I’m a professional software developer with ML experience, albeit not an expert in ML specifically. It would obviously affect the literal value of the embeddings, but there’s no chance it would have a qualitative effect on a reasonably performant model.
It would though and their paper shows as much. The thing many forget is that it isn’t trained visually like us. Little input changes like this have a big impact.
Now eventually if everyone uses the same glazing method the training won’t care but at the moment this is bespoke enough that it can’t be trained well on it. It will always be an arms race though.
No, it wouldn’t, and the paper shows no such thing. Nightshade isn’t “Gaussian blur + sharpen.” It’s based on the use of a different diffusion model to perturb an image (with bounded difference in perceptual similarity) to minimize the distance of the embedding from that of an unrelated concept. It is mathematically optimized and highly specific to the prompt. The clever thing is that you don’t need access to the actual original text-to-image feature extractor because of the transferability between models, and the surprising thing is how few poisoned samples are required to break a model.
Blur+Sharpen isn’t what Nightshade is doing, it’s an example of a passive defense technique that may mess up fine-tuned “invisible” attacks because they rely on making minimal changes to jump category, and that can often come in the form of pretty precise pixel changes. You may have seen past papers about making pandas classify as gibbons. They rely on introducing a noise mask that just makes the image look a little worse quality, but in total is enough to flip the category. They don’t really define their perturbation method in this paper, but there’s some tension between being “invisible” and being resilient to “invisible” corrections like suggested above.
I’m a professional software developer with ML experience, albeit not an expert in ML specifically. It would obviously affect the literal value of the embeddings, but there’s no chance it would have a qualitative effect on a reasonably performant model.
It would though and their paper shows as much. The thing many forget is that it isn’t trained visually like us. Little input changes like this have a big impact.
Now eventually if everyone uses the same glazing method the training won’t care but at the moment this is bespoke enough that it can’t be trained well on it. It will always be an arms race though.
No, it wouldn’t, and the paper shows no such thing. Nightshade isn’t “Gaussian blur + sharpen.” It’s based on the use of a different diffusion model to perturb an image (with bounded difference in perceptual similarity) to minimize the distance of the embedding from that of an unrelated concept. It is mathematically optimized and highly specific to the prompt. The clever thing is that you don’t need access to the actual original text-to-image feature extractor because of the transferability between models, and the surprising thing is how few poisoned samples are required to break a model.
Blur+Sharpen isn’t what Nightshade is doing, it’s an example of a passive defense technique that may mess up fine-tuned “invisible” attacks because they rely on making minimal changes to jump category, and that can often come in the form of pretty precise pixel changes. You may have seen past papers about making pandas classify as gibbons. They rely on introducing a noise mask that just makes the image look a little worse quality, but in total is enough to flip the category. They don’t really define their perturbation method in this paper, but there’s some tension between being “invisible” and being resilient to “invisible” corrections like suggested above.
Oh, blur+sharpen to mitigate Nightshade makes sense, yeah.