Rocco di Torrepadula, Franca (2024) Advancing Edge AI Systems for Smart Cities. [Tesi di dottorato]

Documento PDF ITEE_Thesis_RoccodiTorrepadula.pdf Visibile a [TBR] Amministratori dell'archivio Download (51MB) | Richiedi una copia

Abstract

Smart cities increasingly rely on artificial intelligence (AI), particularly deep learning (DL), to enhance urban services. While DL models achieve excellent performance across various applications, their high resource and energy demands often require centralized, cloud-based solutions, raising privacy and energy concerns. Moreover, the “black-box” nature of DL models poses challenges to interpretability, which is essential in urban environments. Ensuring privacy and interpretability is also a central concern in privacy regulations such as the European General Data Protection Regulation (GDPR). To address these challenges, this thesis explores distributed AI paradigms, focusing on federated learning (FL) and Edge AI as main solutions. The integration of FL in smart cities is hindered by the heterogeneity of the entities involved. Hence, the thesis presents a personalized Federated Learning framework tailored for public transportation (PT) systems, enabling PT companies to collaboratively train AI models without sharing sensitive data. The framework handles data heterogeneity using a personalized mechanism based on publicly available information and, validated on a large-scale urban dataset, demonstrated notable performance improvements over traditional FL methods. While FL enables the distribution of AI models at the edge, the limited computational resources of edge devices pose challenges. In response, the thesis proposes a framework to deploy and interpret XGBoost models within PT predictive systems. By employing tree-based models suited for low-end edge devices, the framework enhances both interpretability and computational efficiency, while achieving performance comparable to DL models, as proved in the experimental campaign. Lastly, the thesis explores knowledge distillation (KD) to reduce the complexity and resource requirements of DL models. It introduces a novel interpretation of KD using Shannon Entropy and proposes a workflow for optimizing the KD process to find energy-efficient configurations tailored to specific edge scenarios, without relying on expensive grid search.

Downloads

Actions (login required)

Modifica documento