAS · CIFAR10 (P=0.2, 0.4, 0.6, 1) | accuracy: 89.05 -> 89.92; 89.39 -> 90.60; 89.90 -> 90.81; 89.94 -> 91.03 | Helped |
Surprising findings
- Substituting the top-5% FIM-valued parameters causes catastrophic accuracy drops, while random 5% substitution has minor effects, supporting FIM as an information-content measure.
Limitations
- Assumes all clients share the same model architecture, limiting heterogeneous-model settings.
- Evaluated only on small image classification benchmarks with a 6-layer CNN.
Experimental setup
Datasets: CIFAR-10 CIFAR-100
Benchmarks: CIFAR-10 CIFAR-100
Metrics: top-1 accuracy
Baselines: FedAvg-P FedProx-P FedPer Ditto LG-FedAvg FedBABU FedProto FedALA FedPAC
Builds on: FedAvg FedPer FedBABU FedALA FedPAC
Closest papers by meaning
- Understanding the Statistical Accuracy-Communication Trade-off in Personalized Federated Learning with Minimax Guarantees (ICML 2025 · similarity 0.91)
- PeFLL: Personalized Federated Learning by Learning to Learn (ICLR 2024 · similarity 0.91)
- Flexible Sharpness-Aware Personalized Federated Learning (AAAI 2025 · similarity 0.90)
- Look Back for More: Harnessing Historical Sequential Updates for Personalized Federated Adapter Tuning (AAAI 2025 · similarity 0.90)
- PFedCS: A Personalized Federated Learning Method for Enhancing Collaboration among Similar Classifiers (AAAI 2025 · similarity 0.90)
- Adaptive Latent-Space Constraints in Personalized Federated Learning (NeurIPS 2025 · similarity 0.89)
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.