Using a watermark is a common practice to make sure an image created or clicked by someone is not used without the prior permission of the creator. This is a standard practice which prevents the images from being used without proper licensing, however, it is not as robust as previously thought. Google has recently presented and published a research titled, “Making Visible Watermarks More Effective” which portrays how watermarks can easily be removed using AI and how to circumvent this problem.
Image credits: Google research
One may think that removing a watermark from an image entirely would be quite hard, however, given ample amount of images with a similar watermark, an AI can quite easily analyse the watermark’s lines and curves and implement an algorithm accordingly to remove it. Google has published their research with the intent of exposing this vulnerability and also patching it effectively.
The research admits that the usage of thin lines and shadows makes removal of a watermark difficult albeit given a single image. Using the same watermark over a span of images, reveals the consistent manner being used to apply it. Utilizing this characteristic, a multi-image optimization problem is employed to find the watermark’s decomposition in the image and use this to segregate the watermark (foreground) from a clean subset of the rest of the images (background). Thus, any image containing an analyzed watermark can be rid free of it. Even changing the position or opacity of the watermark has no effect. This spells bad news for stock image providers and curators but Google has also described a method of circumventing this problem.
Gif credits: Google reasearch
To sidestep the problem of effective watermark removal, Google employs a technique called ‘Warping.’ As the vulnerability being exploited by Google is based on the consistency of watermarks, to counter the problem, inconsistencies are introduced while layering the watermarks onto an image. Warping a watermark slightly will cause problems in detecting them and substantially hinder the working of a watermark removing algorithm. This is due to the fact that the algorithm removing the watermark will also need to estimate the warp field being applied to the watermarks for that image which is quite a difficult task.
Image credits: Google research
The research remarks that many other types of randomizations can be used to make watermark removal less effective, however, warping is one of the easiest methods to employ. The research concludes by saying that there is no guarantee that such randomized methods will not be bypassed in the near future, however, randomization will make removal of watermarks more effective.