A Fair Contribution Measurement Method for Federated Learning.

Area of Science:

• Machine Learning
• Data Privacy
• Cooperative Game Theory

Background:

• Federated learning (FL) enhances data privacy and security but struggles with active client participation.
• Existing Shapley value-based contribution evaluation methods in FL are computationally expensive and impractical due to additional model training.
• Non-Independent and Identically Distributed (Non-IID) data in FL negatively impacts global model accuracy and participant contribution measurement.

Purpose of the Study:

• To develop a fair federated learning contribution measurement scheme that avoids additional model computations.
• To accurately measure participant contributions in federated learning, particularly addressing the challenges posed by Non-IID data.
• To improve the overall accuracy and efficiency of federated learning models.

Main Methods:

• Introduced a novel aggregation weight to enhance the accuracy of contribution measurement in federated learning.
• Developed a fair federated learning contribution measurement scheme that bypasses the need for extra model computations.
• Evaluated the scheme's performance on MNIST and Fashion MNIST datasets.

Main Results:

• The proposed method accurately computes participant contributions in federated learning scenarios.
• Experimental results demonstrate a significant improvement in model accuracy compared to existing baseline algorithms.
• The new scheme achieves comparable time costs to existing methods while enhancing accuracy.

Conclusions:

• The developed fair federated learning contribution measurement scheme effectively addresses the limitations of existing methods.
• The novel aggregation weight successfully improves the accuracy of contribution measurement, even with heterogeneous (Non-IID) data.
• This approach offers a practical and efficient solution for encouraging client participation and enhancing federated learning performance.