About the project
Large Language Models (LLMs) like GPT-4 are transforming how we use information but also exposing new privacy and security risks. This project explores how sensitive data can leak from deployed LLMs and their applications, developing auditing and mitigation methods to make LLMs-based systems safer, more transparent, and accountable across domains.
Large Language Models (LLMs) such as GPT-4 and Claude have revolutionised how we generate, search, and interact with information. However, their growing capability also raises critical privacy and security concerns. Trained on vast datasets, LLMs may inadvertently memorise and reveal sensitive personal or confidential information, creating new forms of privacy leakage.
You'll investigate how such risks emerge and develop techniques to mitigate them, including machine unlearning to remove unwanted training data influences and differentially private fine-tuning to ensure formal protection guarantees. You'll also explore how the rise of Retrieval-Augmented Generation (RAG) and agentic AI systems, which allow models to access external data and act autonomously, introduces new layers of vulnerability and accountability challenges. The aim is to build a comprehensive understanding and practical framework for auditing and safeguarding LLMs across their lifecycle.
This project will be carried out within an interdisciplinary team spanning Law, Finance, Human-Computer Interaction, and industry experts in AI governance and cybersecurity, offering a unique opportunity to conduct impactful, real-world research at the intersection of technology, ethics, and regulation.
