Kvíz spojený s průvodcem · Střední úroveň

Neviditelné vodoznaky pro obrázky AI a zvukový kvíz

Testuje, jak jsou neviditelné vodoznaky jako SynthID a AudioSeal vložené, jak je detektory hodnotí a které útoky je mohou odstranit.

Otázka 1 z 8

Jak se liší neviditelné obrazové a zvukové vodoznaky od textového vodoznaku, se kterým je průvodce kontrastuje?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. Jak se liší neviditelné obrazové a zvukové vodoznaky od textového vodoznaku, se kterým je průvodce kontrastuje?
    • They are stamped as a visible logo in one corner of the content
    • They add imperceptible changes to pixel or sample values instead of steering word choice
    • They rely entirely on file metadata fields that travel alongside the content
    • They require viewers to install a special codec before the content will play
  2. Která organizace vyvinula AudioSeal?
    • Google DeepMind
    • OpenAI
    • Meta
    • Adobe
  3. Detektor nenajde v obrázku žádný vodoznak. Co můžete správně uzavřít?
    • Only that no mark from that system was found; the image could still be AI-generated
    • That the image is proven to be a genuine photograph
    • That the image was edited in a photo editor after it was taken
    • That the watermark has moved into the file's metadata
  4. Který útok průvodce popisuje jako schopný oslabit nebo odstranit vodoznaky obrázku?
    • Viewing the image on a phone screen
    • Renaming the image file
    • Changing the file's creation date
    • Regenerating the image through a diffusion model
  5. Kde skrývá akademická metoda Tree-Ring svůj vodoznak?
    • In a fine-tuned text encoder
    • In the model's starting noise
    • In a hidden EXIF tag
    • In a faint visible grid over the image
  6. What is the purpose of the 'attack layer' used when training a learned watermarking system?
    • To speed up inference on consumer hardware
    • To encrypt the payload so only the owner can read it
    • To simulate edits such as compression and cropping so the decoder learns to recover the mark anyway
    • To make the watermark visible for debugging
  7. How do C2PA content credentials differ from invisible watermarks?
    • They are embedded directly in audio samples
    • They are produced by the diffusion model's starting noise
    • They cannot be removed under any circumstances
    • They are signed metadata that can be lost when a file is re-saved
  8. Why do providers often restrict who can query their watermark detectors?
    • An attacker with detector access can optimize perturbations until the mark is no longer detected
    • Detectors are too large to distribute
    • Laws forbid sharing any detection software
    • Detectors only work without an internet connection