AI System Hacking Methodologies, Frameworks, and Risk Implications COASP Practice Questions (Page 6)

Questions 26–30

1. 26expert · hard

   A security consultant is conducting an offensive assessment of an AI-based medical diagnosis system. The consultant has completed the reconnaissance phase and is now in the weaponization phase. The consultant has identified that the model is a convolutional neural network (CNN) and that the system accepts image inputs. The consultant wants to craft an attack that will cause the model to misdiagnose a benign tumor as malignant. Which approach is MOST appropriate for the weaponization phase?

   • ACreate an adversarial image by adding a small perturbation to a benign tumor image that is imperceptible to humans but causes the model to output a malignant classification.
   • BPoison the training data by injecting malignant images into the model's training set.
   • CDirectly send the benign tumor image to the model's API to see if it is misclassified.
   • DPerform model inversion to reconstruct the training data and find a benign image that is misclassified.
   Reveal Solution Discussion Select an answer first
2. 27expert · hard

   A red team is conducting an assessment of an AI-powered autonomous vehicle. The team has identified that the vehicle's object detection model is a deep neural network. During the exploitation phase, the team wants to cause the vehicle to misidentify a stop sign as a yield sign. Which technique is MOST appropriate for this phase?

   • APlace a physical adversarial patch on the stop sign that is designed to fool the model.
   • BPoison the training data by adding images of stop signs labeled as yield signs.
   • CSteal the model to understand its weaknesses, then craft an adversarial image.
   • DPerform model inversion to reconstruct the training data and find a stop sign image that is misclassified.
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A security analyst is documenting an attack on an AI-based facial recognition system. The attacker used a technique that involves wearing specially designed glasses to fool the system into recognizing the attacker as a different person. The analyst wants to classify this attack using MITRE ATLAS. Which technique is this?

   • AData Poisoning, because the attacker is poisoning the training data by wearing the glasses.
   • BModel Inversion, because the attacker is inverting the model to find the glasses design.
   • CAdversarial Input Manipulation, because the glasses are a physical adversarial input.
   • DModel Stealing, because the attacker is stealing the model's ability to recognize faces.
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A financial institution is considering whether to deploy an AI model for loan approvals. The compliance team is concerned about the risk of model theft, where an attacker could replicate the model by querying it. Which business risk is MOST directly associated with model theft?

   • AImmediate system downtime because the attacker will overload the API.
   • BLoss of competitive advantage because the proprietary model's logic can be replicated and used by competitors.
   • CIncreased regulatory fines because the model's training data is exposed.
   • DHigher operational costs because the model will require more frequent retraining.
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   A red team is using MITRE ATLAS to document an attack on a company's AI-powered supply chain management system. The attack involved compromising a third-party data provider to inject malicious data into the training pipeline. The red team needs to identify the most relevant ATLAS tactic and technique. Which mapping is correct?

   • ATactic: Exfiltration; Technique: Exfiltration Over Web Service
   • BTactic: Execution; Technique: User Execution
   • CTactic: Initial Access; Technique: Supply Chain Compromise
   • DTactic: ML Attack Staging; Technique: Poison Training Data
   Reveal Solution Discussion Select an answer first