Chiaro, Diletta (2024) Federated Learning under Operational Constraints: A Study on Statistical and Device Heterogeneity. [Tesi di dottorato]

Documento PDF FInal_PhD_diss_DilettaChiaro.pdf Visibile a [TBR] Amministratori dell'archivio Download (30MB) | Richiedi una copia

Abstract

Traditional centralized Machine Learning (ML) frameworks require users to upload their personal data to a centralized server for model training, which poses a significant threat to data privacy. To address this issue, Federated Learning (FL) has emerged as a promising paradigm that enhances privacy-preserving ML. As a decentralized ap- proach, FL allows multiple clients to collaboratively train a shared global model without exposing their individual data, thereby ensuring that user privacy is maintained. How- ever, FL is not without its limitations. Key challenges include statistical heterogeneity, where client data is Non-Independent and Identically Distributed (Non-IID), as well as device heterogeneity and increased communication overhead. These issues have be- come the focal point of recent research efforts, given the growing interest and adoption of FL in recent years. This thesis delves deeply into these challenges, offering a com- prehensive analysis of FL and proposing innovative approaches to mitigate the non-IID data issues that often arise. Key solutions introduced include FL-Enhance, KAFE and FLOWS. Moreover, a novel framework named SHELOB-FFL is proposed to address device heterogeneity while tackling the challenges posed by statistical heterogeneity. The thesis concludes by presenting a practical application of FL, showcasing the devel- opment of FL-FD, a system designed for Fall Detection. This application highlights the real-world utility of FL in addressing distributed learning challenges while maintaining privacy and efficiency. As interest in FL continues to grow, ongoing research and development will be critical to fully realize its capabilities and establish it as a foundational technology in secure and distributed ML.

Downloads

Actions (login required)

Modifica documento