Darktrace funded studentships - 03

The purpose of this project is to detect privacy leaks in multimodal medical data, and prevents unauthorised disclosure of confidential personal health information to an external or untrusted environment. In medical data, this leakage is permanent since one cannot cancel the health history or DNA sequence, meaning the exposure can have lifelong consequences. The project has two phases: one to develop an interpretable framework to protect the privacy of multimodal medical data, and two leveraging a dynamically quantified sensitivity knowledge graph to monitor unknown leaks.

Multimodal medical data, such as genomic sequences, radiological images and clinical records, have significant potential to advance precision medicine but also raise major privacy concerns. Traditional anonymisation and differential privacy methods apply uniform protection, overlooking differences in sensitivity across modalities and features. This often reduces diagnostic value and leaves data vulnerable to leakage.

Such methods also lack medical-specific modelling and intent recognition, making them less effective against complex threats like multimodal splicing and cross-system data transfers. Rule-based protection and single unsupervised detection methods cannot fully address the risk of hidden leakage. Additionally, many privacy mechanisms lack clinical interpretability, hindering regulatory compliance and the widespread adoption of healthcare AI.

The project aims to integrate differential privacy (DP), statistical learning and large language models (LLMs) to realise multimodal medical data protection, and develop differentiated DP solution for detecting unknown leakages, via knowledge graph built by combing modality-level hierarchical clustering, entropy weighting and feature-level Shapley value analysis. The core dataset that will be trained includes public benchmarks such as TCGA-BRCA for multi-omics analysis and MIMIC-IV for clinical logs and electronic health records. Auxiliary resources include but are not limited to the SNOMED CT medical terminology system, the CSECICIDS2018 cybersecurity dataset, and synthetic leakage samples generated by models such as generative adversarial networks and diffusion models.

94% of Loughborough’s research impact is rated world-leading or internationally excellent. REF 2021