Bypassing Defensive Controls GOAA Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   An AI-powered facial recognition system is used to control access to a secure facility. A penetration tester wants to bypass the system by presenting a printed photo of an authorized user. The system uses a deep learning model to verify identity. Which type of attack is the tester attempting?

   • AA presentation attack, which is a physical spoofing attempt that does not involve modifying the model's input data.
   • BAn adversarial example attack, which involves adding imperceptible perturbations to the photo.
   • CA membership inference attack, which determines whether a face is in the training set.
   • DA model inversion attack, which reconstructs the authorized user's face from the model's outputs.
   Reveal Solution Discussion Select an answer first
2. 7expert · hard

   An organization has deployed an AI-based spam filter that is vulnerable to adversarial examples. The security team is evaluating two countermeasures: adversarial training and defensive distillation. The team has a limited budget and must choose one. Which factor is most important in this decision?

   • ADefensive distillation is always more effective than adversarial training and requires less data.
   • BAdversarial training is less computationally expensive than defensive distillation.
   • CDefensive distillation is the only countermeasure that does not require retraining the model.
   • DAdversarial training directly addresses the specific attack vector by incorporating adversarial examples into training, while defensive distillation may not be effective against all attack types.
   Reveal Solution Discussion Select an answer first
4. 8application · medium

   A security team is conducting a red-team exercise against an AI-powered endpoint protection platform that uses a deep learning model to detect malicious files. The team wants to test whether the model can be evaded by adversarial examples. They have access to the model's architecture and weights. Which attack technique is most appropriate to use?

   • AFast Gradient Sign Method (FGSM) to generate adversarial examples using the model's gradients.
   • BTransfer attack using a surrogate model trained on a public dataset.
   • CRandom noise injection to the file without using model information.
   • DZeroth-order optimization to estimate gradients through query feedback.
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   What is the key idea behind transferability in model evasion attacks?

   • AAn adversarial example crafted for one model is likely to fool another model trained on similar data
   • BAn attack only works if the attacker has full access to the target model's architecture
   • CA model's weights are transferred to another model to improve its accuracy
   • DThe attacker transfers the target model's training data to a different location
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A security team is testing an AI-powered malware detector that analyzes API call sequences. The team wants to generate a malware sample that evades detection while preserving its malicious functionality. They have white-box access to the detector's model. Which technique is most appropriate?

   • AEncrypt the malware binary so the detector cannot analyze the API calls.
   • BUse a gradient-based attack to modify the API call sequence in a way that minimizes the model's confidence in the malware class.
   • CAdd random API calls to the sequence to confuse the model.
   • DRemove all API calls that are commonly associated with malware.
   Reveal Solution Discussion Select an answer first