Research project Privacy-Preserving Techniques for Large Language Models

Large language models (LLMs) have led to impressive breakthroughs in artificial intelligence (AI) and natural language processing (NLP). LLMs consist of enormous amounts of parameters trained to process human language. This learning is achieved by processing vast amounts of text.

Multiple studies have shown that LLMs memorize information in their training data that can then leak. These privacy issues are worsening as LLMs grow and consume more training data. The risks are especially dire in domains where data are sensitive, such as the clinical domain. At the same time, these are the domains where AI can have the most beneficial societal impact. This project aims to study the threats to privacy that come with LLMs and investigate privacy-preserving techniques to mitigate these risks. Doing so is crucial if LLMs are to be used in an ethical and legal manner.

This is Thomas Vakili’s PhD project. His main supervisor is Hercules Dalianis, and his co-supervisor is Aron Henriksson.