Cuestionario sobre protección del artista de glaseado y belladona

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

1. ¿Qué amenaza está diseñada principalmente para contrarrestar Glaze?
   • Theft of physical paintings
   • Style mimicry through fine-tuning a model on an artist's work
   • Screenshots of artwork on social media
   • Deletion of artwork from websites
2. ¿Cómo difiere Nightshade de Glaze en su objetivo?
   • Nightshade watermarks images for copyright registration
   • Nightshade encrypts images so nobody can view them
   • Nightshade poisons training data so models learn wrong associations between concepts and images
   • Nightshade blocks web crawlers from a website
3. ¿Cuál es el objetivo de optimización de Glaze al perturbar una obra de arte?
   • A version of the artwork converted into a different style
   • A blank white image
   • A random image from the internet
   • The artist's signature
4. ¿Qué evita que las perturbaciones sean evidentes para los espectadores humanos?
   • The tools only change image metadata
   • The changes are applied only to the image border
   • The tools reduce image resolution
   • A perceptual constraint, such as an LPIPS budget, limits visible change
5. ¿Qué encontró un estudio de 2024 de ETH Zurich y Google DeepMind?
   • Glaze makes images impossible to train on under any conditions
   • Simple methods like noisy upscaling could remove Glaze-style protections enough for mimicry
   • Nightshade works only on printed art
   • Perturbations become stronger over time
6. Why is protection from these tools not permanent?
   • A published image cannot be updated while attack methods and models keep changing
   • The software licenses expire after a year
   • The perturbations fade from the file over time
   • Models are required by law to ignore protected images
7. What does transferability mean in the context of these tools?
   • Moving an image from one website to another
   • Transferring copyright to the tool's developers
   • Converting an image between file formats
   • A perturbation optimized against one encoder still working on other models
8. Why did the Nightshade researchers argue a modest number of poisoned images could affect a model?
   • Because every model is trained on only 100 images
   • Because poisoned images are copied automatically across the internet
   • Because each concept appears in relatively few training images compared with the whole dataset
   • Because models trust poisoned images more than clean ones