[azalemeth]

I'm friends with a professor of steganography. Apparently most cinema watermarking is based on very heavily error correcting codes within the wavelet domain that are specifically designed such that they are resistant to collusion attacks, i.e. the statistical properties of the "indistinguishable from random" noise are such that it is highly correlated among different viewers such that they are very much more likely to have bits in common rather than bits different. I'm relatively sure that the obvious things like taking the mean of two images (or randomly picking one of them) have been considered.

Put it this way -- You've got huge amounts of cover data (a hard drive's worth) and a desire to encode at most, what, 128 bits of data, across about two hours, with as much redundancy as possible. There are plenty of patents that explain in detail how.

My friend considers this a moderately distasteful problem, and mostly works on steganalysis, identifying where steganographic techniques have been used, as he thinks it's more interesting and frequently more morally justified...