Trustworthy Aggregation for Federated Learning as a Service Over Wireless Networks

Thumbnail Image
Ads, Mohamed Magdi Mohamed Mohamed
Machine Learning , Wireless Networks , Federated Learning , Edge Computing
This thesis investigates the challenges of implementing Federated Learning (FL) in diverse wireless edge networks, focusing on mitigating the impact of device heterogeneity, communication impairments, and trustworthiness concerns. Our research addresses the limitations of traditional FL approaches, which often suffer from slow convergence and reduced accuracy due to varying data quality, transmission rates, and error-prone channels. we propose novel risk-aware and accelerated FL frameworks that leverage device trustworthiness metrics and dynamic aggregation schemes. By classifying clients based on location-dependent performance and trustworthiness profiles, the frameworks prioritize participation from devices with high transmission rates while progressively incorporating data from cell-edge clients to enhance data diversity. In Addition, the frameworks employ debiasing techniques to account for transmission errors. To address the issue of unreliable clients, we explore two distinct validation approaches. The first approach utilizes a dedicated validation dataset to identify and eliminate untrustworthy clients, ensuring data integrity and model accuracy. Given the potential lack of such a dataset, the second approach investigates alternative mechanisms without relying on a separate validation dataset. The effectiveness of these frameworks and validation mechanisms is demonstrated through extensive simulations in a range of wireless settings, including conventional terrestrial cellular networks and emerging 6G non-terrestrial networks with unmanned aerial vehicles. Results showcase the superior performance of the proposed frameworks compared to conventional FL approaches across these diverse environments, highlighting their adaptability and robustness.
External DOI