The point of the article is that, for sites with a huge image corpus, if they apply some randomization to the watermark for every image, it's much more challenging to recover a un-marked image.
Right now, sites like shutterstock and other image sites apply the same watermark to every image, making it quite easy to computationally extract the watermark, and then apply the inverse transform to marked images.
If the watermark is permuted on a per-image basis, it becomes much harder, since you can't extract the watermark from a single image.
Right now, sites like shutterstock and other image sites apply the same watermark to every image, making it quite easy to computationally extract the watermark, and then apply the inverse transform to marked images.
If the watermark is permuted on a per-image basis, it becomes much harder, since you can't extract the watermark from a single image.