AI Attack Taxonomies and Classification Models COASP Practice Questions (Page 4)

Questions 16–20

1. 16foundation · easy

   In the AI attack lifecycle, which stage involves the attacker gathering information about the target model's architecture, training data, or APIs?

   • AExploitation
   • BWeaponization
   • CImpact
   • DReconnaissance
   Reveal Solution Discussion Select an answer first
2. 17foundation · easy

   An attacker modifies input data at inference time to cause a model to misclassify a sample. This is an example of which type of adversarial machine learning attack?

   • APoisoning attack
   • BModel extraction attack
   • CEvasion attack
   • DInference attack
   Reveal Solution Discussion Select an answer first
4. 18expert · hard

   A red team is planning an attack on a facial-recognition system used for physical access control. The team's objective is to gain unauthorized entry. The team has access to a photo of an authorized employee but cannot modify the model or its training data. Which attack strategy is most likely to achieve the objective, and which lifecycle stage does it primarily represent?

   • APerform an inference attack to reveal the employee's biometric data, representing the exploitation stage
   • BExtract the model to find a universal bypass, representing the reconnaissance stage
   • CPoison the training data with the employee's photo, representing the delivery stage
   • DCraft a physical adversarial patch to place on the employee's photo, representing the weaponization stage
   Reveal Solution Discussion Select an answer first
5. 19expert · hard

   An offensive AI consultant is tasked with assessing a financial trading AI system. The consultant has identified that the system uses a reinforcement learning (RL) agent that interacts with a simulated market environment for training. The consultant has access to the training environment but not the production environment. Which attack taxonomy-based approach should the consultant prioritize to cause the most significant financial impact?

   • APoisoning the training environment to manipulate the RL agent's policy, causing it to make poor trading decisions in production
   • BExtraction attacks to steal the RL agent's policy
   • CInference attacks to determine which market data was used in training
   • DEvasion attacks on the production model by crafting adversarial market data
   Reveal Solution Discussion Select an answer first
6. 20application · medium

   During a red-team engagement, an analyst maps an attack against a computer-vision model used for access control. The team first studies the model's confidence scores on various inputs to understand its weaknesses, then crafts a physical sticker that causes the model to misidentify an authorized person. Which two stages of the AI attack lifecycle does this sequence primarily represent?

   • AExploitation followed by installation
   • BWeaponization followed by delivery
   • CDelivery followed by exploitation
   • DReconnaissance followed by weaponization
   Reveal Solution Discussion Select an answer first