Adversarial defence for Large Language Models (LLMs)

About the Project

Supervisory Team: Prof Stuart Middleton

Large Language Models (LLMs) are powerful but can be tricked - Adversarial Defence protects them. You will improve LLM robustness to prompt-based attack or jailbreaking, exploring novel algorithms for adversarial defence inspired by recent success of adversarial pre-prompt training, reinforcement learning from human feedback (RLHF) and adding safety layers to LLM architectures.

Large Language Models (LLMs) are changing the landscape of NLP applications as we know it today. However, as LLM applications proliferate methods to attack LLMs that induce deliberate error, bias, toxicity and misinformation are also growing.

This PhD will explore adversarial defence methods in the context of LLMs fine-tuned for few-shot applications.

The ambition is to improve LLM robustness in both text-only and multi-modal downstream applications. The nature of LLM robustness will first be explored in the context of prompt-based attack and jailbreaking, followed by an exploration of novel adversarial defence approaches inspired by recent success with ideas such as adversarial pre-prompt training, reinforcement learning from human feedback (RLHF) and adding safety layers to training architectures.

You will join the School of Electronics and Computer Science within the University of Southampton, ranked in the top 100 universities worldwide (QS worldwide ranking 2025). We will support the development of your future career and give you opportunities including teaching assistantships, professional networking via leading organizations such as Alan Turing Institute, £31M RAI UK ecosystem, and access to Future Worlds to explore commercialisation of your research.

Entry requirements

You must have a UK 2:1 honours degree, or its international equivalent.

Experience with Machine Learning is essential.

Fees and funding

We offer a range of funding opportunities for both UK and international students. Horizon Europe fee waivers automatically cover the difference between overseas and UK fees for qualifying students.

Competition-based Presidential Bursaries from the University cover the difference between overseas and UK fees for top-ranked applicants.

Competition-based studentships offered by our schools typically cover UK-level tuition fees and a stipend for living costs for top-ranked applicants.

Funding will be awarded on a rolling basis, so apply early for the best opportunity to be considered.

For more information, please visit our postgraduate research funding pages.

How to apply

Apply now

You need to:

• choose programme type (Research), 2026/27, Faculty of Engineering and Physical Sciences
• select Full time or Part time
• search for programme PhD Computer Science (7089)
• add name of the supervisor in section 2 of the application

Applications should include:

• research proposal
• your CV (resumé)
• 2 academic references
• degree transcripts and certificates to date
• English language qualification (if applicable)

Contact us

Faculty of Engineering and Physical Sciences

If you have a general question, feps-pgr-apply@soton.ac.uk.

Project leader

If you wish to discuss any details of the project informally, please sem03@soton.ac.uk