A traditional watermark is a clearly visible logo or pattern that can appear on anything from paper money in your wallet to a postage stamp, with the goal of preventing the creation of a counterfeit document.
You may have seen a watermark in the preview of your graduation photos, for example. However, in the case of artificial intelligence, a little twist is required, as is usually the case in this field.
See also: Apple invests millions of dollars every day in artificial intelligence (AI)
In the context of artificial intelligence, watermarking can allow a computer to detect whether text or an image is generated by artificial intelligence. Creative art creates an excellent basis for reproducing fake information and other content. Although watermarks are invisible to the naked eye, they can combat the misuse of AI-generated content and can be integrated into machine learning programs developed by major technology companies, such as Google. Other major companies in the space, from OpenAI to Meta to Amazon, have committed to developing digital watermarking to combat disinformation.
That’s why computer science researchers at the University of Maryland (UMD) set out to examine how easy it is for malicious actors to add or remove watermarks. Soheil Feizi, a professor at UMD, told Wired that his team’s findings confirm his suspicion that there are no reliable applications . The researchers were able to easily bypass current watermarking methods during testing and found that it’s even easier to add fake emblems to images that weren’t created by AI. However, UMD developed a special watermark that is nearly impossible to remove from content without compromising intellectual property. Through this application, product theft is detected.
In a collaborative research effort by the University of California, Santa Barbara and Carnegie Mellon University, researchers have discovered that they can easily remove watermarks through simulated attacks. The paper presents two methods for removing watermarks: a destructive and a constructive approach. In destructive attacks, malicious users can treat watermarks as part of the image. By changing elements such as brightness, contrast, or even using JPEG, we can get rid of a watermark. Even simply rotating the image can make it disappear. The problem is that while these methods remove the watermark, they also affect the quality of the image, making it unrecognizable. In a constructive attack, watermark removal requires sensitivity and uses techniques such as Gaussian blurring.
See also: Paint Cocreator: Released to Insiders with artificial intelligence
Although AI-generated watermarking needs to be improved, we can imagine a scenario where digital watermarking is a competitor to hackers.
With the 2024 US presidential election soon to take center stage, artificial intelligence could play a huge role in influencing political opinion with misleading advertising. The Biden administration has commented on the issue, expressing legitimate concerns about how artificial intelligence can be used for subversive purposes, especially in the area of disinformation .
Source: engadget.com
