Cross-Domain Federated Data Modeling on Non-IID Data.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Federated learning (FL) offers privacy-preserving distributed training but suffers from performance degradation due to data heterogeneity across non-IID (non-independent and identically distributed) data.
• Current FL aggregation methods often use simple averaging, neglecting individual party contributions and limiting overall model effectiveness.

Purpose of the Study:

• To propose a novel cross-domain federated data modeling (CDFDM) scheme to address data heterogeneity and improve model aggregation in federated learning.
• To enhance the prediction performance and stability of federated models under non-IID conditions.

Main Methods:

• Developed a shared model that dynamically adjusts shared data allocation based on user data size to mitigate heterogeneity.
• Integrated an attention mechanism into the model aggregation phase to assign weights based on user contributions, optimizing performance.
• Validated the CDFDM scheme through extensive experiments on MNIST and CIFAR-10 datasets.

Main Results:

• The proposed CDFDM scheme significantly outperforms existing methods in federated learning under non-IID data distributions.
• CDFDM demonstrates improved model prediction accuracy compared to conventional approaches.
• The federated model trained using CDFDM exhibits greater stability in prediction accuracy throughout the training process.

Conclusions:

• The CDFDM scheme effectively tackles the challenges of data heterogeneity and suboptimal aggregation in federated learning.
• The integration of a shared model and attention mechanism leads to superior and more stable performance in privacy-preserving distributed training.
• CDFDM offers a promising advancement for practical applications of federated learning in diverse data environments.