FedDAP: Domain-Aware Prototype Learning for Federated Learning under Domain Shift

Researchers introduce FedDAP, a federated learning framework that addresses domain shift challenges by constructing domain-specific global prototypes rather than single aggregated prototypes. The method aligns local features with prototypes from the same domain while encouraging separation from different domains, improving model generalization across heterogeneous client data.

FedDAP tackles a fundamental challenge in federated learning environments where participating clients hold data from distinct domains. Traditional federated learning assumes relatively uniform data distributions across clients, but real-world deployments frequently encounter domain shift—where data characteristics vary significantly between participants. This degradation in model performance directly impacts practical applications from healthcare systems using data from different hospitals to computer vision systems trained on images from diverse sources.

The innovation centers on prototype learning, a technique using class-wise feature representations to guide model training. Existing prototype-based FL methods construct single global prototypes by aggregating local prototypes indiscriminately, which erases critical domain-specific information. FedDAP instead preserves domain context by creating domain-specific global prototypes through similarity-weighted fusion of local prototypes within the same domain.

The framework implements dual alignment: local features align with prototypes from matching domains while being pushed away from prototypes of different domains. This approach enhances domain-specific learning locally while enabling better cross-domain generalization globally. Validation across DomainNet, Office-10, and PACS datasets demonstrates effectiveness in practical scenarios.

The significance extends to industries deploying federated learning infrastructure. Organizations managing distributed data collection—particularly in healthcare, finance, and computer vision—can leverage improved privacy-preserving model training with better performance under realistic data heterogeneity. The open-source availability accelerates adoption and encourages further research in domain-aware federated learning techniques.

• →FedDAP creates domain-specific global prototypes rather than single aggregated ones, preserving domain information during federated aggregation
• →Dual alignment mechanism ensures local features match same-domain prototypes while separating from different-domain prototypes
• →Framework demonstrates improved performance on three diverse datasets addressing real-world domain shift challenges
• →Approach enhances both domain-specific learning at local level and cross-domain generalization at global level
• →Open-source code availability enables industry adoption for privacy-sensitive applications with heterogeneous data distributions