Federated Learning Under Evolving Distribution Shifts.

Area of Science:

• Machine Learning
• Distributed Systems
• Artificial Intelligence

Background:

• Federated learning (FL) enables collaborative model training without centralizing raw data.
• Existing FL methods often assume static client data distributions, which is unrealistic.
• Real-world FL scenarios involve dynamic, non-trivial changes in client data over time, even between training and testing.

Purpose of the Study:

• To develop FL algorithms capable of training models on time-evolving client data.
• To enhance FL system robustness against evolving data distribution shifts.
• To achieve generalization to future target data in dynamic FL environments.

Main Methods:

• Proposed FedEvolve algorithm: Explicitly models temporal evolution by learning representation transitions between consecutive client data domains.
• Proposed FedEvp algorithm: Learns an evolving-domain-invariant representation by aligning current data with continuously updated prototypes from all past domains.
• Extensive experiments conducted on both synthetic and real-world datasets.

Main Results:

• FedEvolve and FedEvp demonstrated significant performance improvements over traditional FL baselines.
• The proposed algorithms effectively captured evolving patterns in client data distributions.
• The methods showed robustness and strong generalization capabilities under evolving distribution shifts.

Conclusions:

• The proposed FedEvolve and FedEvp algorithms successfully address the challenge of dynamic client data distributions in federated learning.
• These novel approaches enable FL systems to generalize effectively to future data despite temporal shifts.
• The findings highlight the importance of accounting for data evolution in realistic FL applications.