Pre-presentation quiz

What does "perturbation" mean in the context of adversarial examples?

• A large, random modification that changes the meaning of the input

• A small, intentional change made to an input to mislead a model

• A type of data used to improve model training

• A method to visualize model performance

What is the main purpose of applying a perturbation to an input?

• To make the input more realistic

• To improve the accuracy of the model

• To make the model produce an incorrect output

• To reduce overfitting

What are vulnerabilities in AI systems?

• Simple programming errors or syntax mistakes

• Structural weaknesses that cause incorrect or exploitable outputs

• Network latency problems in AI hardware

• Overfitting issues during training

Why do overfitting, memorization, and brittleness make AI “vulnerable”?

• Because people don’t use Wi-Fi

• Because AI wants to rest

• Because the AI doesn’t actually understand — it just copies patterns

• Because AI hates learning

Do AI vulnerabilities exist at the intersection of model architecture, training data quality, and deployment context?

• Yes

• No

These are characteristics of LLM hallucinations except:

• Generating fluent and coherent text

• Producing factually incorrect or fabricated content

• Always caused by software or coding bugs

Are deepfakes and hallucinations different because of human involvement?

• Yes

• No

Deepfakes does not involve human manipulation

• Yes

• No

Models purposely lie when they hallucinate, generating factually incorrect outputs through probabilistic mechanisms without intentional deception

• Yes

• No

What is intrinsic hallucination?

• Output that is intentionally misleading by human manipulation

• Output that contains spelling or grammar errors

• Output that is fluent and coherent

• Output that contradicts the source content