Reproduce “Fair Federated Learning under Domain Skew with Local Consistency and Domain Diversity” on OpenResearch

About this paper

Yuhang Chen, Wenke Huang, Mang Ye

This work introduces Federated Parameter-Harmonized and Aggregation-EquAlized Learning (FedHEAL), a framework designed to enhance performance fairness in federated learning under domain skew. FedHEAL consistently reduced the standard deviation of accuracy across domains by up to 5.60% and improved average accuracy by up to 15.12% when integrated with existing FL methods, outperforming prior fairness-focused approaches.

Key insights

• The discovery of 'Parameter Update Consistency' (PUC) reveals that certain model parameters consistently update in the same direction for a given client, signaling their importance for adapting to that client's domain.
• Performance fairness under domain skew can be improved by harmonizing parameter updates at a granular level, preventing dilution of critical domain-specific knowledge.
• Aggregating client models with weights that account for domain diversity and 'fitting gaps' between the global and local models leads to a more balanced and equitable global model convergence.

Summary by alphaXiv.