I'm not so sure about that. I believe the problem has been solved, the solution just isn't widely known yet. I read a paper on arxiv probably a year ago that describes a method that seems pretty straightforward and secure, but I've never seen anything about it since. It involved, essentially, throwing out any records which could actually contribute to a change in a statistical measure. You basically end up finding what aspects of the data are actually identifiable, and throw out any records that contain that. It's guaranteed not to screw up your observations because, by definition, if something is statistically significant it has to show up often enough that it CANT be used to single out a source.