AI and Machine Learning Fundamentals from an Offensive Security Perspective COASP Practice Questions (Page 7)

Questions 31–35

1. 31expert · hard

   A security team is assessing the risk of an AI system that uses a model to detect fraudulent transactions. The team has identified that the model is trained on historical transaction data that includes a small number of fraudulent examples. An attacker has been able to craft transactions that are misclassified as legitimate. Which combination of factors is most likely contributing to the success of the attacker's evasion attacks?

   • AThe model's decision boundary is not robust, and the attacker has knowledge of the model's behavior
   • BThe training data is imbalanced, and the attacker is using a data poisoning attack
   • CThe model is overfitted, and the attacker is using a model extraction attack
   • DThe model's API is rate-limited, and the attacker is using a denial-of-service attack
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A penetration tester is conducting an engagement against a company that uses a machine learning model to filter spam emails. The tester has already performed reconnaissance and identified the model's input features (email headers, content, and sender reputation). The tester now wants to craft an email that bypasses the filter without disrupting the email server. Which step of the hacking methodology does this activity primarily represent?

   • AExploitation — crafting an adversarial email to evade the model's classification.
   • BPersistence — maintaining access to the model's environment.
   • CPrivilege escalation — gaining higher access to the model's infrastructure.
   • DReconnaissance — gathering more information about the model's training data.
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A security engineer is reviewing a machine learning system that uses a supervised classifier to filter spam emails. The engineer wants to understand how an attacker could cause the classifier to misclassify a specific phishing email as legitimate without altering the training data. Which attack characteristic is the engineer most likely considering?

   • APerforming a denial-of-service attack on the model's API
   • BExtracting the model's training data through query responses
   • CCrafting adversarial examples that exploit the model's decision boundary
   • DInjecting malicious samples into the training set
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A security team is defending a machine learning model that classifies images. They have implemented adversarial training, but an attacker has still managed to craft successful evasion attacks. The team suspects the attacker has access to the model's architecture and weights. Which type of attack is the attacker most likely performing?

   • ABlack-box evasion attack
   • BWhite-box evasion attack
   • CModel inversion attack
   • DData poisoning attack
   Reveal Solution Discussion Select an answer first
6. 35application · easy

   A security analyst is evaluating the risk of a machine learning model that predicts credit default. The analyst is concerned that an attacker could exploit the model to learn whether a specific individual is in the training dataset. Which type of attack is the analyst most concerned about?

   • AData poisoning attack — corrupting the training data.
   • BEvasion attack — crafting inputs to cause misclassification.
   • CModel extraction attack — stealing the model's parameters.
   • DMembership inference attack — determining if a specific record was part of the training data.
   Reveal Solution Discussion Select an answer first