Research map / Federated learning and privacy / Personalized and heterogeneous FL
Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning
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 MPFT, a one-round federated domain adaptation method that uploads CLIP-embedding prototypes from clients, trains a global adapter on the server, and then locally adapts with knowledge distillation.
Problem
Federated domain adaptation, where clients have distinct data domains but share a label space; averaging-based aggregation diverges under domain heterogeneity and hurts both in-domain and out-of-domain accuracy.
Approach
Replace weight averaging with server-side training of a global adapter on a pooled set of per-domain, per-class prototypes (mean, cluster-center, or random-sampled frozen CLIP embeddings), so the global model fits the union of client distributions in a single communication round, with Gaussian-noise differential privacy on prototypes.
Key results
- On DomainNet subset-50, MPFT (Random, rate=0.3) reaches 0.8236 ood acc and 0.8294 ind acc in one communication round, versus 0.7902 ood / 0.7345 ind for FedAvg after 24 rounds.
- On PACS, MPFT (Random, rate=0.3) reports 0.9887 ood and 0.9919 ind accuracy in one round, while FedAvg reaches 0.9725 ood / 0.9887 ind over 32 rounds.
- MPFT (Average) training time is about 1.9s on DomainNet subset-50 versus 1933.5s for FedAvg, though average sampling gives lower ind/ood accuracy than cluster or random sampling.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace FedAvg with MPFT Random sampling at rate=0.3 | Federated domain adaptation · DomainNet Subset-50 | ood acc: 0.7902 -> 0.8236 | Helped |
| Use MPFT Random sampling at rate=0.3 instead of local training | Federated domain adaptation · DomainNet Subset-50 | ind acc: 0.8609 -> 0.8294 | Hurt |
| Increase sampling rate from 0.1 to 0.3 with Random sampling (ablation) | MPFT · DomainNet Subset-50 | ood acc: 0.7953 -> 0.8236 | Helped |
| Increase sampling rate from 0.1 to 0.3 with Cluster sampling (ablation) | MPFT · DomainNet Subset-50 | ind acc: 0.7957 -> 0.8294 | Helped |
| Use Cluster sampling instead of Average sampling (ablation) | MPFT · DomainNet Subset-50 | ood acc: 0.8077 -> 0.8204 | Helped |
| Add prototype noise with q=0.1, s=0.5 (ablation) | MPFT Average · DomainNet Subset-50 | ood acc: 0.8077 -> 0.8083 | Helped |
| Add prototype noise with q=0.1, s=5 (ablation) | MPFT Average · DomainNet Subset-50 | ood acc: 0.8077 -> 0.7898 | Hurt |
| Increase KD weight during few-shot local adaptation (ablation) | MPFT local adaptation · few-shot local adaptation | ood and ind accuracy | Mixed |
| Decrease the global convergence threshold (ablation) | MPFT global adapter initialization · global adapter initialization | ood and ind accuracy | Mixed |
Surprising findings
- Mean (average) prototype sampling underperforms cluster and random sampling, with the mean prototype lying far from individual domain-class prototypes.
- Certain Gaussian noise settings slightly improve accuracy over the noise-free model, which the authors attribute to reduced cross-dataset bias.
Limitations
- Prototype quality depends heavily on the pretrained encoder's feature extraction.
- Prototypes may still leak statistical information enabling membership or attribute inference attacks.
- Evaluation limited to image classification on DomainNet and PACS with CLIP encoders.
Experimental setup
Datasets: DomainNet PACS
Benchmarks: DomainNet subset-50 DomainNet subset-100 DomainNet subset-150 PACS
Metrics: in-domain accuracy out-of-domain accuracy communication rounds training time communication cost
Baselines: local training FedAvg FedProx Ditto MOON FedProto DBE
Builds on: FedAvg DBE FedProto CLIP
Closest papers by meaning
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.90)
- UFDA: Universal Federated Domain Adaptation with Practical Assumptions (AAAI 2024 · similarity 0.89)
- Gains: Fine-grained Federated Domain Adaptation in Open Set (NeurIPS 2025 · similarity 0.87)
- TTA-FedDG: Leveraging Test-Time Adaptation to Address Federated Domain Generalization (AAAI 2025 · similarity 0.86)
- Dual-Personalizing Adapter for Federated Foundation Models (NeurIPS 2024 · similarity 0.86)
- Adaptive Latent-Space Constraints in Personalized Federated Learning (NeurIPS 2025 · 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.