• Asifall@lemmy.world
      link
      fedilink
      arrow-up
      20
      arrow-down
      1
      ·
      1 year ago

      Not really, if you read the paper what they’re doing is creating an image that looks like a dog, is labeled as a dog, but is very close to the model’s version of a cat in feature space. This means manual review of the training set won’t help.

        • Asifall@lemmy.world
          link
          fedilink
          arrow-up
          11
          ·
          1 year ago

          I don’t think the idea is to protect specific images, it’s to create enough of these poisoned images that training your model on random free images you pull off the internet becomes risky.

    • SmoothOperator@lemmy.world
      link
      fedilink
      arrow-up
      16
      ·
      1 year ago

      Hmm, sounds more like they are adding structures to the images such that what is clearly a picture of a dog registers as a picture of a cat to an AI. I suppose this can be done by altering the pixels in a way invisible to humans, but visible to AI, adding a cat into the “ghost pixels”.

      • Mirodir@discuss.tchncs.de
        link
        fedilink
        arrow-up
        13
        ·
        edit-2
        1 year ago

        I went and skimmed the paper because I was curious too.

        If my skimming is correct, what they do is similar to adversarial attacks on classifiers, where a second model learns to change as few pixels as possible to confuse a classifier into giving a wrong prediction.

        Looking at the examples of dogs and cats: They find pictures of dogs where by making only minimal changes, invisible to the naked eye, they can get the autoencoder to spit out (almost) the same latent representation as an image of a cat would have. Done to enough dog-images, this will then confuse the underlying diffusion model to produce latent representations of cat images when prompted to generate a dog. Edit for clarity: Those generated latent representations would then decode into cat images.

        If my thinking doesn’t fail me, this attack could easily be thwarted by unfreezing the pretrained autoencoder. In the paper that introduced latent diffusion they write that such approaches already exist. If “Nightshade” takes off, I’m sure those approaches would be refined and used. Even just finetuning the autoencoder for a few epochs first should be enough to move the latent representations of the poisoned dog images and those of the cat images they’re meant to resemble far enough apart to make the attack meaningless.

        Edit: I also wonder how robust this attack is against just adding an imperceptible amount of noise to the poisoned images.