German researchers have shown that AI models used to help diagnose medical conditions can be queried to reveal whether a specific patient's data was in their training set. At the individual level, they report near-perfect success.
Health27 September 20265 min readSources 2
German researchers report that medical diagnosis AI models can be tricked into revealing whether a specific patient's data trained them. Their Nature paper, published on Wednesday, reports near-perfect success on individual patients. The finding lands as the FDA, the EU and the WHO are still writing the rules for auditing these devices.
Health27 September 20266 min readSources 4
Medical diagnosis AI is being sold faster than regulators can audit it. Three recent studies point to the core problem: the models are hard to check, hard to secure and hard to quantify.
Health27 September 20265 min readSources 5
A study from the University of Texas at Austin tracked 82 young adults for several months with consumer wearables. It found that physical activity lengthened REM latency, the time it takes to enter REM sleep, according to the university's announcement.
Health27 September 20266 min readSources 2
Discriminative medical AI models can be queried to reveal whether an individual patient's data was used to train them, with "near-perfect attack success" at the individual level, according to a Nature paper published Wednesday and reported by The Register.
Health27 September 20264 min readSources 2
Two research findings published 14 months apart show medical AI failing at opposite ends of the same task: diagnosing patients early, and keeping the patients in its training data anonymous.
Health27 September 20264 min readSources 4