Model Extraction & IP Protection Quiz

1. Which defense technique adds random noise to model outputs?

Model quantization

Output perturbation

Feature scaling

Batch normalization

2. True or False: Keeping a model completely private is always the best defense against extraction.

True

False

Submit

4. Which defense is most effective against model extraction via API queries?

Returning detailed confidence scores

Implementing strict prediction thresholding and rate limiting

Using smaller models

Increasing training data size

Submit

6. What is a backdoor attack in machine learning?

Inserting hidden triggers into a model to cause misclassification on specific inputs

Stealing the training dataset

Deleting model parameters

Reducing model accuracy

Submit

8. Which of these is a white-box attack?

Attacker has no knowledge of the model

Attacker has full access to model weights and architecture

Attacker can only see final predictions

Attacker has limited query budget

9. What is model fingerprinting?

Creating a unique signature of a model to detect theft

Removing all identifying information from a model

Encrypting model parameters

Compressing model size

Submit

11. What is model extraction?

Stealing a trained model by querying it and reverse-engineering its behavior

Removing features from a dataset before training

Deleting model weights after deployment

Extracting training data from model logs

12. What is an adversarial perturbation?

A small, intentional noise added to inputs to cause misclassification

A large change in model weights

A deletion of training samples

A modification of the loss function

Submit

14. How can differential privacy help defend against model extraction?

By encrypting all user queries

By adding noise to predictions to obscure model behavior

By removing all features from data

By deleting the model after each query

15. What is the primary goal of membership inference attacks?

To steal the model's source code

To determine if a specific sample was in the training set

To modify model predictions

To delete the model

16. Which of the following is NOT a typical model extraction threat?

Functional stealing via repeated API queries

Hyperparameter inference

Training data augmentation

Architecture reverse-engineering

17. API rate limiting serves what purpose in model protection?

It prevents legitimate users from accessing the model

It reduces the speed at which attackers can query the model

It improves model accuracy

It compresses model size

18. What is a key defense against model extraction?

Returning only class labels without confidence scores

Increasing model complexity

Using larger datasets

Training longer

19. How do prediction confidence scores aid model extraction attacks?

They reveal nothing about the model

They provide training labels for surrogate models

They encrypt the model weights

They prevent overfitting

20. Which attack method involves training a surrogate model by observing the target model's predictions?

Poisoning attack

Knockoff nets / substitute model attack

Adversarial perturbation

Gradient inversion

SecAI+ Model Extraction and Theft Prevention Quiz

1. Which defense technique adds random noise to model outputs?

2.

What first name or nickname would you like us to use?

2. True or False: Keeping a model completely private is always the best defense against extraction.

3. Model distillation can be exploited for ____.

4. Which defense is most effective against model extraction via API queries?

5. Poisoning attacks corrupt the ____ to compromise model performance.

6. What is a backdoor attack in machine learning?

7. Black-box attacks rely on ____.

8. Which of these is a white-box attack?

9. What is model fingerprinting?

10. Watermarking in models serves to ____.

11. What is model extraction?

12. What is an adversarial perturbation?

13. Adversarial examples are inputs designed to ____.

14. How can differential privacy help defend against model extraction?

15. What is the primary goal of membership inference attacks?

16. Which of the following is NOT a typical model extraction threat?

17. API rate limiting serves what purpose in model protection?

18. What is a key defense against model extraction?

19. How do prediction confidence scores aid model extraction attacks?

20. Which attack method involves training a surrogate model by observing the target model's predictions?