Heterogeneous Federated Dynamic Graph HyperNetwork for Image Classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Federated learning (FL) facilitates privacy-preserving collaboration but struggles with heterogeneous models and non-IID data.
• Real-world FL systems face challenges with newly joined and abnormal clients, impacting stability and convergence.
• Existing methods often fail to adequately address model heterogeneity and dynamic client behaviors in FL.

Purpose of the Study:

• To propose HFedDGHN, a Heterogeneous Federated Dynamic Graph HyperNetwork, for robust and personalized FL.
• To jointly model inter-client relations and personalized parameter generation for heterogeneous FL.
• To improve adaptation to new clients and enhance robustness against abnormal clients.

Main Methods:

• A graph structure learner adaptively constructs a dynamic collaboration graph based on client correlations.
• A graph-convolutional hypernetwork generates personalized parameters for heterogeneous model architectures.
• Meta-learning enables efficient generalization and adaptation for newly joined clients.

Main Results:

• HFedDGHN demonstrates superior accuracy compared to state-of-the-art personalized and heterogeneous FL methods.
• The dynamic graph construction effectively isolates abnormal clients, improving system robustness.
• The framework shows improved scalability and adaptation capabilities for real-world FL deployments.

Conclusions:

• HFedDGHN effectively addresses heterogeneity, non-IID data, and dynamic client issues in federated learning.
• The proposed method achieves state-of-the-art performance in image classification tasks.
• HFedDGHN offers a robust, scalable, and adaptable solution for practical federated learning systems.