RELEASE-AI: Protecting Sensitive Data Across The AI Lifecycle Disclosure Risks and Mitigations in Trusted Research Environments

TY - CONF

T1 - RELEASE-AI

T2 - Data and Digital Health

AU - Crespi Boixader, Alba

AU - Li, Simon

AU - Liley, James

AU - Ward, Laura

AU - Cole, Christian

AU - Smith, Jim

PY - 2025/9/23

Y1 - 2025/9/23

N2 - Trusted Research Environments (TREs) provide secure access to personal and sensitive data, such as Electronic Healthcare Records (EHRs), for approved research. The increasing use of Artificial Intelligence (AI) and Machine Learning (ML) in health research introduces new challenges for managing privacy risks of individuals’ data. We present a comprehensive framework that embeds mitigation strategies throughout the entire AI project lifecycle, structured across six project phases: design, governance, development, evaluation, disclosure control, and release. This framework aims to empower all stakeholders - researchers, project teams, output checkers, and TRE staff - with clear, phase-specific recommendations on which measures and checks are necessary before model release to help identify potential disclosure risks to data. It promotes early identification of risks with corresponding mitigations and ensures responsibilities are clearly assigned to relevant actors at each stage, from initial planning through to deployment and monitoring. Mitigation strategies include good AI/ML practices, both in terms of code and documentation, privacy-enhancing techniques during training and evaluation, restricting model access via secure query systems, licencing agreements, and adversarial attack testing using tools like SACRO-ML. This process highlights the need to train everyone involved appropriately with relevant role-specific material. A novel tiering system for disclosure control is proposed, categorising AI projects based on the likelihood of attack and associated sensitive data leakage risks. By integrating a lifecycle-focused risk management process with a scalable disclosure control tiering system, this approach enables innovative AI research while maintaining rigorous data protection standards and public trust.

AB - Trusted Research Environments (TREs) provide secure access to personal and sensitive data, such as Electronic Healthcare Records (EHRs), for approved research. The increasing use of Artificial Intelligence (AI) and Machine Learning (ML) in health research introduces new challenges for managing privacy risks of individuals’ data. We present a comprehensive framework that embeds mitigation strategies throughout the entire AI project lifecycle, structured across six project phases: design, governance, development, evaluation, disclosure control, and release. This framework aims to empower all stakeholders - researchers, project teams, output checkers, and TRE staff - with clear, phase-specific recommendations on which measures and checks are necessary before model release to help identify potential disclosure risks to data. It promotes early identification of risks with corresponding mitigations and ensures responsibilities are clearly assigned to relevant actors at each stage, from initial planning through to deployment and monitoring. Mitigation strategies include good AI/ML practices, both in terms of code and documentation, privacy-enhancing techniques during training and evaluation, restricting model access via secure query systems, licencing agreements, and adversarial attack testing using tools like SACRO-ML. This process highlights the need to train everyone involved appropriately with relevant role-specific material. A novel tiering system for disclosure control is proposed, categorising AI projects based on the likelihood of attack and associated sensitive data leakage risks. By integrating a lifecycle-focused risk management process with a scalable disclosure control tiering system, this approach enables innovative AI research while maintaining rigorous data protection standards and public trust.

KW - Trusted Research Environment

KW - Machine Learning

KW - Artificial Intelligence

KW - Release

KW - Disclosure Control

M3 - Poster

Y2 - 23 September 2025 through 23 September 2025

ER -