Federated Learning - Contextual Efficiency Optimisation and Adversary

Abstract

Artificial Intelligence (AI) has been widely studied, applied, and achieved impressive successes in various sectors. Among these successes, Federated Learning (FL) stands out as an innovative paradigm that enables distributed devices to collaboratively train machine learning models without transferring sensitive data to central servers. This de centralised paradigm, crucial for privacy preservation, also introduces significant prac tical challenges, particularly due to the heterogeneity among participating clients, which include variations in computational power, distinct data distributions, and diverse user behaviours, all of which complicate the training process in efficiency and convergence. Existing FL frameworks often apply a uniform learning configuration across all clients, neglecting individual device characteristics and user contexts, inevitably leading to in efficient utilisation of hardware resources and increased energy consumption within the federation. Moreover, optimisation techniques intended to exhaust computation potential of the system, can inadvertently expose systems to vulnerabilities, compromising security and reliability at both software and hardware levels. This thesis addresses the critical gap of the contexts of FL including heterogeneous hardware specifications and optimisation technique employments, by proposing adaptive, context-aware methodologies that optimise efficiency, while analysing and mitigating as sociated security risks within federated learning environments. First, a Mixed-Precision Over-the-Air Federated Learning (MP-OTA-FL) framework is introduced, allowing cli ents to operate at different precision levels based on their hardware capabilities. Exper imental results show that MP-OTA-FL significantly improves energy efficiency and task performance, particularly for resource-constrained devices. Second, a novel Retrieval Augmented Generation (RAG)-based user profiling framework is proposed, dynamically optimising client precision decisions based on contextual factors such as hardware con straints and user preferences, achieving higher global model accuracy and user satisfac tion. Finally, the thesis explores hardware vulnerabilities introduced by FL efficiency optimisations, presenting a remote Rowhammer attack vector exploiting physically in duced perturbations on client devices, demonstrating practical security risks under real istic conditions. Collectively, these contributions provide comprehensive strategies for contextually optimised, secure, and efficient FL deployment, balancing performance and security considerations. The findings offer foundational guidance for developing robust FL systems tailored to diverse and resource-limited real-world scenarios.