Artificial Intelligence Governance is enabling the development of AI within SeRP-powered TREs. AI models can be prone to privacy attacks and data leakage which allow for inference or reconstruction of the training data used to develop them.
There are also various characteristics in these models which can make them more vulnerable to these privacy attacks.
Release of these models from a TRE pose a unique risk. In collaboration with SeRP tenant, Dementias Platform UK (DPUK), a community of data providers, researchers and public representatives was established to review the risks involved in AI model release. DPUK-funded staff held a series of workshops highlighted the important role that TREs can play in the safe development and sharing of AI models in sensitive healthcare data.
Through this work, an AI model risk evaluation framework has been developed that has been deployed across multiple SeRP TREs to allow AI models to be securely developed and assessed.
An AI model risk evaluation framework
1. Training and accreditation to address a lack of researcher knowledge.
2. AI risk impact assessments to aid TRE governance personnel.
3. AI experts, clinicians and public involvement to evaluate risks and benefits.
4. Tools to generate synthetic data for privacy and utility evaluation.
5. Privacy attack simulations on AI models to provide confidence prior to release.
6. Continuous development of an AI Risk Index based on rankings of ongoing AI development and release.
7. Secure hosting for AI model releases and encryption for safe querying.
8. Researcher agreements and AI model licensing for distribution and utility controls.
9. Application of ‘FAIR’ principles in AI model libraries within TRE’s.
More information
Further information produced by the DPUK team about this work can be found at AI_Risk_Report.
AI hosting
SeRP’s Innovation Team in Collaboration with DPUK has enabled the ability to share, federate and deploy models so that we can facilitate AI research in a safe and secure way. This world leading governance coupled with our technological abilities has allowed us to lead the way in AI research within TREs.
Share
Share the model via the TRE to enable open reproducible science while keeping data secure.
Federate
Federate between different data environments to enable fine tuning or validation of AI models.
Deploy
Host the model via the TRE and allow it to be securely queried externally with restrictions.