Google Research Finds Way to Automatically Erase Watermarks on Photos, Also Offers Solution

Google recently demonstrated research that makes the process of identifying and removing watermarks from images automatic, which would obviously introduce a huge problem for stock agencies. Luckily, they also showed an easy solution for protecting the images.

The exploit relies on the consistency of the process: whereas a single image would prove difficult to analyze, a library of images in which the same watermark is applied in the same fashion allows a computer to quickly find what feature is consistent across all images, i.e. the watermark:

google-watermark-removal 

So, the exploit relies on two things: having a large library of images to analyze and that library having a consistent watermark across its images; in other words, large stock providers are most at risk. It turns out the solution is rather simple, however: by introducing small, random warping of the watermark each time it's applied, one can break the consistency the algorithm relies on, thereby causing it to leave artifacts when applied to such an image:

google-watermark-removal-2

As you can see, the artifacts appear wherever the watermark was warped from the average (or base) version. While you likely don't need to worry about this exploit being applied to your own images, it's good to see that the research also produced more robust methods for protecting photos.

[via Gizmodo]

Posted In: 
Log in or register to post comments

6 Comments

Justin Berrington's picture

Their protection method isn't all that great when 90% of the watermark is still removed in a pretty clean manner. With their exploit you could still get the majority of the work done with a script and then do a lot less manual work than you previously would have needed to remove it. So yes, a thief wouldn't be able to batch process a full library but they could certainly speed up the work

Alex Cooke's picture

Sure, but I think the point was to make it so it still required manual intervention, so the state of affairs hasn't shifted that much.

Pawel Paoro Witkowski's picture

I guess it would be simpler to add some random scaling/translating by few pixels - this will automatically make things so much more difficult for the algotirhm

Spy Black's picture

The bigger question of course is why spend time and money on research and development of such technology in the first place? What is the end goal and benefit for Google here?

If you know how it can be broken, then you know how to fix it.... and patent the fix.

Spy Black's picture

...except there's more than one way to skin a cat. Method X may not be of use for method Y and Z.