HCFL: hybrid contribution-driven federated learning for fair and efficient optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Federated Learning (FL) facilitates privacy-preserving collaborative training across decentralized data.
• Existing FL client selection relies on limited metrics, hindering performance with sparse, heterogeneous data like in healthcare.
• This leads to reduced global model generalization and inefficient training.

Purpose of the Study:

• To propose an adaptive FL framework addressing limitations in conventional client selection.
• To enhance global model generalization and training efficiency in data-scarce, heterogeneous environments.
• To ensure fairness and prevent participant exclusion in FL.

Main Methods:

• Developed a hybrid contribution evaluation mechanism for client selection and resource management.
• Implemented a performance-based evaluation measuring the impact of client updates on global optimization.
• Introduced a coverage-based evaluation estimating data diversity in latent space without raw data exposure.

Main Results:

• The proposed adaptive FL framework demonstrated superior performance over existing baselines.
• Achieved significant improvements in training efficiency and data utilization.
• Showcased enhanced fairness in client selection and resource allocation.

Conclusions:

• The hybrid evaluation mechanism effectively balances client contribution and data diversity.
• The adaptive FL framework offers a robust solution for privacy-preserving collaborative learning in challenging data settings.
• This approach enhances both the efficiency and fairness of Federated Learning models.