Savoia, Martina (2025) Federated Learning for Edge Intelligence: From Privacy Preservation to Continual Adaptation. [Tesi di dottorato]

Documento PDF Savoia_Martina_38.pdf Visibile a [TBR] Amministratori dell'archivio Download (19MB) | Richiedi una copia

Abstract

In traditional Machine Learning (ML), data from individual devices or organizations are collected and transmitted to a central server, typically in the cloud, where the model is trained. Although effective, the centralized paradigm presents critical limitations. First, it raises significant privacy concerns, as users have to share sensitive data. Second, the transmission of large volumes of raw data results in high communication and bandwidth costs. Finally, centralized systems represent a single point of failure, making them vulnerable to attacks and outages. Edge Computing addresses some of these issues by bringing computation closer to data sources, such as smartphones or IoT devices, thereby reducing latency and improving privacy. Federated Learning (FL) enables distributed training across devices, combining scalability with privacy preservation. Moreover, FL reduces bandwidth consumption and supports personalized AI models, while maintaining a shared global model. In parallel, Continual Learning (CL) is emerging as a complementary paradigm that allows models to adapt over time as new data become available, without requiring retraining from scratch. This combination is particularly promising in real-world scenarios where data evolve dynamically, as it improves efficiency and reduces energy consumption. This thesis explores the principles, strengths, and challenges of FL, and investigates its integration with CL. It introduces novel approaches such as Eco-FL, which reduces energy consumption through intelligent client selection, AGRIFOLD, a realistic agricultural case study addressing non-IID data, and FEDCOM, which integrates CL into FL to efficiently incorporate new data during training, saving energy by avoiding full retraining. Furthermore, it presents a survey on the application of Federated Continual Learning (FCL) in Digital Twins (DTs), an emerging and highly relevant domain for these techniques. Overall, FL represents a rapidly evolving research area. Ongoing research is essential to address open challenges related to privacy, energy efficiency, real-time data adaptability, and practical and sustainable Artificial Intelligence.

Downloads

Actions (login required)

Modifica documento