Re:seriously? (Score 1)

According to the brain cells I still retain, whatever their final number, it seems to me that to find out by this method that someone's medical data is inside the model, you need to test this by already having said medical data in hand. If so you can't by this method learn whether the person had the disease. What you can learn is whether they signed the consent form to participate in lending their personal data to the model builders.

I think you are mostly right. But who is to say that some of the data set isn't publically available? Lets say I know some set of things about you that (almost) uniquely identify you (zipcode, gender, doctor, genetic heritage, birthday, etc..) that was used in training the model. A lot of this stuff would be provided in an insurance application, job application, etc. And even more is scrape-able or inferable from the web. Maybe that data alone is enough to help me match. But if not, then I can fuzz over the whatever medical data until something lands above the confidence level I am looking for.

Bam no job or no insurance or no whatever you want. They don't have to be 100% accurate. Its better for them to overcharge, exclude or otherwise mistreat a few people that are healthy than to accept 1 person with a medical condition.

Also .. this applies to everything that GANs get trained on .. not just medical data or faces. Although since training sets, while large, tend to be much much smaller than the actual population so I am not sure how valuable this would be.

Assuming the underlying research is true.