AI Artistry: Exploring DeepMind and Google Cloud's Invisible Image Watermarking Collaboration

0:00The first thing to note here is that I really think this is a collaboration between Google Cloud and DeepMind. Now, a lot of people are like, how is the collaboration? It's all just Google, right? DeepMind is a company Google owns. Google Cloud is a department Google owns. And what I think a lot of people don't understand is just how massive Google is and how these different entities are like essentially different companies. Pretty much Google Cloud works on very different things. DeepMind is its own sort of segmented company. And, you know, this really was brought to my attention a little bit earlier this year when I believe it was Google Cloud made a$300 million investment into Anthropic.

0:35and people are like, oh, look, Google invested into Anthropic. Even though, you know, they haven't barred, that means they're worried that Google Barred is going to be a failure. This is before Google Barred had come out. And the rest of the Google team and over at Google Ventures are like, hey, just so everyone knows, we're a completely different organization than Google Cloud. Google Ventures did not invest into Anthropic. That was more like a cloud play, right? They wanted to get their cloud compute. It was a heavily funded AI company. Anyways, they knew it was going to be doing some compute.

1:01So Google Cloud wanted to snap some of that up, not miss out on it. But, you know, they were like assuring everyone, don't worry, we have high hopes, we think BART is going to be great still. So that was kind of funny to me way back then to see, and it's kind of playing out again today. So Google's specialized AI research division is, of course, DeepMind that was acquired. It was a UK-based firm before they acquired it. And they're creating essentially a new tool for watermarking AI-generated images, and it's called SynthID. So the tool is in beta and accessible to a select group of users on Vertex AI, Google's platform dedicated to building AI applications and models.

1:38And interestingly, this tool is, you know, tailored exclusively by Imagine, Google's text image model that is only available on Vertex AI. So Synth ID is designed to essentially embed a digital watermark directly into an image's pixels. Although the watermark is essentially invisible to the human eye, it can be detected by specialized algorithms. this move goes beyond google's earlier announcement about embedding metadata to essentially identify visual media um created by ai see the problem with the metadata is it's fairly easy to get around uh metadata right you can take a screenshot of the image now all the metadata is gone you can you know download the file and change the metadata yourself there's all sorts of things you can do to change metadata um but this goes a step further when by you know embedding it in the pixels generating something in the pixels that can be detected by algorithms.

2:30Essentially, it's very, very difficult to make it so that's undetectable. And it's going to be interesting to see if Google does similar, or if other AI image generating platforms do similar things. Like imagine if MidJourney did this, it would be very, very difficult to ever get that out. Now, the thing I can think of is if you grab one of these images that has the kind of like secret embedded watermark in the pixels and you ran it through another AI platform, because there's a lot of platforms where you can like upload one image and make like a similar one. That would be one way to get around to this.

3:04However, if the big ones like Mid Journey start doing it, then really if you want one of the top of the line images that's generated, it's gonna have that embedded. So this is very interesting. And I think that might be the solution to a lot of people's kind of concern about like how do you actually identify some of this stuff, images. They're already doing this to some degree with text, although OpenAI is not very transparent about how they do that. And they also shut down their program to detect AI-generated text. So that one is a little bit, I don't know too many details on that, but this is very interesting.

3:35So DeepMind kind of elaborated on the importance of this development in a blog post and they said, quote, while generative AI can unlock huge creative potential, it also presents new risks like enabling creators to spread false information, both intentionally or unintentionally. Being able to generate AI, Being able to identify AI-generated content is critical to empowering people with knowledge of when they're interacting with generated media and for helping prevent the spread of misinformation. So I think what really makes SynthID noteworthy is its resilience. This was developed by DeepMind, and it was further redefined in collaboration with Google Research.

4:14And the tool's watermark remains detectable even after various image manipulations, like applying filters or altering colors, right? So you can't just throw this thing in Photoshop and, you know, adjust the colors on it or in Lightroom or something and expect that this will go away. So the tool essentially uses a pair of AI models trained together on a diverse data set for both watermarking and identification. That said, I will also say, though, Synth ID is not like perfect. While it can't, you know, guarantee 100 % detection of watermarked images, it does distinguish between images that might contain a watermark and those that are highly likely to.

4:51So I think that's so interesting, right? Like we're so used to watermarks like a photographer or you know like I stock or someone just stamping a massive watermark on the image You can't get away from it in this case It's like so undetectable to like this might have a watermark or like this image has a high probability of having a watermark It's so interesting, right? It's not just like a yes or no detection So and I'm sure that will get better in the future But in terms of future applications deep mind says quote synth ID isn't foolproof against extreme image manipulations, but it does provide a promising technical approach for empowering people and organizations to work with AI generated content responsibly.

5:29This cool tool could also evolve alongside other AI models and modalities beyond imagery, such as audio, video and text, right? So this could this concept could be applied to everything. And I think that's what's really interesting about this is not necessarily like this technology is perfect today, although it's probably fairly decent. It's the concept of like hidden embedded watermarks in all types of AI AI-generated content, eventually I'm assuming whether regulation or self-regulation, this is going to be included in all of the corporate models. And then you're probably going to get a bunch of open source models to get around that.

6:03I don't want my images to have AI watermarks, so I'm going to use this open source model. There's definitely going to be that whole underlying area as well. So I think the concept of watermarking in generative art is not very revolutionary. So for instance, French startup Imiteg, which launched in 2020, offers a watermarking tool resilient to resizing, cropping, and editing. And this is actually similar to SynthID. Another firm, Steg.ai, employs an AI model specifically to apply resilient watermarks. So the push for watermarking AI-generated content, I think, has gained a lot of traction. China's cyberspace administration recently mandated that all generative AI, including text and images, be marked clearly.

6:48And additionally, during recent US Senate committee hearings, Senator Christian Siena emphasized the importance of kind of transparency and generative AI, particularly through watermarking. So people are talking about this in legislation in the United States and China and other places. Several other AI tech giants and startups are also making strides in watermarking. Microsoft committed to watermarking AI generated images and videos using cryptographic methods, um shutterstock and also mid journey have adopted guidelines to indicate ai generated content clearly but i would also say you know like a universal standard for watermarking is still not a real thing in today's day and age currently synth id is um you know confined to working with google's imaging and isn't universally applicable deep mind is considering making synth id available for third-party use soon, although the adoption by open source AI image generators would often, you know, lack the protective features of commercial platforms.

7:49And so that kind of remains an open question. In any case, I think this is a really important topic, whether that's for images, text, or video, watermarking and knowing if an image is genuine or AI, I think is going to be a big thing that we're going to grapple with in the future.