Adversarial Attacks CT-AI Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   What is the main idea behind adversarial training as a defense mechanism?

   • ATraining the model on a dataset that includes adversarial examples so it learns to be robust against them.
   • BEncrypting the model's parameters to prevent attackers from understanding them.
   • CUsing a separate model to detect adversarial inputs before they reach the main model.
   • DRegularizing the model to reduce overfitting on the training data.
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A company is developing an AI-based system for autonomous driving. They are aware that adversarial attacks could cause the system to misidentify traffic signs. Why is it important to consider adversarial attacks during the testing phase?

   • ATo ensure the system is robust against malicious inputs that could cause accidents
   • BTo make the system more interpretable
   • CTo reduce the computational cost of the model
   • DTo improve the system's accuracy on normal inputs
   Reveal Solution Discussion Select an answer first
4. 18application · medium

   A company uses an AI model for email filtering. They want to defend against adversarial attacks that add subtle changes to emails to bypass the filter. Which defense mechanism is most effective for text-based models?

   • AAdversarial training with text perturbations
   • BImage preprocessing techniques
   • CAdding more training data
   • DModel compression
   Reveal Solution Discussion Select an answer first
5. 19foundation · easy

   Which of the following is an example of an attacker capability that would be considered in a threat model?

   • AThe attacker has access to the model's architecture and parameters.
   • BThe attacker has a high level of education in computer science.
   • CThe attacker is located in a different country than the AI system's servers.
   • DThe attacker has a strong motivation to cause financial loss to the company.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   How does Projected Gradient Descent (PGD) differ from the Fast Gradient Sign Method (FGSM)?

   • APGD uses a random initialization of the perturbation, while FGSM starts from the original input.
   • BPGD is a white-box attack, while FGSM is a black-box attack.
   • CPGD generates adversarial examples that are always more imperceptible than FGSM.
   • DPGD is a multi-step iterative method that projects the perturbation back into a valid range after each step.
   Reveal Solution Discussion Select an answer first