| dc.description.abstract | Federated learning (FL) is a promising paradigm in machine learning (ML) that addresses the challenges posed by centralised data storage and processing. However, deploying FL faces challenges such as resource limitations in wireless networks and privacy concerns. This thesis focuses on developing over-the-air federated learning (OTA-FL) systems, with a heightened emphasis on preserving privacy and ensuring secure communication. The overarching objective is to strike a balance between the learning performance and the enhancement of privacy and security. The specific contributions are: 1) The optimal misaligned power allocation for differentially private FL (DPFL) systems (MPA-DPFL) is studied. With MPA-DPFL, data privacy is safeguarded through local DP (LDP), achieved by injecting artificial noise into the local gradients before transmission. The closed-form solution for optimal power allocation is derived, and the scenarios employing the proposed scheme outperform unbiased scenarios are explored, which theoretically demonstrates the effectiveness of the proposed approach. 2) The optimal device scheduling for differentially private OTA-FL (S-DPOTAFL) is explored under the unbiased OTA-FL (UB-OTA-FL) scenarios with central DP (CDP), where participants' privacy is protected by channel noise. The closed-form solution that achieves a balance between the improved alignment coefficient and the decreased number of participants is obtained. Conditions are explored under which S-DPOTAFL outperforms scenarios without considering device scheduling. 3) An optimal design is explored to jointly optimise device scheduling, the alignment coefficient, and aggregation rounds for differentially private over-the-air FedAvg (OTA-FedAvg) systems (O-DP-OTA-FedAvg) with constrained sum-power and privacy budgets. This investigation extends the relationship between the alignment coefficient and the number of participants to a more general scenario where all devices have distinct peak transmission power. Additionally, the study delves into the trade-off between reducing local training errors and increasing aggregation distortion under a limited sum-power budget. 4) Device scheduling schemes, intended to bolster privacy and security through the utilisation of channel noise in channel-weighted OTA-FL systems, are investigated across three levels of channel noise: (1) channel noise is sufficient to protect privacy and security with all device participation; (2) channel noise is sufficient to protect privacy and security with partial device participation; (3) channel noise is insufficient to protect privacy and security with any device participation. 5) A novel secure and privacy-preserving OTA-FL (SP-OTA-FL) framework is introduced. In this framework, a subset of devices are designated as jammers to transmit artificial noise for enhancing privacy and security. A thorough investigation is carried out to examine optimal designs incorporating transmission design and device scheduling for SP-OTA-FL in both unbiased OTA-FL (UB-OTA-FL) and biased OTA-FL (B-OTA-FL) scenarios. Theoretical analyses, encompassing privacy, security, and convergence analysis, are conducted to evaluate the impact of each scheme on the improvement of privacy and security, as well as learning performance. Based on these results, optimisation problems are formulated and addressed via closed-form solutions or heuristic algorithms. Simulation results demonstrate that the proposed schemes effectively enhance learning accuracy while also enhancing privacy and security in FL. | en_US |
