Adversarial Machine Learning Techniques COASP Practice Questions (Page 7)

Questions 31–35

1. 31application · medium

   A company is deploying an AI-based spam filter. They want to make it robust against adversarial emails that insert special characters or extra spaces to evade detection. Which defensive technique is most directly suited to this threat?

   • ARestricting the API to only allow authenticated users.
   • BIncreasing the model's complexity to capture more patterns.
   • CAdversarial training with perturbed email examples.
   • DUsing a larger training dataset of benign emails.
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A security researcher is evaluating a malware detection model. They want to generate malware samples that evade detection while preserving the malware's functionality. Which technique is most appropriate?

   • AEncrypt the malware payload.
   • BUse a gradient-based attack to add small perturbations to the malware's feature vector.
   • CRecompile the malware with a different compiler.
   • DAppend random bytes to the malware file.
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A security team is hardening a computer vision model against adversarial patches (physical stickers placed in the scene). Which defensive approach is most likely to be effective?

   • ARestricting the camera's field of view.
   • BAdversarial training that includes images with simulated patches.
   • CApplying a Gaussian blur to all input images.
   • DUsing a larger input image size.
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A security researcher is testing a malware detection model that uses API call sequences as features. They want to generate adversarial malware that evades detection while preserving the malware's behavior. Which approach is most appropriate?

   • AInserting benign API calls into the sequence to confuse the model.
   • BRemoving all API calls from the malware.
   • CEncrypting the malware payload.
   • DChanging the malware's file extension.
   Reveal Solution Discussion Select an answer first
6. 35expert · hard · select all that apply

   A company is deploying an AI model for autonomous driving. They are concerned about adversarial attacks on the perception system. Which of the following defensive techniques are appropriate to implement? Select all that apply.

   • AInput sanitization to filter out unusual sensor readings.
   • BRegularly updating the model with new attack patterns from a red team.
   • CAdversarial training with simulated physical-world perturbations.
   • DUsing multiple sensor modalities (camera, LiDAR, radar) and cross-validating.
   • EIncreasing the model's inference speed.
   Check answers Discussion Select an answer first