Nice research! Was that trick with mutation observers necessary however? If they add a constant noise vector, all you have to do is creating a canvas with known data, then call toDataURL() and compare what you get to the original data?