Research map / Label-efficient and robust learning / Domain adaptation and generalization
UFDA: Universal Federated Domain Adaptation with Practical Assumptions
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 Universal Federated Domain Adaptation (UFDA), a black-box, label-set-relaxed FDA setting, and HCLD^2 (one-hot pseudo-labels with GMM-based contrastive disambiguation plus a Mutual-Voting Decision), evaluated on Office-Home, Office-31 and VisDA+ImageCLEF-DA.
Problem
Federated domain adaptation usually requires white-box model parameters or gradients and aligned label sets across source and target domains, which is impractical and risky; UFDA relaxes these to black-box source APIs with possibly inconsistent source label sets and a fully unknown target label set.
Approach
Use one-hot outputs from multiple black-box source APIs to form multi-candidate pseudo-labels (PHL), disambiguate them via GMM-based split of contrastive self-entropy (GCLD), and separate shared from unknown classes with a cluster-level mutual-voting score (MVD) combining source and target views.
Key results
- HCLD^2 reaches 75.41% average accuracy on VisDA+ImageCLEF-DA, compared with 80.58% for UMAN (the strongest prior UniMDA baseline in the table).
- HCLD^2 averages 77.18% on Office-Home versus 77.97% for UMAN, and 89.28% on Office-31 versus 93.08% for UMAN.
- Ablation on Office-31 shows the full MVD raises average accuracy from 83.17% (without MVD) to 89.28%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HCLD² instead of UMAN | Office-Home · Office-Home | accuracy: 77.97 -> 77.18 | Hurt |
| Use HCLD²* instead of HCLD² | HCLD² on Office-Home · Office-Home | accuracy: 77.18 -> 78.42 | Helped |
| Use HCLD² instead of UMAN | Office-31 · Office-31 | accuracy: 93.08 -> 89.28 | Hurt |
| Use HCLD²* instead of HCLD² | HCLD² on Office-31 · Office-31 | accuracy: 89.28 -> 91.02 | Helped |
| Use HCLD² instead of UMAN | VisDA+ImageCLEF-DA · VisDA+ImageCLEF-DA | accuracy: 80.58 -> 75.41 | Hurt |
| Add GCLD to PHL-only training (ablation) | PHL classifier on VisDA+ImageCLEF-DA · VisDA+ImageCLEF-DA | accuracy: 41.6 -> 66.6 | Helped |
| Add MVD to PHL + GCLD (ablation) | HCLD on VisDA+ImageCLEF-DA · VisDA+ImageCLEF-DA | accuracy: 66.6 -> 75.4 | Helped |
| Add GCLD to PSL-only training (ablation) | PSL classifier on VisDA+ImageCLEF-DA · VisDA+ImageCLEF-DA | accuracy: 53.9 -> 62.3 | Helped |
| Add MVD to PSL + GCLD (ablation) | PSL with GCLD on VisDA+ImageCLEF-DA · VisDA+ImageCLEF-DA | accuracy: 62.3 -> 68.3 | Helped |
| Use PHL rather than PSL with GCLD and MVD (ablation) | HCLD on VisDA+ImageCLEF-DA · VisDA+ImageCLEF-DA | accuracy: 68.3 -> 75.4 | Helped |
| Remove target view from MVD (ablation) | MVD on Office-31 · Office-31 | accuracy: 89.28 -> 86.16 | Hurt |
| Remove source view from MVD (ablation) | MVD on Office-31 · Office-31 | accuracy: 89.28 -> 88.44 | Hurt |
Surprising findings
- One-hot pseudo-hot labels (PHL) perform worse than soft pseudo-labels without GCLD (41.6% vs 53.9% avg) but significantly outperform them once GCLD is applied (75.4% vs 68.3%).
Limitations
- Still assumes source label sets are known and the decentralized clients communicate black-box models during training.
- Performance is below the UMAN baseline on several tasks, e.g., VisDA+ImageCLEF-DA average 75.41% vs 80.58%.
- HCLD^2 with soft pseudo-labels is sensitive to the label-set configuration (e.g., drops when multi-source label-set intersection is empty).
Experimental setup
Datasets: Office-Home Office-31 VisDA2017 ImageCLEF-DA
Benchmarks: Office-Home Office-31 VisDA+ImageCLEF-DA
Metrics: accuracy
Baselines: DANN RTN OSBP UAN DCTN MDAN MDDA UMAN
Builds on: MoCo UniMDA UMAN DCTN
Closest papers by meaning
- Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning (ICLR 2025 · similarity 0.89)
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.87)
- Gains: Fine-grained Federated Domain Adaptation in Open Set (NeurIPS 2025 · similarity 0.86)
- TTA-FedDG: Leveraging Test-Time Adaptation to Address Federated Domain Generalization (AAAI 2025 · similarity 0.86)
- Revisiting Source-Free Domain Adaptation: Insights into Representativeness, Generalization, and Variety (CVPR 2025 · similarity 0.86)
- Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias (CVPR 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.