New Study Proves AI Watermarks Carry Little Significance And Can Be Easily Removed

A classic watermark is one that’s visible on logos and patterns of all kinds which tend to pop up on things ranging from pictures and currency to postage stamps and beyond.

So much takes place in today’s world and the goal of experts and technology has always been to evade such means of counterfeiting. But wait, the concept is a little different when it comes down to the era of AI technology.

It is a twist of a whole new kind and thanks to researchers who took part in a new study, we’re getting some more exciting insights regarding this front.

When you think of AI, watermarks give computers the chance to alert users if any font or pictures were created via AI means. Now you’re probably wondering why would anyone opt to watermark pictures in the first place, right?

As far as AI is concerned, this kind of content gives rise to a new type of breeding ground that produces deep fakes and more kinds of misinformation. Hence, it’s definitely invisible to the regular eye but felt that they still go the extra mile in terms of reducing the incidents of the technology misusing content produced through AI means.

Moreover, this might be included in some cases in various types of machine learning ordeals that are made via big tech platforms such as Google.

On the other hand, top players in the market including Meta and ChatGPT’s OpenAI also vowed to work on the right type of watermarking technology to prevent the spread of inaccurate information that’s become commonly published online.

This is one of the many reasons why leading researchers hailing from UMD made it a point to study how simple it was for bad actors to include or get rid of watermarks.

One professor also mentioned how the findings of the study proved that you won’t be finding any type of watermarking apps that are reliable at this point in time.

The authors said their tests proved how simple it was to evade watermarks and how easy it was to include fake logos in pictures that were not designed using AI technology.

Beyond such tests, another interesting aspect did come into play including how the UMD team produced a watermark that’s incredibly difficult to get rid of, until and unless you plan to compromise the property itself.

Such an app opens up new horizons and gives rise to numerous possibilities by which stolen products may be detected. Meanwhile, another similar research conducted in the US saw researchers conclude how simple it was to carry out watermark removal attacks.

Two major means of getting rid of watermarks were highlighted, where one was productive while the other was destructive. In terms of the latter, bad actors may end up treating the emblem in a manner that makes it appear like it’s a part of the picture.

It was shocking yet amazing to see how changing small things like hue, brightness, rotating, and even JPEG compression enable watermark removal with sheer ease. But the leading catch in this situation has to do with how such methods also end up playing with the picture’s quality and ensuring it’s worse than before.

In other attacks designated as constructive, the process of removing the watermark gets a little tricky and is definitely more sensitive. It makes use of a technique that’s similar to a Gaussian blur.

Yes, more studies are surely required to better hold these results as true and reliable. Still, it’s not hard to imagine a scenario where such kinds of watermarking truly turns into a new type of competition conducted against the likes of threat actors.

Photo: FreePik
Read next: 40% of Teens Are Using ChatGPT According to This Survey
Previous Post Next Post