One-shot Federated Learning without server-side training.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Privacy

Background:

• Federated Learning (FL) is a privacy-preserving machine learning approach.
• High communication costs in traditional FL limit its efficiency.
• One-shot FL methods reduce communication but often rely on knowledge distillation, requiring extra training and data.

Purpose of the Study:

• To propose a novel one-shot federated learning algorithm for cross-silo settings.
• To enable parameter aggregation in a single round without server-side training.
• To address challenges posed by extremely non-identical data distributions across clients.

Main Methods:

• Introduced Model Aggregation via Exploring Common Harmonized Optima (MA-Echo).
• MA-Echo iteratively updates local model parameters towards a common low-loss area.
• Ensures local model performance is maintained on their respective datasets.

Main Results:

• MA-Echo outperforms existing one-shot FL methods.
• Demonstrated effectiveness even with non-overlapping label sets across clients.
• Achieved superior performance on image classification tasks.

Conclusions:

• MA-Echo offers an effective solution for one-shot federated learning in challenging cross-silo settings.
• The algorithm successfully handles non-identical data distributions without performance degradation.
• MA-Echo represents a significant advancement in efficient and private machine learning.