Research map / Label-efficient and robust learning / Domain adaptation and generalization
Multi-Source Collaborative Gradient Discrepancy Minimization 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 MCGDM, a federated domain generalization method that minimizes intra-domain gradient discrepancy (original vs. RandAugment images) and inter-domain gradient discrepancy (against other clients' classifier heads), extended to federated domain adaptation via pseudo-labeling.
Problem
Federated domain generalization: learning a domain-invariant model from decentralized, isolated source domains (privacy constraints) that generalizes to an unseen target domain, and extending this to federated domain adaptation.
Approach
Instead of sharing style or data across clients, align classifier-head gradients between original and augmented images within each client (intra-domain) and between the current client's classifier and other domains' classifier heads (inter-domain), so local models update toward domain-invariant directions.
Key results
- On PACS, MCGDM reaches 86.2% average accuracy, versus 85.1% for COPA and 81.3% for FedAvg.
- On VLCS, MCGDM reaches 78.1% average accuracy, above CASC at 76.2%.
- On Digit-5 federated domain adaptation, MCGDM reaches 97.1% average accuracy, versus 95.4% for COPA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MCGDM instead of CDG | federated domain generalization · PACS | accuracy (Avg): 85.8 -> 86.2 | Helped |
| Use MCGDM instead of DAEL | federated domain generalization · Office-Home | accuracy (Avg): 66.1 -> 66.8 | Helped |
| Use MCGDM instead of CASC | federated domain generalization · VLCS | accuracy (Avg): 76.2 -> 78.1 | Helped |
| Use MCGDM instead of COPA | federated domain adaptation · Digit-5 | accuracy (Avg): 95.4 -> 97.1 | Helped |
| Use MCGDM instead of CAiDA | federated domain adaptation · Office-Caltech10 | accuracy (Avg): 98.4 -> 98.6 | Helped |
| Add RandAug to the FedAvg baseline (ablation) | PACS FedDG · PACS | accuracy (Avg): 81.3 -> 83.9 | Helped |
| Add intra-domain gradient matching to RandAug baseline (ablation) | PACS FedDG · PACS | accuracy (Avg): 83.9 -> 85.5 | Helped |
| Add inter-domain gradient matching to RandAug baseline (ablation) | PACS FedDG · PACS | accuracy (Avg): 83.9 -> 85.4 | Helped |
| Add both intra-domain and inter-domain gradient matching (ablation) | PACS FedDG with RandAug · PACS | accuracy (Avg): 83.9 -> 86.2 | Helped |
| Replace gradient matching with KL alignment (ablation) | PACS FedDG alignment strategy · PACS | accuracy (Avg): 86.2 -> 84.8 | Hurt |
| Replace gradient matching with SupCon alignment (ablation) | PACS FedDG alignment strategy · PACS | accuracy (Avg): 86.2 -> 84.8 | Hurt |
| Replace gradient matching with CoFC alignment (ablation) | PACS FedDG alignment strategy · PACS | accuracy (Avg): 86.2 -> 84.2 | Hurt |
Limitations
- Gradient matching restricted to classifier heads rather than full model or feature extractor, a deliberate trade-off for cost.
- Inter-domain matching relies on classifier heads from other domains from the previous round, which are shared with clients.
Experimental setup
Datasets: PACS VLCS Office-Home Digit-5 Office-Caltech10
Benchmarks: PACS VLCS Office-Home Digit-5 Office-Caltech10
Metrics: accuracy
Baselines: DeepAll JiGen EISNet L2A-OT DDAIG FACT MixStyle StyleNeo RSC CDG DAEL FedAvg
Builds on: FedAvg RandAugment COPA MixStyle
Closest papers by meaning
- Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.92)
- Federated Unsupervised Domain Generalization Using Global and Local Alignment of Gradients (AAAI 2025 · similarity 0.90)
- Principled Federated Domain Adaptation: Gradient Projection and Auto-Weighting (ICLR 2024 · similarity 0.88)
- FedWon: Triumphing Multi-domain Federated Learning Without Normalization (ICLR 2024 · similarity 0.87)
- DiPrompT: Disentangled Prompt Tuning for Multiple Latent Domain Generalization in Federated Learning (CVPR 2024 · similarity 0.87)
- Gradient-Guided Annealing for Domain Generalization (CVPR 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.