Abstract
The next generation of wireless networks will increasingly depend on machine learning (ML) to enable efficient on-device intelligence and adaptive decision-making at the network edge. However, resource limitations, communication delays, and privacy concerns prevent edge devices from transferring their full datasets to a centralized cloud for model training or inference. To overcome these challenges, federated learning (FL) has emerged as a distributed ML paradigm that enables collaborative model training without sharing raw data. FL allows edge devices to train their local models and transmit trained parameters to a central parameter server (PS) for aggregation and updating of the global model. However, wireless constraints such as fading effects, privacy issues, limited communication resources (such as bandwidth and power), and scalability pose significant challenges. This dissertation focuses on improving the practicality of FL by jointly optimizing learning and communication processes under realistic wireless and hardware limitations. First, an optimal global aggregation method is proposed for FL over fading channels, integrating differential privacy (DP) to protect client updates while preserving model utility. Theoretical analysis and experiments demonstrate robust convergence under unreliable and noisy wireless environments. Second, to further enhance scalability in wireless FL, an autoencoder-based over-the-air computation (AirComp) system is introduced. Unlike conventional analog AirComp, the proposed digital AirComp framework employs neural encoders and decoders to enable end-to-end learning-driven communication with improved spectral efficiency, and it is further extended to support higher-order modulations. Third, a communication- and storage-efficient federated split learning (CSE-FSL) framework is developed to reduce transmission load on edge devices and storage demands on the parameter server. A case study demonstrates the feasibility of FSL in real-world edge computing systems. Together, these contributions address the key challenges of fading, privacy, scalability, and communication efficiency in federated learning. The proposed methods bridge theoretical insights and practical implementations, advancing the development of communication-efficient, privacy-preserving, and hardware-aware federated learning for future intelligent wireless networks.
