Thanks to artificial intelligence, removing a watermark from a photo has never been easier

In the era of generalization of artificial intelligence, based on the watermark (watermark in English) protecting a photo is no longer such a good idea. This is an observation implemented by a quality site called WatermarkRemover.io, which allows you to remove a watermark from an image without losing quality. Artificial intelligence developed by PixelBin.io, a specialist in image hosting and distribution on the Internet, can specifically identify tags attached to photos sold online. And the result can be amazing, as this example presents PetaPixel and this is based on the stock image.

This innovation worries photographers who may depend on these watermarks to protect their photos before a sale or distribution deal. Stock image platforms rely heavily on these digital signatures to protect their catalogs from abuse. But with the generalization of such tools, these protections will no longer exist, and it will become easier and easier to remove watermarks from photos. To render them completely useless? It’s not that simple.

These smart AIs are actually very stupid. Watermark removers are probably the result of machine learning (machine learning) based on the comparison: photos with and without watermark. Artificial intelligence has learned to go from one to another with these patterns, to realistically reconstruct what is “underneath” the signs. The example above proves that it works very well on stock photo watermarks, but you can also quickly get the AI ​​wrong.

As an example, I was thinking about a solution for automating adding a watermark to a copy of an ID card. The context is the numerous scams found on the internet, particularly classified ad sites where crooks collect copies of identity documents to steal identities. I know this scenario well because I was cheated on two years ago:

One scam on the Internet can hide another

One piece of advice I got after this article was to never hand out a copy of your ID without noting it to clearly link to the purpose. By writing the date, site, or recipient’s name on the copy of the document, it can remain legible, but not so easily redirected if you encounter a bad actor.

Online Fraud: How to Protect and Respond to Identity Theft

But is this advice still valid when faced with such tools? Just to be sure, I tested it by starting with a popular site that should remove the photo of a fake government-issued ID, the Shortcuts mark in the middle, and the watermarks. As you can see, the result is not as satisfactory as the stock photo above. The AI ​​removed some elements, but the original ID watermarks and some text that shouldn’t have been touched disappeared. The message I wrote in big red letters remained intact.

This makes perfect sense, AI has never been trained on before/after images with this kind of coarser markup. Not to mention the lack of transparency in the text, which would have allowed him to better “guess” how to reconstruct the original image. Are these watermarks safe? Not really, because one can imagine various machine learnings that could eventually lead to a pretty effective tool for erasing letters without touching the bottom. Even if it is very easy to fool it again by choosing a different font or, for example, writing manually with a pen on a tablet.

This cat-and-mouse game is already impossible in the long run, which is probably why the French government is working on a better option. With the new ID card that integrates a chip and France Identity in particular, it will be possible to create one-time proofs of identity that prevent the transfer of a copy of the card. In any case, the promise is that we will have to wait for all the infrastructure to be in place and for these supporting documents to have legal value, which is not yet true today.

France Identity, the application that will have a card to play in everyday life

Alpha testing of the France Identity app on iOS: a la carte access

Going back to watermarks in photos, there is no denying that tools that can remove them cleanly will proliferate in the coming years. They will certainly be upgraded to handle more and more work, so we can assume that this method of protecting images on the web will lose all interest, and probably soon.