• QueenMidna@lemmy.ca
    link
    fedilink
    English
    arrow-up
    12
    ·
    5 days ago

    Highly unlikely that any amount of significant processing is happening on-device. Wouldn’t be cost effective

    • CheeseNoodle@lemmy.world
      link
      fedilink
      English
      arrow-up
      16
      arrow-down
      1
      ·
      5 days ago

      You could piss off the company running them though, don’t these cameras use AI? Get some QR codes that lead to nepenthies traps full of their own AI generated images of nonsense license plates. Waste their money and poison their training data.

      • QueenMidna@lemmy.ca
        link
        fedilink
        English
        arrow-up
        1
        ·
        4 days ago

        Image element recognition doesn’t magically make actions happen, such as following a QR code link. The on-board AI would be more along the lines of pattern recognition such as reading text, license plates, people counts, etc. You could poison it by putting up the license plates but unless someone is actively looking for those plates it’ll just be a small bit of metadata attached to the time code. The fact that they’re stationary means it’ll probably be ignored as well

        • CheeseNoodle@lemmy.world
          link
          fedilink
          English
          arrow-up
          1
          ·
          4 days ago

          Image recognition AI has been found to be vulnerable to prompt injection via text present in the image.

          • QueenMidna@lemmy.ca
            link
            fedilink
            English
            arrow-up
            2
            ·
            3 days ago

            For LLMs, sure, but there’s no way they’re running LLMs on-device. Just classic AI models at the most