AFed: Algorithmic Fair Federated Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training without data centralization, enhancing user privacy.
• FL introduces unique fairness challenges because traditional debiasing methods require centralized data access, which is impractical in FL.
• Diverse client data in FL can exacerbate fairness issues related to sensitive group attributes.

Purpose of the Study:

• To develop a framework for promoting group fairness in federated learning settings.
• To address the challenge of training fair models in FL without direct access to local client data.
• To propose methods that circumvent restricted data access by learning the global data distribution.

Main Methods:

• Introduction of the AFed framework for group fairness in FL.
• AFed-G: A server-side conditional generator learns the global data distribution.
• AFed-GAN: Client-side conditional GAN improves upon AFed-G for bias mitigation.
• Augmentation of client data with generated samples to remove bias.

Main Results:

• Theoretical analysis supports the validity of the proposed AFed methods.
• Empirical results on real-world datasets show significant fairness improvements with AFed compared to baseline methods.
• The proposed approaches effectively mitigate bias in federated learning models.

Conclusions:

• The AFed framework offers a practical solution for achieving group fairness in federated learning.
• Learning the global data distribution is a viable strategy to overcome data privacy constraints in debiasing.
• AFed provides substantial fairness gains, demonstrating its effectiveness in diverse, real-world FL scenarios.