The rapid growth of the Internet of Things (IoT) will connect billions of devices to sixth-generation and beyond networks, generating massive amounts of edge data that demand on-device learning to meet stringent latency and privacy requirements. While machine learning (ML) can leverage distributed data and computational resources, centralized training requires frequent data transmission to servers, introducing privacy risks, communication overhead, and power constraints; challenges that intensify in hierarchical wireless networks (HWNs) with multi-tier aggregation. Hierarchical Federated Learning (HFL) has emerged as a promising paradigm that enables distributed learning without sharing raw data; however, HFL still suffers from slow convergence, additional aggregation overhead, and degraded performance under heterogeneous and highly non-IID device data. This dissertation builds on the Clustered Federated Learning (CFL) paradigm to develop several resource-efficient and personalized learning frameworks tailored for HWNs. The proposed Clustered Multitask Federated Distillation (CMFD) leverages knowledge distillation to enhance system compatibility and scalability in heterogeneous Industrial IoT environments, enabling collaboration among devices with diverse capabilities. A two-phase client-selection and two-level model-aggregation design is further introduced for IoT and vehicular applications, improving fairness, resource utilization, latency, and energy efficiency through greedy and round-robin scheduling. To address the scarcity of labeled data at the edge, the Clustered Federated Semi-Supervised Learning (CFSL) framework integrates semi-supervised learning and pseudo-label generation using specialized cluster models, enhancing learning accuracy and convergence speed. Beyond CFL-based approaches, this research proposes PFL-TPP, a Personalized Federated Learning framework that employs transformer pruning and hypernetwork-driven personalization to reduce computation and communication costs while preserving model accuracy in wireless edge environments. All frameworks jointly optimize system-level parameters, including transmission power, bandwidth allocation, and latency, by modeling them as integrated cost functions within HWNs. Extensive experiments demonstrate that the proposed CFL-based schemes achieve faster convergence, lower latency, and reduced energy consumption under data heterogeneity. CFSL improves labeling accuracy and reduces labeling delay, while CMFD enhances scalability and energy efficiency. Finally, the transformer-based PFL-TPP framework delivers substantial improvements in accuracy, resource efficiency, and training speed, demonstrating its practicality for real-world personalized FL at the wireless edge.
To address these challenges, this research builds on the Clustered Federated Learning (CFL) paradigm to develop several novel frameworks, including Clustered Multitask Federated Distillation (CMFD), fair client selection and model aggregation, and Clustered Federated Semi-Supervised Learning (CFSL). By clustering devices with similar data distributions and assigning specialized models to each cluster, these frameworks enhance learning efficiency and personalization while effectively managing data heterogeneity and resource constraints in HWNs. Specifically, CMFD integrates knowledge distillation to improve system compatibility and scalability in heterogeneous Industrial IoT environments, enabling collaboration among devices with diverse capabilities. A two-phase client selection and two-level model aggregation scheme is further introduced for IoT devices and intelligent vehicles, enhancing fairness and resource utilization through a combination of greedy and round-robin selection, while reducing latency and energy consumption. Recognizing that much device data is unlabeled, CFSL integrates semi-supervised learning to generate accurate pseudo-labels using specialized models, thereby enhancing the use of unlabeled data and accelerating convergence. Beyond CFL, this research introduces transformer-based models, proposing PFL-TPP (Personalized Federated Learning with Transformer Pruning and Personalization) for wireless edge networks that integrates dynamic pruning of feed-forward layers with hypernetwork-driven personalization, effectively reducing computation and communication costs while preserving model personalization. Overall, this research jointly optimizes critical network parameters, including transmission power, bandwidth allocation, energy consumption, and latency, by modeling them as system-level costs within HWNs. These optimizations are seamlessly integrated into both the clustered and transformer-based FL frameworks, ensuring faster convergence, higher model accuracy, and improved energy efficiency across diverse deployment scenarios. Extensive experiments validate the effectiveness of all proposed frameworks across hierarchical and edge-intelligent networks. The CFL-based schemes, CMFD, CFSL, and the fair client-selection and model-aggregation framework, consistently achieve accelerated convergence, lower latency, and reduced energy consumption. CMFD enhances accuracy and energy efficiency under data heterogeneity, while the two-phase client-selection design improves fairness and communication efficiency. CFSL further improves testing and labeling accuracy with reduced labeling delay. Finally, the PFL-TPP framework delivers substantial improvements in accuracy, energy efficiency, and training speed, demonstrating its scalability for real-world PFL at the wireless edge.
| Date of Award | 2025 |
|---|
| Original language | American English |
|---|
| Awarding Institution | - HBKU College of Science and Engineering
|
|---|