Hmm... it seems to me that if the watermark does not cause audible distortion, then shouldn't there be a simple algorithm to remove the watermark: namely, to watermark the same file again with different information?

I suppose this requires you to know the algorithm used to watermark the file in the first place.  All the algorithms discussed here: http://www.ece.uvic.ca/~aupward/w/watermarking.htm seem like they would be pretty trivially removed by the applying them again with different watermark data.

I could imagine that one could choose from different frequency bands, and hide the data in only one of them, so that in order to remove it you'd need to know which band it's in, or over-write the data in *all* bands. But in order to make the ratio of noise generated by watermark-all-bands vs watermark-one-band large enough that watermarking all bands would introduce unreasonable noise, but just watermarking one wouldn't, it seems the bands would have to be very small.  How much data could you really fit in them (and what are the chances that any one of the bands happens to look "watermarked" with something meaningful)?

But perhaps there is some other way to parametrize the algorithm so that if the parameters are not equal, the watermarks can coexist.  I would be interested in seeing it.  But my feeling is that the amount of data you can robustly store is small enough that splitting it into a reasonable number of bands to thwart removers would make the capacity unreasonably small.

Jeremy