Using Class and Domain Information to Address Domain Shift in Federated Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) is challenged by heterogeneous client data distributions, leading to domain shift.
• This domain shift causes divergent local models and degrades overall global performance in FL.
• Existing methods struggle to effectively handle the complexities of domain variation in FL.

Purpose of the Study:

• To propose a novel federated learning framework addressing the domain-shift problem.
• To decouple and collaboratively learn domain-invariant and domain-specific representations.
• To enhance the robustness and performance of the global model against domain variations.

Main Methods:

• Introduced a class- and domain-aware FL framework.
• Employed a cross-gated feature separation (CGFS) module to decouple domain and class features.
• Utilized heterogeneous prototype contrastive learning (HPCL) for feature discriminability and a gradient-reweighted hierarchical aggregation (GHA) strategy for server aggregation.

Main Results:

• The proposed framework successfully decouples and learns domain-invariant and domain-specific representations.
• Experimental results on two FL datasets with domain shift demonstrate superior performance.
• The method consistently outperforms existing state-of-the-art approaches in handling domain variation.

Conclusions:

• The proposed class- and domain-aware FL framework effectively mitigates the impact of domain shift.
• The novel CGFS, HPCL, and GHA modules contribute to improved global model robustness and performance.
• This approach offers a promising solution for building more resilient federated learning systems.