Researchers from Google have released a new method for watermarking images in an effort to prevent computer algorithms from removing the watermark. Within the last few months, these scientists studied the possibility that a computer algorithm, or even machine learning, may be used to remove watermarks. After all, they were able to develop a software that could remove objects from photographs back in 2015. Thus, researchers employed by the search giant, attempted to design a program that will detect logos or any similar features that are used as watermarks and then isolate the actual image from the file. In order to get a clean version of the picture, the program will find image gradients or structures that are consistent across different files. After several stages of optimization, a watermark pattern will then be recovered, which can be used in a wide range of photo-editing software to remove this feature from photographs.
These researchers warn concerned parties, especially those working in stock photography, that cases of image piracy may substantially rise if certain individuals with malicious intent design an algorithm with a similar function. It is not enough that the position and opacity of the watermark are changed across the pictures since the software can still detect the patterns. Thus, they recommend that photographers utilize warping when applying watermarks. Warping refers to subtle changes made to the image structure, like modifications in the geometric formation of the watermark, which may not be noticeable to the human eye. These modifications may fool the algorithm used by content pirates since AI relies heavily on studying and recognizing patterns in order to perform its function.
The results obtained in the preliminary tests were promising. While a substantial portion of the watermark was still removed by the algorithm, it left visible artifacts indicating that the material was possibly copyrighted by somebody else. It also ruins the quality of the picture, which should prevent people from reselling the image. All of these were achieved despite the lack of noticeable differences between the photograph that used warping and those that did not. Researchers hope that it will make future attempts in stealing images from the stock photography collections more difficult in the near future.