DWFed: A statistical- heterogeneity-based dynamic weighted model aggregation algorithm for federated learning.

Area of Science:

• Machine Learning
• Distributed Systems
• Artificial Intelligence

Background:

• Federated Learning (FL) trains global models on decentralized data, achieving high performance.
• Statistical heterogeneity from non-independent and identically distributed (non-IID) data significantly degrades FL performance due to model divergence.
• This heterogeneity is a critical challenge limiting FL's widespread application.

Purpose of the Study:

• To propose a novel algorithm, DWFed, to mitigate performance degradation in Federated Learning caused by statistical heterogeneity.
• To quantitatively define an index for statistical heterogeneity and utilize it for adaptive model aggregation.

Main Methods:

• Developed a dynamic weighted model aggregation algorithm (DWFed) for Federated Learning.
• Quantitatively defined an index to measure statistical heterogeneity.
• Applied the heterogeneity index to dynamically weight local models during aggregation to constrain divergence.

Main Results:

• DWFed demonstrated improved performance of federated models in heterogeneous settings.
• Experiments showed enhanced robustness of federated models when using DWFed on public benchmark datasets.
• The proposed method effectively constrains model divergence caused by non-IID data.

Conclusions:

• DWFed offers an effective solution to the critical challenge of statistical heterogeneity in Federated Learning.
• The algorithm improves both the performance and robustness of FL models in real-world, non-IID data scenarios.
• This work advances the practical applicability of Federated Learning by addressing data distribution challenges.