Skip to main content
Back to industries

Industry

Health and life sciences

On the UCI Diabetes readmission dataset, differential privacy at its tightest setting costs minority subgroups 2.6 times more accuracy than the majority, while the membership leakage it was applied to prevent is already close to zero.

How we work in this sector

Privacy and fairness are usually audited by separate teams on separate timetables, which is how a trade-off between them stays invisible. We put both on one report card for the same model. The same discipline applies to the data underneath, where an ontology decides whether a knowledge graph can be trusted to answer a question it was not built for.

Research in this sector

Every figure in the work below is reproducible from an open repository. There are 6 studies here.