TransPECT: Securing AI NLP-Transformer Models for Safe Release in TREs

AI language models are increasingly used in research to analyse unstructured text such as clinical notes and administrative records. While these models can be trained securely within TREs, they may unintentionally memorise fragments of the training data, creating potential disclosure risks if models are exported.

TransPECT will develop tools and methods for evaluating these risks before models are released. Working with DataLoch – a Trusted Research Environment in South-East Scotland – the project will extend semi-automated disclosure control approaches, including SACRO-ML, to assess transformer-based language models.

The project will evaluate different model architectures, including large language models and lightweight alternatives, and test privacy-enhancing techniques such as differential privacy and machine unlearning.

Public and stakeholder engagement will help shape standards for evaluating model risk. Through workshops and reference groups, the project will ensure that proposed approaches are transparent, understandable and aligned with public expectations.

By the end of the project, TransPECT will:

• Develop a toolkit for assessing disclosure risks in transformer language models
• Extend SACRO-ML to support the evaluation of text-based AI models
• Compare privacy risks across different model architectures
• Test mitigation strategies, including differential privacy and machine unlearning
• Produce practical guidance for TRE governance of AI model release

Project information

Lead organisation: University of Edinburgh
Principal investigator: Dr Arlene Casey
Project duration: 12 months
Project partners: DataLoch, Health Informatics Centre, King’s College London, Public Health Scotland
Funding provided: £422,343
Primary contact email: Arlene.Casey@ed.ac.uk