Securing Digital Content: Steg.AI's Invisible Watermarking

0:00Claiming ownership of digital content has become a bit of a complex issue especially with new technologies like AI you know creating content and also digital collectibles right you have the nfts in any case creators often kind of struggle to put some sort of mark or watermark on a creation because really there's obviously the problem of widespread online theft of images videos and pretty much all other content right getting passed off as your own and then of course there's a problem with ai of all the content you've ever created getting sucked into the giant vacuum that is ai models and being trained and then of course being able to spit out something else based off of what you created so existing ways of marking content as your own can be really easily fooled by something you know as simple as just changing the file type of the image or something like that so in response to this whole kind of arena of problems that people are solving or having with this steg.ai has developed a new method for hiding invisible but very strong marks in content to protect it from theft.

1:04So the company which has been developing a deep learning approach to watermarking uses AI to embed an almost invisible mark on digital assets. And this is, you know, kind of define the usual countermeasures like imagery sizing and resaving. So unlike conventional watermark methods, Steg.ai's technology offers a more resilient and unobtrusive solution to intellectual property protection. And this is what they said, quote, imagine a generative AI company, you know, it creates an image and Steg watermarks it before delivering it to the end user. That's, you know, Eric Wiengrowski, Steg's co-founder and co-author of the 2019 CVPR paper outlining the company's kind of innovative approach.

1:47But he wrote to TechCrunch an email and said, the end user might post the AI generated image on social media. Copies of the deployed image will still contain the steg.ai watermark even if the image is resized, compressed, screenshotted, or has its traditional metadata deleted. So this is what they're kind of trying to do here. At its core, steg.ai leverages a pair of machine learning models that customizes the watermark to the image. So the encoding algorithm identifies optimal points to modify the image in a way that remains imperceptible to human eyes but can be easily detected by the decoding algorithm so it's kind of described as a largely immutable or invisible qr code and the watermark essentially processes specific and you know the watermarking process specifics remain kind of under wraps right they don't explain exactly how they do this but this is actually really cool and for those um that have seen this it reminds me of um a while back there's a bunch of projects that come out where essentially they have these really cool pieces of artwork that are built around QR codes.

2:52So it's like, you know, a Japanese village, but there's like a QR code embedded in the image that you can scan. And a lot of people are making these like really cool kind of artistic QR codes. This is interesting because this goes a step further where essentially they're embedding a QR code into the image, but it's imperceptible the way they build it in. So it's a special kind of QR code that can be scanned. Now I am going to, and I could be completely wrong here. I'm think this is a really cool use case. I think this is a really cool technology, but I'm going to play the devil's advocate here for a second and tell you how I believe you could get around this.

3:24Now, if every major image generating AI starts doing this, then there's probably not a chance. Well, actually there is with open source, but in any case, here's how you could get around this, right? So you get, you get, you know, Steg AI to generate you an awesome picture. It's got the Steg AI watermark and you're gonna say you want to like be um you know you want to be controversial and say this is a real image you found a picture of the the pope smoking weed and you're gonna do a big pope smokes weed expose article say it's real right it's it's fake so what you do you get steg ai to generate this thing then you take this thing over to mid journey you upload it and then you just say you know recreate like regenerate this image essentially but make it so that the pope has like a green jacket right so of course mid journey goes takes the image it recreates it but it's slightly different i'm pretty sure doing that would kill the stag watermark in there it just imperceptible changes that it it creates would probably kill that qr code or that watermark um and you know inevitably if you ran it through a few times you maybe you yeah if you run it through this thing different ais a few times you get that and the ai is still generating a really realistic thing now if someone like mid journey or dolly 2 actually which in my opinion what's more likely is that steg.ai is going to implement with or integrate with other image generators or those image generators will create something similar right the same idea concept technology if steg.ai or mid journey if it's essentially integrated into all image generators and it's like well now no matter where you go there's always going to be some sort of embedded watermark essentially that's invisible but um the caveat there is that there are a lot of open source image generators right um stability ai is working on a bunch there's a number and essentially you could use your own now that being said that's not to say that even the open source ones won't eventually embed uh similar technology into their projects if this became a really big thing that everybody wanted maybe they would embed it and if you wanted the best image generators the best open source image generators you're going to have to like they would still have that and so maybe it's not like a centralized company that's verifying that you know how everything was created but maybe they just have like a website where you can put put an image in there it's maybe called watermark checker ai watermark checker dot ai or dot com whatever right you get the idea and you can imagine a company where you essentially could upload an image and it would scan that image and maybe it would run it by steg.ai, Dolly2, MidJourney, Stability.ai, and it would check and tell you, yes, there's a watermark from this AI company on there, so this thing was AI generated.

6:08That would be incredibly interesting. Now, some, you know, like, I think some people are saying that, well, steg.ai essentially is saying that even if you crop it or resize it or whatever, I'm not sure if they said crop it, but they said if you resize it, that it will get rid of it won't get rid of the watermark but i'm not sure about cropping because i'm not sure if this thing is a watermark that covers they're not super clear but they're not saying i'm not sure if there's a watermark that covers every inch all over it or if it's a watermark that's like embedded in part of it that you could crop around for example right and if you could crop around it then of course the solution would be take something that has a steg watermark somewhere on it um and then you could say you know you you'd use essentially like an automated fill maybe to automatically fill all around the image and then I don't know expand it and then try to crop a part that didn't have the watermark from before I don't know there's workarounds um so I do think this is a really cool technology but I wouldn't I would just say you still you know you could still put this thing in all of the AI checkers and it says nope it's perfectly human made and it could still be AI generated so this whole thought experiment of how to get around this is just to shake your confidence in the fact that it's actually not always possible to detect AI, and it is a very complex issue.

7:27But I do give them some props for tackling this problem. I think this is an important issue, and it is really interesting. So I think what is interesting here is while the exact data capacities of the watermark is undisclosed, if it resembles a QR code, I think it could potentially hold a few kilobytes, which is enough for URL hash and other plain text data. So each page of a document or frames in a video could contain unique codes amplifying this capacity. Several applications can, I think, benefit from this technology, including, of course, stock photography providers, Instagram creators, movie studios, and also businesses needing to, you know, mark confidential documents.

8:07Steg.ai's ability to provide really robust and some subtle watermarks, I think, presents quite an attractive, you know, it is pretty attractive for those entities. but I will say that the company's journey is not without its challenges. So initially they actually discovered that many of their product ideas were not hitting the mark and they soon recognized the value of robustness. So a critical differentiator in their technology since then is that the company has found traction among entities where leaked information could stir consumer interests. You know we're talking about consumer electronic brands for example.

8:41So So Steg.ai's offering is packaged as an enterprise level SaaS integration integrated seamlessly into a digital asset management platform. So this solution essentially ensures that all media is marked and tracked as part of the regular handling process. And that really kind of is eliminating the need for manual watermarking. So I think in addition to tracing images back to their source, the technology could also potentially detect changes made to an image over time. So the app or API could also offer a confidence level indicating that an image hasn't been manipulated. And this is a feature that could be, I think, pretty invaluable to editorial photography managers.

9:22But yeah, I just think kind of in a climate where tech firms are really kind of exploring ways to watermark AI content. I think that Steg.ai's solution could be a pretty critical measure. They did secure$1.2 million in NSF grants and some angel investments. And I think the company has recently announced a$5 million seed round, which was led by Paladin Capital Group, which I think really is going to be very interesting to see how they use that to grow in the space. So definitely a company that will be very interesting to follow in the future.