Comment by jsheard
3 years ago
Is it possible to reliably detect whether an image is poisoned? If not then it achieves the goal of punishing entities which indiscriminately harvest data.
3 years ago
Is it possible to reliably detect whether an image is poisoned? If not then it achieves the goal of punishing entities which indiscriminately harvest data.
You can use older images, collected from before the "poisoning" software was released. Then you don't have to.
This, of course, assumes that "poisoning" actually works. Glaze and Nightshade and similar are very much akin to the various documented attacks on facial recognition systems. The attack does not exploit some fundamental flaw in how the systems work, but specific characteristics in a given implementation and version.
This matters because it means that later versions and models will inevitably not have the same vulnerabilities. The result is that any given defensive transformation should be expected to be only narrowly effective.
AI's have learned much tougher things. You just need a small data set of poisoned images to learn it's features.
It's roughly in the same spot as reliably detecting if you have permission to use the image for your data training set in the first place.
If it doesn't matter, then neither does the poisoning matter.