Uncertainty-Aware Aggregation for Federated Open Set Domain Adaptation.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Open set domain adaptation (OSDA) methods typically require simultaneous access to source and target data.
• Real-world scenarios often involve distributed data across clients due to privacy constraints.
• Existing federated learning (FL) methods lack robust OSDA capabilities for handling unknown classes.

Purpose of the Study:

• To develop a novel federated OSDA (FOSDA) algorithm capable of training models on distributed data while addressing privacy concerns.
• To enable the recognition of known and unknown classes in target domains within a federated learning framework.
• To overcome the limitations of existing FL methods in handling unknown classes during domain adaptation.

Main Methods:

• Developed a federated OSDA (FOSDA) algorithm integrating an uncertainty-aware mechanism for global model aggregation.
• Implemented a strategy that prioritizes source clients with high uncertainty while maintaining consistency.
• Incorporated a federated class-based weighted strategy to preserve source client category information.

Main Results:

• Comprehensive experiments on three benchmark datasets demonstrated the effectiveness of the proposed FOSDA algorithm.
• The uncertainty-aware mechanism successfully reduced aggregation uncertainty in federated settings.
• The federated class-based weighted strategy effectively maintained crucial category information.

Conclusions:

• The proposed FOSDA algorithm offers a viable solution for privacy-preserving OSDA in distributed environments.
• FOSDA effectively addresses the challenge of unknown classes in federated learning.
• The method shows significant promise for real-world applications requiring decentralized and secure domain adaptation.