Research map / Label-efficient and robust learning / Test-time adaptation
TTA-FedDG: Leveraging Test-Time Adaptation to Address 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 TTA-FedDG, a framework that adds test-time adaptation to federated domain generalization, and FedSPL, which combines feature-reordering mixing during local training with a teacher-student model using confidence-separation-based strong/weak pseudo-label selection and contrastive learning at test time.
Problem
Federated domain generalization (FedDG) models trained only on source client data fail to generalize well when unknown target client domains differ significantly from source domains, and prior FedDG methods ignore the unlabeled test data available at inference.
Approach
Treat unlabeled test data from the unknown client as training signal at test time: combine feature mixing by feature reordering (histogram-style matching preserving higher-order statistics) during local training with a teacher-student model that uses a confidence separation score (top-1 minus top-2 probability) to split pseudo-labels into strong and weak groups, plus a PCA-based contrastive loss.
Key results
- On PACS with ResNet18, FedSPL reaches 88.77% average accuracy versus 82.26% for FedAvg and 66.80%-type TTA baseline ATP at 75.21% on OfficeHome (67.03% avg vs ATP 66.80%).
- On Digit-5, FedSPL achieves 87.76% average accuracy, versus 82.65% for ATP and 82.37% for GA.
- Ablation on PACS shows FMR alone gives 82.26% and MBSL alone gives 87.50% average accuracy, versus 80.22% for FedAvg.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FedSPL instead of Tent | PACS · PACS Avg | accuracy: 86.83 -> 88.77 | Helped |
| Use FedSPL instead of ATP | OfficeHome · OfficeHome Avg | accuracy: 66.80 -> 67.03 | Helped |
| Use FedSPL instead of ATP | Digit-5 · Digit-5 Avg | accuracy: 82.65 -> 87.76 | Helped |
| Add FMR to FedAvg training (ablation) | FedSPL · PACS Avg | accuracy: 80.22 -> 82.26 | Helped |
| Add MBSL test-time adaptation to FedAvg (ablation) | FedSPL · PACS Avg | accuracy: 80.22 -> 87.50 | Helped |
| Add FMR to MBSL (ablation) | FedSPL · PACS Avg | accuracy: 87.50 -> 88.77 | Helped |
| Use Ls rather than no loss components (ablation) | PACS loss ablation · PACS Avg | accuracy: 82.26 -> 88.03 | Helped |
| Use Lw rather than no loss components (ablation) | PACS loss ablation · PACS Avg | accuracy: 82.26 -> 85.95 | Helped |
| Use Lc rather than no loss components (ablation) | PACS loss ablation · PACS Avg | accuracy: 82.26 -> 85.57 | Helped |
| Add Lw to Ls, without Lc (ablation) | PACS loss ablation · PACS Avg | accuracy: 88.03 -> 88.13 | Helped |
| Add Lc to Ls, without Lw (ablation) | PACS loss ablation · PACS Avg | accuracy: 88.03 -> 87.97 | Hurt |
| Add Lc to Lw, without Ls (ablation) | PACS loss ablation · PACS Avg | accuracy: 85.95 -> 86.33 | Helped |
Surprising findings
- TTA does not always help: when the initial model is poor or the gap between target and source domains is large, test-time adaptation may yield no gain or degrade performance.
- The contrastive loss Lc can reduce accuracy on domains far from the training domains (e.g., Sketch), though it improves average accuracy.
Limitations
- Performance can degrade when the target domain is far from source domains or the initial model is weak.
- Test-time training introduces instability, which the method only partially mitigates with auxiliary losses and augmentation.
- Evaluation limited to three image benchmarks with ResNet backbones and leave-one-domain-out splits.
Experimental setup
Datasets: PACS OfficeHome Digit-5
Benchmarks: PACS OfficeHome Digit-5
Metrics: accuracy average accuracy across domains
Baselines: FedAvg FedProx FedCSA FedNova AM RSC FedSAM HarmoFL Scaffold ELCFS GA DSBN
Builds on: FedAvg Tent ATP MixStyle FedDG-GA
Closest papers by meaning
- Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.90)
- Benchmarking Algorithms for Federated Domain Generalization (ICLR 2024 · similarity 0.90)
- Efficiently Assemble Normalization Layers and Regularization for Federated Domain Generalization (CVPR 2024 · similarity 0.88)
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.88)
- Federated Unsupervised Domain Generalization Using Global and Local Alignment of Gradients (AAAI 2025 · similarity 0.87)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (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.