Date of Award:

8-2024

Document Type:

Dissertation

Degree Name:

Doctor of Philosophy (PhD)

Department:

Electrical and Computer Engineering

Committee Chair(s)

Rose Qingyang Hu

Committee

Rose Qingyang Hu

Committee

Todd Moon

Committee

Zhen Zhang

Committee

Bedri Cetiner

Committee

Ziqi Song

Abstract

Artificial intelligence (AI) is transitioning from a long development period into reality. Notable instances like AlphaGo, Tesla’s self-driving cars, and the recent innovation of ChatGPT stand as widely recognized exemplars of AI applications. These examples collectively enhance the quality of human life. An increasing number of AI applications are expected to integrate seamlessly into our daily lives, further enriching our experiences.

Although AI has demonstrated remarkable performance, it is accompanied by numerous challenges. At the forefront of AI’s advancement lies machine learning (ML), a cutting-edge technique that acquires knowledge by emulating the human brain’s cognitive processes. Like humans, ML requires a substantial amount of data to build its knowledge repository. Computational capabilities have surged in alignment with Moore’s law, leading to the realization of cloud computing services like Amazon AWS. Presently, we find ourselves in the era of the IoT, characterized by the ubiquitous presence of smartphones, smart speakers, and intelligent vehicles. This landscape facilitates decentralizing data processing tasks, shifting them from the cloud to local devices. At the same time, a growing emphasis on privacy protection has emerged, as individuals are increasingly concerned with sharing personal data with corporate giants such as Google and Meta. Federated learning (FL) is a new distributed machine learning paradigm. It fosters a scenario where clients collaborate by sharing learned models rather than raw data, thus safeguarding client data privacy while providing a collaborative and resilient model.

FL has promised to address privacy concerns. However, it still faces many challenges, particularly within wireless networks. Within the FL landscape, four main challenges stand out: high communication costs, system heterogeneity, statistical heterogeneity, and privacy and security. When many clients participate in the learning process, and the wireless communication resources remain constrained, accommodating all participating clients becomes very complex. The contemporary realm of deep learning relies on models encompassing millions and, in some cases, billions of parameters, exacerbating communication overhead when transmitting these parameters. The heterogeneity of the system manifests itself across device disparities, deployment scenarios, and connectivity capabilities. Simultaneously, statistical heterogeneity encompasses variations in data distribution and model composition. Furthermore, the distributed architecture makes FL susceptible to attacks inside and outside the system.

This dissertation presents a suite of algorithms designed to address the challenges effectively. Mew communication schemes are introduced, including Non-Orthogonal Multiple Access (NOMA), over-the-air computation, and approximate communication. These techniques are coupled with gradient compression, client scheduling, and power allocation, each significantly mitigating communication overhead. Implementing asynchronous FL is a suitable remedy to solve the intricate issue of system heterogeneity. Independent and identically distributed (IID) and non-IID data in statistical heterogeneity are considered in all scenarios. Finally, the aggregation of model updates and individual client model initialization collaboratively address security and privacy issues.

Share

COinS