Edit: this is a normal algorithm for files in general, but it appears to be a image-specific algorithm used in this case. (A la imageDNA)
Given that Google Image has reverse search, you can almost guarantee that they have approximate hashes.
I remember a episode of some legal drama where our hero the judge had his laptop "hacked" and child poem images installed on it. (Luckily it was make believe so he was able to "delete" them before the other judges arrived with the tipoff but that's not the point)
I think we shall soon enter a legal landscape where mere possession of digital products does not imply legal possession. It's going to be an interesting world.
There are plenty of ways to screw with people.
Either of those two goals could fail and it would still be implausible that an attacker could trigger a false positive without using an image derived from actual known child pornography.
That said, the Google/Microsoft tool does seem to work based on a known image corpus.