Qi, Pian (2025) Privacy-Oriented Federated Learning: From Fundamental Challenges to Practical Unlearning Solutions. [Tesi di dottorato]

Documento PDF Phd_Thesis_Pian.pdf Visibile a [TBR] Amministratori dell'archivio Download (12MB) | Richiedi una copia

Abstract

Federated Learning (FL) enables collaborative model training without exposing raw data, making it a compelling paradigm for privacy-sensitive applications. Despite its promise, real-world FL systems face critical challenges: statistical heterogeneity, system constraints, communication bottlenecks, and growing concerns around robustness and trust. This dissertation explores FL from a privacy-oriented systems perspective, aiming to bridge the gap between theoretical guarantees and practical, deployable solutions. The first part presents foundational analyses of fairness, trust, and privacy in heterogeneous FL environments. It further investigates lightweight, resource-aware FL strategies suited for edge devices, and discusses integration with large language models under real-world constraints. The second part introduces a suite of system-level innovations that enhance privacy, resilience, and adaptability across the FL lifecycle. FLAIR addresses scalability and privacy in Industry 4.0 by enabling distributed warehouses to collaboratively fine-tune a multimodal Contrastive Language–Image Pretraining (CLIP) model for industrial product retrieval, using adapter modules and data augmentation to preserve efficiency and privacy. FLAME targets collaborative Distributed Denial of Service (DDoS) detection under strict privacy constraints. It filters client updates using Jensen-Shannon Divergence and applies kernel density estimation–based aggregation to mitigate feature distribution bias, improving robustness against zero-day attacks in non-Independent and Identically Distributed (non-IID) settings. FuGuard introduces a client-level federated unlearning mechanism. It synthesizes generative surrogates to simulate missing data and uses optimal transport to guide constraint-aware forgetting. FuGuard balances forgetting effectiveness, retention performance, and storage overhead, supporting scalable and compliant data removal. Together, these contributions advance the state of FL by proposing scalable, privacy-preserving, and attack-resilient techniques applicable to sensitive domains such as industrial Internet of Things (IoT).

Downloads

Actions (login required)

Modifica documento