Towards Trustworthy Intelligence: Safeguarding Privacy and Security in Machine Learning

Abstract

The recent proliferation of machine learning (ML) has introduced heterogeneous learning paradigms, driving significant advancements in model capabilities and expanding vast applications in various domains. However, alongside the exciting achievements, emerging learning paradigms bring unforeseen security challenges in realistic scenarios that were previously unaccounted for in traditional ML settings. An abundance of growing security blind spots in different paradigms has unveiled escalating security challenges, exposing ML systems to multifaceted risks that necessitate immediate attention and defensive research in various real-world scenarios.

In the thesis, we systematically address three crucial security challenges across multiple machine learning paradigms through a series of contributions. Fundamentally, we identify the main risks in the distributive nature of recent machine learning systems and seek solutions for two principal security concerns in such systems: (1) the preservation of data privacy during training and inference, and (2) the assurance of model security against malicious behaviors. Specifically, defenses are investigated in the pervasively exploited self-supervised learning (SSL), privacy-preserving federated learning framework, and vision-language models. Firstly, we propose a non-invasive SSL backdoor detection method DeDe that is effective with no reliance on the training dataset or the knowledge of trigger type. Secondly, we develop a general framework PriRoAgg that simultaneously satisfies the robustness and privacy requirements in federated learning protocols. We formalize a novel security notion of aggregated privacy to characterize the minimum amount of user information leakage in FL protocols. Thirdly, we introduce a privacy-preserving federated prompt personalization protocol SecFPP for multi-modal language models.

Date of Award	2025
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Kai CHEN (Supervisor) & Songze LI (Supervisor)