A team of Google researchers succeeded in developing an algorithm to remove the watermark from non-royalty-free photos. A breach of security easily overshadowed by their demonstration.
Researchers from the American giant Google have looked at a security flaw affecting photographs of image banks, such as Getty images or Shutterstock among others and more precisely on the apparent watermarks used to protect them and thus avoid use Not authorized. The story is told in detail in their explanatory article "On the Effectiveness of Visible Watermarks" presented at a conference in Hawaii last July. Tali Dekel and Michael Rubinstein expose with two other confreres the lack of reliability of images a priori protected by a visible watermark. The two authors also specified the results of their research in an article published on the Google blog.
An algorithm and many images
The team thus discovered that the watermark method did not make images unusable. It has even gone further by generating a successful algorithm to remove them. For this purpose, they have realized a software capable of identifying several models of watermarks, thanks to a consequent number of public examples found on the Internet. Using a process, he then succeeds in separating the part of the watermark from the rest of the image. It is because of the details that make up the watermark such as its opacity, shape and shadow, or the effects of its color gradation, that the algorithm is able to remove the watermark or similar from any photo Which contains it.
• Read also: A penis appears on the home page of Bing, Microsoft uses Photoshop
A temporary remedy
But the researchers did not remain there and proposed a temporary solution to solve this problem. Instead of changing the location of the watermark or its opacity, which the algorithm can bypass, you need to add specific random elements to it. Rubinstein and Dekel point out that it is necessary to make changes incorporating "random geometric changes to the watermark", that is, deforming the text or logo used to protect the image. So, when an algorithm like that of Google is accustomed to locate constant elements, it then no longer manages to remove those that differ from the original.
This alternative defense is still imperfect and more sophisticated algorithms can circumvent it. In the meantime, this proposal is always more convincing than the current ones and allows for a time, to protect a little better the image bank companies.