Glaze - new method to protect images from AI learning?

[Mich-666]

So apparently, there is this new paper (non peer-reviewed) from Chicago University that aims to disrupt the AI training by protecting image with additional layer that should theoretically confuse AI learning algorithm.

https://arxiv.org/pdf/2302.04222.pdf

What do you think about it? Is this the same nonsense as before or can this potentially work?

The original site of the project here: http://glaze.cs.uchicago.edu/

[vladmandic]

adversarial attacks at nothing new when it comes to ml.

[zrichz]

it's an interesting idea, but the images on page 11 of the paper show that significant noise in the "cloaked" images will currently be a limiting factor in artists posting cloaked versions of their art. I imagine simply adding a little gaussian noise will remove a lot of this protection, but interesting to see how this develops :)

[Mich-666]

So the download link is up now, they just released it. Wonder what all the fuss was about.
https://glaze.cs.uchicago.edu/download.html

[pkuliyi2015]

Can we remove the protection by simply pass the image through the VAE？

[catboxanon]

That might've been what was done here to defeat it (original parent tweet was deleted unfortuantely so I'm posting the reply): https://twitter.com/JordanCMeyer/status/1636367657694789632

[spaghettiSystems]

The authors test gaussian noise and jpeg and they don't defeat their approach.
A VAE that preserves image details likely wouldn't discard the extra style information this adds. Unfortuantely their public release is only for windows and mac so I can't test this.

[5KilosOfCheese]

Here is a thought.
What stops me from basically doing what we did to NFT bros? Just basically do the equivalent of taking a screenshot of the image. Quick reading of that, this only applies to using the original image file for training - not the media by itself. Basically this is a really fancy watermark calculated specific to the artwork. (And by that logic, if you know how to make one, you know how to decode one).

So if you take that 4000x2000 resolution picture, shrink it down to size you can manage training and which is still manageable; like 1000x500px all that glaze is going to be lost. Look I think artists should have right to somehow opt-out or protect themselves, but this ain't a way to achieve it. I have used basic machine vision systems where, where certain patterns were used to force the system to ignore something and to calibrate the systems. The system primarily ran within the smart camera units, central unit only relaying commands according to their reports, and saving the images.

To me this seems like something one could simply bypass: by them mentioned screencap method; fiddling with resolution; or simply crushing or flattening the image by adjusting it's properties. Narrow the range of information it has so only get the desired features.

However:
"Second, a system like Glaze that protects artists faces an inherent challenge of being future-proof. Any technique we use to cloak artworks today might be overcome by a future countermeasure, possibly rendering previously protected art vulnerable. While we are under no illusion that Glaze will remain future-proof in the long run, we believe it is an important and necessary first step towards artist-centric protection tools to resist invasive AI mimicry. We hope that Glaze and followup projects will provide some protection to artists while longer term (legal, regulatory) efforts take hold." (Shan, Shawn et al. GLAZE, 2023).

[SmxMullah]

I think it is pretty difficult to future these kind on technologies since they are evolving at a very high pace , this can be future proofed is there is a way to update the photo after it is posted too . But we dont have the means to it yet . I think only other possible way might be legally enforcing these AI companies to use filters from their side .

[catboxanon]

It's quite ironic this utilizes SD to even work. I think the paper's wording tries to mask this, the website certainly seems to at least. How would artists feel if they found out that the tool they're using is directly built on top of what they're trying to fight against? The application downloads the SD weights directly to a user's machine. Clearly AI is only good if there's a benefit for them.

Also, IANAL, but violating the GPL license in this way would require them to release the full source, no? It's all part of one single binary executable. https://twitter.com/ravenben/status/1636439335569375238

[BaaMeows]

"Clearly AI is only good if there's a benefit for them" its. their. art.
if you used public domain images for your models it would be fine.
are you dense??
why am I even here I need a walk slkejlsfekjeslf

[Mich-666]

They took down the downloads, that was quick.

The code for their frontend is still available here: https://glaze.cs.uchicago.edu/software/glaze-fontend.zip

[mfarhadhossain]

@Mich-666 Do you have the codebase? I need for r&d purposes.

[Mich-666]

@Mich-666 Do you have the codebase? I need for r&d purposes.

You mean from the link above? It's only frontend but here it is:

https://limewire.com/d/hqKZq#RgwCwuMqDs

[FullBleed]

I never thought there would be a market for artists to actively push for their work to be ignored and to have their influence diminished.

I seriously want my ad blocker to have an option to block "art" that incorporates these techniques. If they truly want to have less influence then I'm more than happy to oblige them.

[chobeat]

they want money and recognition, not influence. Try to steal from a shop and then say: "Ok, keep your goods, I thought you wanted some influence on the things we consume but apparently you want to be ignored"

[huzpsb]

https://github.com/huzpsb/DeTox/
AFAIC it can be filtered out with ease...