Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Learning with Domain Shift Eraser
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 FDSE, a federated learning framework that decomposes each layer into a shared domain-agnostic extractor and a personalized domain-specific skew eraser, combining consensus-maximized aggregation of shared parts with similarity-aware personalization of skew erasers.
Problem
Cross-domain federated learning where clients hold data from different domains, so domain shifts cause conflicting updates and degrade the aggregated model's performance and representation consistency.
Approach
Views the forward pass as iterative deskewing: each layer is split into a DFE (shared) and a DSE (personalized) sub-convolution; a statistics-matching regularizer pulls each DSE's output statistics toward global statistics, DFE and head are aggregated by min-norm layer-wise consensus, and DSE modules are aggregated via self-attention over client similarity.
Key results
- FDSE reaches 83.81% ALL accuracy on PACS, versus 82.02% for Ditto and 81.58% for FedBN.
- On Office-Caltech10, FDSE achieves 87.15% ALL accuracy and 91.58% AVG accuracy, the best among compared methods.
- FDSE halves the per-round communication (24.87M vs 49.52M parameters) and reduces FLOPs relative to FedBN and Ditto.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FDSE instead of PartialFed | AlexNet federated learning · DomainNet, ALL | accuracy (%): 75.92±0.24 -> 76.77±0.41 | Helped |
| Use FDSE instead of Ditto | AlexNet federated learning · Office-Caltech10, ALL | accuracy (%): 84.12±1.32 -> 87.15±2.06 | Helped |
| Use FDSE instead of Ditto | AlexNet federated learning · PACS, ALL | accuracy (%): 82.02±1.32 -> 83.81±1.70 | Helped |
| Add consensus-maximization aggregation (ablation) | FDSE · DomainNet, Office-Caltech10, and PACS, ALL | accuracy (%): 75.66 -> 76.77; 84.44 -> 87.15; 82.92 -> 83.81 | Helped |
| Add similarity-aware aggregation (ablation) | FDSE · DomainNet, Office-Caltech10, and PACS, ALL | accuracy (%): 76.57 -> 76.77; 85.23 -> 87.15; 83.40 -> 83.81 | Helped |
| Add consistency regularization (ablation) | FDSE · DomainNet, Office-Caltech10, and PACS, ALL | accuracy (%): 76.49 -> 76.77; 86.43 -> 87.15; 83.40 -> 83.81 | Helped |
Surprising findings
- Personalized methods like FedBN and Ditto outperform most non-personalized methods, and FedHeal/FedFA beat FedAvg in nearly all settings.
Limitations
- Evaluated only on image classification with AlexNet and three domain datasets.
- Consistency regularization and personalized aggregation add hyperparameters (lambda, beta, tau) that must be tuned; performance is sensitive to tau.
Experimental setup
Datasets: Office-Caltech10 PACS DomainNet
Benchmarks: Office-Caltech10 PACS DomainNet
Metrics: testing accuracy (ALL) testing accuracy (AVG) communication cost FLOPs training time
Baselines: Local FedAvg LG-FedAvg FedProx Scaffold FedDyn MOON Ditto PartialFed FedBN FedFA FedHeal
Builds on: FedAvg FedBN FedFA FedHeal multi-gradient descent algorithm
Closest papers by meaning
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.89)
- A Simple Data Augmentation for Feature Distribution Skewed Federated Learning (CVPR 2025 · similarity 0.88)
- FLUX: Efficient Descriptor-Driven Clustered Federated Learning under Arbitrary Distribution Shifts (NeurIPS 2025 · similarity 0.87)
- Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity (CVPR 2024 · similarity 0.87)
- Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.86)
- Federated Graph Learning under Domain Shift with Generalizable Prototypes (AAAI 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.