Research map / Federated learning and privacy / Personalized and heterogeneous FL
Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization
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 gPerXAN, a federated domain generalization method that explicitly mixes instance and batch normalization (global IN, local BN) with a classifier-guided regularizer, achieving strong results on PACS, Office-Home and Camelyon17 without sharing client data.
Problem
Federated domain generalization: training a global model across clients each holding a single source domain so it generalizes to unseen clients/domains, without data leakage or high communication/computation cost.
Approach
Explicitly assemble IN and BN outputs via learnable mixing weights (XAN), personalize the BN sides locally while aggregating IN sides globally (PerXAN), and add a regularizer that passes client features through the frozen global classifier so features are directly usable for classification.
Key results
- gPerXAN reaches 87.94% average accuracy on PACS and 71.01% on Office-Home over unseen clients, 1.02 and 1.15 points above the second-best method.
- On Camelyon17, gPerXAN averages 94.1%, about 2 points above FedDG-GA (92.3%).
- Adding the regularizer improves FedAvg and PerXAN but not ELCFS or CCST, which the authors attribute to redundancy with information-sharing.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use gPerXAN instead of FedDG-GA | FedDG · PACS | average accuracy: 86.92 -> 87.94 | Helped |
| use gPerXAN instead of FedDG-GA | FedDG · Office-Home | average accuracy: 69.86 -> 71.01 | Helped |
| use gPerXAN instead of FedDG-GA | FedDG · Camelyon17 | average accuracy: 92.3 -> 94.1 | Helped |
| use PerXAN instead of BN (ablation) | gFedAvg on PACS · PACS | average accuracy: 83.68 -> 87.94 | Helped |
| add the guiding regularizer (ablation) | FedAvg and PerXAN on PACS · PACS | average accuracy | Helped |
| add the guiding regularizer (ablation) | ELCFS and CCST on PACS · PACS | average accuracy | Hurt |
Surprising findings
- The regularizer does not help information-sharing methods (ELCFS, CCST), suggesting redundancy with shared global knowledge.
- Centralized DG methods MixStyle and RSC are inconsistent and sometimes harmful when placed in FedAvg.
Limitations
- Suitability may be limited to normalization-based architectures.
- Evaluation focused on cross-silo FL; cross-device scenarios not tested.
- Robustness to attacks remains unexplored.
Experimental setup
Datasets: PACS Office-Home Camelyon17
Benchmarks: PACS Office-Home Camelyon17
Metrics: test accuracy
Baselines: FedAvg FedAvg w/ MixStyle FedAvg w/ RSC ELCFS CCST COPA FedDG-GA
Builds on: FedAvg IBN-Net (I-BN) DSON MixStyle FedDG-GA
Closest papers by meaning
- Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.89)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.88)
- TTA-FedDG: Leveraging Test-Time Adaptation to Address Federated Domain Generalization (AAAI 2025 · similarity 0.88)
- Federated Unsupervised Domain Generalization Using Global and Local Alignment of Gradients (AAAI 2025 · similarity 0.87)
- Improving Federated Domain Generalization Through Dynamical Weights Calculated from Data Influences on Global Model Update (AAAI 2025 · similarity 0.87)
- Benchmarking Algorithms for Federated Domain Generalization (ICLR 2024 · similarity 0.87)
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.