Let's say Facebook (a) didn't offer advertisers direct personal information, only access to demographic profiles, and (b) set some minimal standard so that you can't target profiles that result in fewer than 15 people. (What's that data set that does that already?)

Can't a sufficiently large advertising aggregator run a sufficiently large set of multivariate ads, do Statistical Magic, and discover the real demographics of each of the users - at least enough to reassociate them with their real-world, extra-Facebook identity? Sure, they'd be violating ToS and probably various laws, but we're talking about attack surface, not public policy (yet).

This paper about "k-anonymity" and "privacy-preserving distributed protocol" might yield some clues, but my brain is fundamentally incapable of reading those once-"beautiful" LaTeX fonts. (I would pay good money for an app/Mechanical Turk/whatever that reprocesses text with a modern font.)