FedDAP: Domain-Aware Prototype Learning for Federated Learning under Domain Shift
In the realm of artificial intelligence and machine learning, Federated Learning (FL) has emerged as a groundbreaking approach that enables decentralized model training across multiple clients while ensuring the privacy of sensitive data. However, a significant challenge arises in real-world FL scenarios where clients often possess data from distinct domains. This leads to severe domain shifts that can degrade the performance of the global model. A recent paper titled “FedDAP: Domain-Aware Prototype Learning for Federated Learning under Domain Shift” seeks to address this critical issue.
Abstract Overview
The authors of the study highlight two key limitations prevalent in existing prototype-based Federated Learning methods:
- Single Global Prototype Limitation: Current methods typically construct a single global prototype per class by aggregating local prototypes from all clients without effectively preserving domain information.
- Domain-Agnostic Feature-Prototype Alignment: Existing approaches enforce a domain-agnostic alignment, compelling clients to align their local features with global prototypes, irrespective of the domain origin.
Introducing FedDAP
To tackle the aforementioned challenges, the authors propose a novel framework named Federated Domain-Aware Prototypes (FedDAP). This innovative approach aims to construct domain-specific global prototypes by aggregating local client prototypes that belong to the same domain. This is achieved through a similarity-weighted fusion mechanism, which ensures that the domain information is retained during the aggregation process.
FedDAP’s framework operates on two principal mechanisms:
- Domain-Specific Prototype Construction: By focusing on local client prototypes within the same domain, FedDAP creates global prototypes that are more representative of the specific characteristics of each domain.
- Dual Alignment Strategy: The model guides local training by aligning local features with the corresponding prototypes from the same domain, while simultaneously encouraging separation from prototypes associated with different domains. This enhances domain-specific learning at the local level.
Experimental Validation
The effectiveness of the FedDAP framework is substantiated through extensive experiments conducted on three diverse datasets: DomainNet, Office-10, and PACS. The results indicate that FedDAP significantly improves the global model’s capability to generalize across varying domains, thereby effectively addressing the challenges posed by domain shifts in Federated Learning scenarios.
Conclusion and Accessibility
In conclusion, the FedDAP framework represents a significant advancement in the field of Federated Learning by introducing domain-aware prototype learning. By focusing on domain-specific global prototypes and enhancing local training through a dual alignment mechanism, FedDAP enhances model performance in the face of domain shifts. For researchers and practitioners interested in exploring this innovative approach further, the implementation code is available at https://github.com/quanghuy6997/FedDAP.