1. Sometimes it is.
2. I already gave substantive reasons why this will commonly occur in environmental medical research -- dense datasets about small populations that capture sensitive information about individuals.
> I do not see how it's takeway from that paper.
Because that's literally the only assumption in the adversarial model other than "access to the published dataset". And correlating with an auxiliary dataset is literally how all actual instantiations of this attack work.
> just sprinkle some private data on it, show it to IRB, they cry out "no, this can't be released, no way!" - and voila, you are safe from review.
1. My proposal doesn't say "IRB happened to say you can't publicly publish this one particular protocol." My proposal says that IRB explicitly judges that no such protocol exists.
2. IRBs are not that arbitrary and are themselves audited.
3. If you're this paranoid about intent, then you can't trust the research anyways. If the research is committing intentional fraud, they could just do it at the data collection step.
The way I see it, either:
A. Even if we can't trust individual scientists, we can more-or-less trust a broad subset of scientists with different and well-aligned motives (as in, a whole group of: the IRB board, the PI, the people reviewing the IRB board) when they say that data can't be shared for privacy reasons; or,
B. We should expect that literally dozens of scientists, all with different motives, will routinely and actively collude to commit wide-spread fraud.
If (A), my approach makes sense. If we max out paranoia and go with (B), then I don't know why we should trust the data regardless of whether it's published publicly...
> important take away from the above paper is that you need to know about all the other public data in the universe in order to really ensure anonymity.
I do not see how it's takeway from that paper.
> You never know what dataset will be leaked/published tomorrow correlating columns D-F in the spreadsheet you gave EPA to someone's first and last name.
You also never know whether somebody doesn't just break into your system, downloads all secret data and publishes them. Yes, you have no guarantee against bad actors and mistakes. But if that's the border condition then no research should be performed at all - after all, once the data exists, there's no guarantee it won't somehow get leaked out.
> The situation where IRB says "no" to public disclosure should explicitly trigger exemption
IRB is a part of the institution, right? So you essentially delegating the decision about the exception to (part of) the institution. What the point of having the rule then if every organization can make their own exceptions on demand? And that makes trivially easy to hide the data in sloppy research - just sprinkle some private data on it, show it to IRB, they cry out "no, this can't be released, no way!" - and voila, you are safe from review.