Guide-yakabatana mibvunzo · Pakati Nhanho

Synthetic Image Detection Quiz

Yedza detector scope, nyowani-jenareta uye recompression shift, calibration, base-rate maalarm enhema, Content Credentials, uye humbowo wongororo.

Nzira dzekutungamira dzinoenderanaKuziva Mifananidzo Yekugadzirwa
Mubvunzo 1 ye 8

A detector marks a photo as probably synthetic. Which conclusion is outside that score alone?

Tarisa mubvunzo wese8 mibvunzo

Mhinduro uye tsananguro zvinoonekwa paunenge uchitora bvunzo.

  1. A detector marks a photo as probably synthetic. Which conclusion is outside that score alone?
    • Whether the image’s depicted claim is true
    • That a model produced a score
    • That further investigation is possible
    • That errors can occur
  2. Why test on screenshots and images from newer generators?
    • Any detector always generalizes to new generators
    • Screenshots prove an image is genuine
    • Compression cannot change model features
    • Their artifacts may differ from the detector’s training data
  3. A raw detector score is 0.9. What is needed before reading it as a 90% probability?
    • A more colorful UI
    • Calibration on relevant held-out data and threshold context
    • A JPEG extension
    • A missing content credential
  4. In the guide’s invented 1,010-image example, 8 synthetic and 10 genuine images are flagged. What fraction of flags are synthetic?
    • Eight of ten
    • Ten of eighteen
    • Eight of eighteen
    • One percent exactly
  5. An image lacks Content Credentials. What follows?
    • It must be AI-generated
    • It must be an unedited camera original
    • It has no source history anywhere
    • Nothing decisive about whether it is synthetic or authentic
  6. A valid provenance record traces an image file. What still needs checking?
    • Whether the visual claim is true in context
    • Nothing; credentials prove every caption
    • Only whether the file is large
    • Only whether a detector also agrees
  7. Why avoid a public accusation from one unvalidated detector flag?
    • Every flag is legally binding
    • The detector cannot be wrong on genuine images
    • False positives can harm real people, especially at low prevalence
    • The image becomes synthetic when flagged
  8. Which evidence should accompany a consequential detection review?
    • Only the displayed confidence number
    • Original files, source history, tool version and independent corroboration
    • Only a cropped thumbnail
    • Only the detector’s marketing description