Bypassing Defensive Controls GOAA Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A red team is conducting an exercise against a biometric authentication system that uses both fingerprint and voice recognition. The system requires both modalities to match for authentication. The team has successfully created a synthetic fingerprint that fools the fingerprint model and a voice recording that fools the voice model individually. However, when they present both together, authentication fails. Which factor is most likely causing the failure?

   • AThe system requires a higher confidence threshold when both modalities are used.
   • BThe system uses a fusion model that checks the consistency between the two modalities, and the synthetic samples are not correlated.
   • CThe synthetic fingerprint and voice recording were not presented simultaneously.
   • DThe fingerprint and voice models are not vulnerable to adversarial examples.
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   A red team is conducting an exercise against a voice-based authentication system that uses a deep learning model to verify speakers. The team has a recording of a legitimate user's voice. They want to test whether the system can be bypassed using a replay attack. Which test would be most aligned with a systematic red-team exercise?

   • ASubmit the recording file directly to the authentication API, bypassing the microphone.
   • BUse a text-to-speech engine to generate a synthetic voice that mimics the user's voice.
   • CAlter the pitch of the recording to see if the model still recognizes the speaker.
   • DPlay the recording through a high-quality speaker to the microphone and observe if the system accepts it.
   Reveal Solution Discussion Select an answer first
4. 23foundation · easy

   Which of the following is a common mitigation for adversarial input attacks on AI models?

   • ARemoving all input validation to avoid interfering with the model
   • BDisabling the AI model and using only human review
   • CAdversarial training, which includes adversarial examples in the training data to make the model more robust
   • DUsing a more complex model with no regularization
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   What is the primary purpose of generating adversarial inputs in the context of AI-driven offensive operations?

   • ATo improve the training accuracy of the target AI model
   • BTo validate the model's robustness against random noise
   • CTo increase the computational efficiency of the AI model
   • DTo cause the AI model to misclassify or malfunction, thereby evading defensive controls
   Reveal Solution Discussion Select an answer first
6. 25expert · hard

   A red team is testing an AI-powered network intrusion detection system (NIDS) that uses a deep learning model to classify network flows. The team has access to the model's training data and can train a surrogate model. However, the NIDS is deployed in a live environment and the team has a limited number of queries before detection. Which strategy is most effective to craft an evasion while minimizing queries?

   • AUse a pure query-based attack that iteratively refines the adversarial flow using the NIDS's responses.
   • BTrain a surrogate model on the same data and use transferability to craft an adversarial flow, then validate with a few queries.
   • CUse a precomputed universal adversarial perturbation that is known to fool all NIDS models.
   • DRandomly mutate network flow features until the NIDS does not flag the flow.
   Reveal Solution Discussion Select an answer first