Warped watermarks leave visible artifacts when removed automatically. Image: Google

Watermarks are widely used by photographers and stock agencies to protect their digital property in an online world where very little stands between an eager image thief and your photography. However, even complex watermarks might not be as secure as you'd think when the same pattern is applied to a large number of accessible images.

A research team at Google recently embarked on a two-part experiment. First, they developed a method of quickly,. effectively and automatically removing watermarks from a large set of images. Then, they found a way to thwart their own automatic system, creating a more secure way to watermark.

Automatic Watermark Removal

It's a tedious task to remove a watermark manually, which can can take even image editing experts several minutes. Even for a computer it is very difficult to automatically detect and remove a watermark on a single image. However, if watermarks are added in a consistent manner to many images, automatic removal becomes much easier.

In the first step, an algorithm identifies which image structures are repeating in an image collection. If a similar watermark is embedded in many images, the watermark becomes the signal and images become the noise. At that point, a few simple image operations can generate a rough estimation of the watermark pattern.

In the second step, the watermark is separated into its image and opacity components while reconstructing a subset of clean images. The end result is a much more accurate estimation of the watermark pattern, which can then easily be removed from the marked images—no manual photo editing required.

Making More Secure Watermarks

As the vulnerability of current watermarking techniques lies in the consistency in watermarks across image collections, the research team at Google developed a method to introduce inconsistencies when embedding watermarks.

They found that simply changing the watermark position or random changes in its opacity do not improve security by much; however, slightly warping the watermark when embedding it in the image did the trick by producing a watermark that is very similar to the original but leaves very visible artifacts when removed by an algorithm. Estimating the random warp field that was applied to the watermark is simply too difficult for current algorithms.

According to the researchers, there is no guarantee that there will not be a way to break randomized watermarking schemes in the future, but randomized warping will make it fundamentally more difficult to automatically remove watermarks from image collections.

More detail and sample images are available on the Google Research Blog.