0

Is there a watermarking scheme for images and audio that is:

  1. Invisible - should be imperceptible for a human.
  2. Non-invertable - cannot be removed by any means without the original.
  3. Robust - in order to destroy the watermark you have to spoil the image or audio perceptually.
  4. Public - the data embedded in the image or audio can be read by anyone without requiring any information other that watermarked image or audio.
  5. Replicated - the data is replicated all over the file such that only a small chunk of the file is needed to read the watermark with reasonable confidence.
user47376
  • 101
  • 1
  • Since your requirement does not include that the quality and content of the original audio/video is unharmed you can add a clear "property of ..." audio/video content to the file all over the place, i.e. make the watermark clearly visible. Make sure that this must destroy enough of the original information in the original data so that it can not be restored by interpolation or similar - see also [this current research](https://venturebeat.com/2017/08/17/google-researchers-made-an-algorithm-to-delete-watermarks-from-photos/). – Steffen Ullrich Sep 03 '17 at 05:41
  • Your'e right, forgot that requirement. edited. – user47376 Sep 03 '17 at 11:04

1 Answers1

1

Quick answer: not easiliy.

Your requirement #1 is in opposition to the others.

Let's look at a counter example. #1 requires the changes to be imperceptible to humans. That means you cannot add to, take away from, or modify the visible image. For starters, that opens the "analog hole", where a photo of the screen masks the watermark data.

Requirement #3 says it needs to be robust. Developing a scheme to remove or obscure a watermark today is done in two ways. You can compare a watermarked file to an original in hopes of deriving a way to reverse every change made by the watermarking process, allowing the attacker to write a tool to "unsign" the images.

But that's not the only valid attack. The attacker can compare two watermarked files against each other. Remember, the attacker doesn't need to remove the watermark data to be successful. All he has to do is introduce enough doubt that you can't prove if the file came from the image distributed to client A or the image distributed to client B. In a courtroom, the content producer is suing client A for copyright infringement, but he can only collect from A if he can conclusively prove it came from A and not B. (The success rate of this attack grows even further when the attacker incorporates more watermarked copies than just two.)

If I could issue visibly different media to A and B, then the info could not be removed so easily. Let's say it's an animated video where I have a scene with a background shelf cluttered with household stuff. I can render it with a blue teapot in video A, and a red vase in video B, and repeat this concept in as many scenes as necessary to help with #5. (The same concept could perhaps be done with a percussion instrument in a music track.) Any copy will uniquely identify the two, but requirement #1 says they cannot be perceptibly different.

John Deters
  • 33,650
  • 3
  • 57
  • 110
  • 1) The attacker has no access to the original, if he did all of this would be pointless. 2) of course there is a tradeoff between visibility and robustness, but we can have both to a reasonable extent. – user47376 Sep 04 '17 at 00:26