Decentralised Federated Learning over Wireless Communication Networks

As the proliferation of IoT devices and smart technologies continues, their potential
remains underutilised due to significant privacy and data sensitivity concerns. In this
thesis, I present a secure, decentralised, and intelligent framework designed to optimise
diverse problems through Federated Learning (FL) schemes and machine learning models
on endpoint devices. Unlike traditional centralised approaches, this framework enables
collaborative problem-solving without data sharing, relying instead on the exchange of
model updates to refine a global solution. The key to this decentralised model is the
establishment of robust and efficient communication networks that manage interactions
and data exchanges between devices.
This work explores Decentralised Federated Learning (DFL), which enhances participant collaboration while ensuring privacy and mitigating the communication bottlenecks
inherent in Centralised Federated Learning (CFL). By improving inter-device communication within the DFL network and optimising the associated learning models, I aim to
boost overall system performance and reliability. Furthermore, the presence of adversarial devices poses significant threats; thus, strategies to exclude untrustworthy devices are
critical to maintaining the integrity and efficiency of the network.
This thesis contributes towards a comprehensive analysis of network communication, geometric configurations, and system robustness. I introduce innovative DFL models and
simulation techniques, demonstrating a robust and server-free FL process. Enhancements in model accuracy have been achieved, leading to an intelligent, low-latency, and
adaptable framework suitable for various important applications, including Autonomous Vehicles (AVs) and IoT systems.
While further advancements are necessary, this thesis marks substantial progress towards
a flexible and distributed DFL scheme. It is anticipated that this foundational work will
encourage continued enhancements in communication efficiencies, fostering more effective
collaboration and sustained privacy in FL environments.