Quiz dotyczący zapamiętywania modeli dyfuzyjnych

Odpowiedzi i wyjaśnienia pojawiają się w miarę rozwiązywania quizu.

1. Czym jest zapamiętywanie w modelu dyfuzyjnym?
   • The model remembering a user's previous prompts
   • The model reproducing specific training images almost exactly
   • The model storing images in a database for search
   • The model forgetting how to draw after training
2. Dlaczego zduplikowane dane zwiększają zapamiętywanie?
   • Duplicates are stored in higher resolution
   • Duplicates are removed automatically before training
   • When an image appears many times, storing it becomes the easiest way to reduce training loss on it
   • Duplicates confuse the model so it outputs noise
3. Co wykazał artykuł Carliniego i współpracowników z 2023 r.?
   • Diffusion models never copy training data
   • Only text models memorize data
   • Memorization only happens in black and white images
   • Over a hundred near-identical training images could be extracted from Stable Diffusion
4. W jaki sposób atak ekstrakcji pozwolił zidentyfikować prawdopodobnie zapamiętane obrazy?
   • It looked for near-identical generations across many seeds for the same caption
   • It read the model's weights directly as images
   • It asked the model which images it had memorized
   • It searched the web for similar photos
5. Które stwierdzenie dotyczące zapamiętywania jest prawdziwe?
   • It is the model's normal behavior for every prompt
   • It is rare but can be targeted by a determined user
   • It only occurs as exact pixel copies
   • It has no relevance to privacy
6. Why is privacy risk sharper when fine-tuning on a small dataset such as medical scans?
   • Small datasets are always public
   • Medical scans are too large to train on
   • Each example has more influence on the model, raising the chance of regurgitation
   • Fine-tuning deletes the base model
7. Which mitigation bounds the influence of any single training example but currently reduces image quality?
   • Differentially private training
   • Increasing the number of sampling steps
   • Using longer prompts
   • Training for more epochs
8. Why is deduplication more effective when it removes near-duplicates, not only exact copies?
   • Exact copies do not affect training
   • Near-duplicates are illegal to store
   • Near-duplicates make images blurrier
   • Slightly edited versions of the same image still act as repeated exposure