GNN-Enhanced Hierarchical Federated Learning in Device-to-Device Networks

The proliferation of intelligent devices and advanced wireless networks has resulted in an explosive growth of data generated at the network edge, creating new opportunities for data-driven services while posing fundamental challenges in privacy preservation, communication efficiency, and adaptability. Federated Learning (FL) provides a privacy-preserving paradigm for collaborative model training across distributed devices without sharing raw data, yet conventional FL architectures suffer from high communication overhead and poor scalability. Device-to-device (D2D) communication enables devices to exchange model updates directly and perform local aggregation, alleviating the communication bottleneck and reducing reliance on the central server. Nevertheless, deploying FL in D2D networks still faces inherent limitations, including heterogeneous device capabilities, non-independent and identically distributed (non-IID) local data, and dynamic network topologies, which can compromise global model convergence and training stability. Moreover, the growing complexity and diversity of edge intelligence tasks further motivate the development of collaborative strategies for efficiently fine-tuning large foundation models within the D2D-assisted FL framework, enabling effective adaptation to diverse downstream tasks. To address these challenges, this thesis investigates a Graph Neural Networks (GNN)-enhanced hierarchical FL architecture in D2D networks, aiming to achieve efficient, adaptive, and scalable federated model training across distributed devices. To begin with, this thesis proposes an asynchronous hierarchical clustered FL method that leverages Graph Convolutional Networks (GCN) to mitigate the impact of heterogeneous computation and communication resources in D2D-assisted FL. In this approach, devices are clustered within the D2D network by formulating the assignment as a graph problem and applying an unsupervised GCN-based strategy that respects D2D connectivity. A global optimizer state is incorporated to reduce model drift caused by non-IID local data, and theoretical analysis establishes an upper bound on the global loss to guarantee convergence. Extensive experiments across diverse network scenarios and datasets demonstrate that the proposed method achieves up to 80% improvement in training efficiency and up to 62% reduction in communication cost. Building on this, a transferable GNN-based clustered FL method is designed to adaptively handle heterogeneity and changing network topologies in dynamic environments. To further mitigate the impact of data heterogeneity and accelerate training, a D2D connectivity-aware dynamic programming algorithm driven by Mutual Information is used to select participating devices within each cluster. Theoretical analysis provides a global convergence bound, and extensive experiments demonstrate 24–78% improvement in training efficiency and 30–88% reduction in communication cost compared to baseline methods. Finally, this thesis develops a multimodal federated parameter-efficient fine-tuning (PEFT) method to enable the efficient adaptation of foundation models in hierarchical D2D-assisted FL architectures. To address the vulnerability of lightweight PEFT parameters to modality heterogeneity and the potential bias propagation through local D2D aggregation, a modality-aware adapter with a shared bottleneck and modality-specific adjustments is devised. Furthermore, a Graph Attention Network (GAT)-driven aggregation scheme is proposed to adaptively weight neighbors according to their modality characteristics in the local D2D topology. Extensive experiments show up to 15.3% higher task performance and up to 64.6% reduction in communication overhead compared to baseline methods, while requiring fewer global rounds, demonstrating significant improvements in efficiency and effectiveness. Collectively, this thesis advances hierarchical FL through GNN-enhanced strategies, paving the way for efficient, privacy-preserving, and scalable deployment of intelligent services in D2D network environments.<p></p>

Author dc:creator

• Yuhong Jiang (21065876)

• Federated Learning
• Device-to-Device Networks
• Graph Neural Networks
• Communication Efficiency

Statement dc:rights

• All rights reserved

Identifier

10779/exe.32685249.v1

OAI identifier oai:identifier

oai:figshare.com:article/32685249