AI Attack Taxonomies and Classification Models COASP Practice Questions (Page 1)

Questions 1–5

1. 1foundation · easy

   In the AI kill chain, which stage corresponds to the attacker achieving their ultimate objective, such as causing financial loss or reputational damage?

   • ACommand and control
   • BExploitation
   • CImpact
   • DDelivery
   Reveal Solution Discussion Select an answer first
2. 2foundation · easy

   Which adversarial ML attack is characterized by the attacker's goal of recovering private training data or sensitive attributes from a model?

   • AEvasion attack
   • BPoisoning attack
   • CInference attack
   • DModel extraction attack
   Reveal Solution Discussion Select an answer first
4. 3expert · hard

   A penetration tester is using an AI attack taxonomy to plan an assessment of a voice recognition system. The tester has physical access to a device that uses the system but no network access to the backend. The goal is to cause the system to execute a malicious command. Which attack taxonomy-based approach is most viable given the physical access constraint?

   • APerform a network-based poisoning attack by intercepting and modifying the training data
   • BExtract the model by querying the device's API
   • CCraft a physical adversarial audio sample and play it to the device to trigger the malicious command
   • DPerform a membership inference attack to determine if specific audio samples were used in training
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   An offensive AI team is scoping a test for a self-driving car's object-detection model. The team has limited time and must choose which attacks to test first. The model is deployed in a safety-critical environment, and the team has no access to the training data. Using a taxonomy, which prioritization is most defensible?

   • APrioritize inference-time attacks that can be executed with black-box access, given the safety-critical context and lack of training-data access
   • BPrioritize extraction attacks because they are the easiest to execute
   • CPrioritize poisoning attacks because they are the most damaging to safety-critical models
   • DTest all attack categories equally because taxonomy categories are equally important
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A cybersecurity instructor is explaining AI attack taxonomies to a class. The instructor wants to illustrate how different taxonomies organize attacks based on the attacker's goal. Which of the following correctly groups attacks by their primary goal?

   • AEvasion and extraction attacks aim to compromise availability; poisoning and inference attacks aim to compromise integrity
   • BPoisoning and inference attacks aim to compromise confidentiality; evasion and extraction attacks aim to compromise integrity
   • CAll attacks aim to compromise availability, as they all disrupt the AI system's normal operation
   • DEvasion and poisoning attacks aim to compromise integrity; extraction and inference attacks aim to compromise confidentiality
   Reveal Solution Discussion Select an answer first