On: Making Visible Watermarks More Effective

[jotato]

I wonder what would happen if I took 1000 copies of the same image that is using a warped mark and ran it through the algorithm....

[lightbyte → jotato]

The algorithm is calculating the mean of the set of images, so putting 1000 copies of the same image will do absolutely nothing

[colejohnson66 → lightbyte]

Couldn't I run the result through Photoshop with some "Content Aware Fill" on the remaining spots?

[jotato → lightbyte]

i mean to remove the watermark on that specific image

[azdle → jotato]

The idea is to use a randomly-warped per-image watermark. So in theory, there's exactly one distinct watermark per image, so 1000 of the same image would have the same watermark.

[Kdparker → azdle]

It would be easy to change that hash by modifying one pixel per image (pick one that is unaffected by the watermark).

[azdle → Kdparker]

I'm not sure what you're suggesting. Change what hash? Of the image? I can't think of how that could help with removing a watermark.

[LukeShu → azdle]

jotato: I believe this can be defeated by tricking the watermarker in to generating 1000 distinct watermarks for the same image.

azdle: That won't work, because "there's exactly one distinct watermark per image, so 1000 of the same image would have the same watermark."

Kdparker: So subtle alter the image, to trick the watermarker in to thinking the image isn't the same, thus generating more distinct watermarks for the (effectively) same image.

[tshaddox → LukeShu]

If you have the original image, the problem is already solved. If you don't have the original image, then all you will be doing is submitting the watermarked image and having the site apply another random watermark to it.

[0x0 → jotato]

They would probably use the same warp every time on a specific image, so that you can't average it out across multiple download sessions.

[AgentME → jotato]

The algorithm takes a bunch of pictures, figures out what's in common between them, and then removes that part. If you give it only many copies of the same image, it will find that the whole image is in common between the images, and subtract out the entire content.