FedDyH: A Multi-Policy with GA Optimization Framework for Dynamic Heterogeneous Federated Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Biology

Background:

• Federated learning (FL) enables privacy-preserving distributed learning, particularly for cross-institutional medical image analysis.
• Real-world data exhibit dynamic heterogeneity (e.g., disease progression, imaging device variations), causing catastrophic forgetting and performance degradation in FL models.
• Existing FL methods inadequately address dynamic heterogeneity, limiting their effectiveness in complex, evolving medical datasets.

Purpose of the Study:

• To propose the FedDyH framework, an innovative solution designed to tackle dynamic data heterogeneity challenges in federated learning.
• To enhance the robustness and accuracy of federated learning models in dynamic, heterogeneous environments by drawing inspiration from biological adaptive regulation.

Main Methods:

• The FedDyH framework employs cross-client knowledge distillation to simulate intercellular information transfer, preserving local features and mitigating knowledge forgetting.
• A dynamic regularization term adaptively adjusts its strength, mimicking regulatory T cells to balance global convergence with local specificity.
• A genetic algorithm (GA) is integrated for adaptive hyperparameter optimization, simulating biological evolution for improved model adaptability and performance.

Main Results:

• The FedDyH framework demonstrated significant accuracy improvements over the SOTA baseline FedDecorr on benchmark datasets: MNIST (+2.59%), Fashion-MNIST (+0.55%), and CIFAR-10 (+5.79%).
• Experimental results validate the framework's effectiveness in addressing data heterogeneity in dynamic environments.
• The study highlights the novelty of using optimization algorithms like GA for hyperparameter tuning in federated learning.

Conclusions:

• The FedDyH framework offers a robust and adaptive solution for federated learning in dynamic, heterogeneous environments, crucial for medical image analysis.
• The biologically inspired approach enhances model stability and predictive accuracy by effectively managing data variations.
• This work advances federated learning by introducing novel methods for handling dynamic heterogeneity and optimizing model performance.