3 ms·
Hi, one of the founders here, "robustness" is our next step. We are currently training a few models based on cross attention and invariant domain learning that
by matyask 2y ago
Hi, one of the founders here, "robustness" is our next step. We are currently training a few models based on cross attention and invariant domain learning that promise to be quite resistant to noising and denoising. The purpose of fingerprint watermarking as we have released it is to act like a signature for validity, some companies already do this but none of them have public products that you can just use. The nest encoder and decoder will use a unique signature for every user. The limitation of Glaze and models like it is that they target specific models/approaches. That being said we are also running experiments on adversarial watermarking that aim to deliver the same image quality as glaze with less computation and more resistance. We want to secure digital identity and copyrights. Watermarking is our initial approach.
- mrbungie 2y agoI'm still trying to understand how a watermarking method may resist simple methods as (1) upscaling and downscaling plus noise, or (2) taking a photo of the image plus compression plus noise, (3) any simple convolutional filter or a combination of those.
- matyask 2y ago1) With an invariant domain representation, 2) and 3) I am not certain yet, but I will find out. Although when it comes to signing with a watermark for validity you would not want the id (50 bits) to be preserved or copied. These are separate use cases / applications.
- janalsncm 2y agoGlaze techniques might be specific to a model but the techniques to defeat it are not. More specifically, I don’t see any reason to believe that I cannot learn a function from A to B where A is the distribution of watermarked files and B is the distribution of non-watermarked files. It’s a trivial exploit.