Adversarial Machine Learning Techniques COASP Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A researcher wants to generate adversarial examples for a text classification model (e.g., sentiment analysis). Which perturbation strategy is most appropriate?

   • AReplace words with synonyms that preserve the sentence's meaning but change the model's prediction.
   • BReverse the order of words in the sentence.
   • CAdd random punctuation at the end of the sentence.
   • DRemove all stop words from the sentence.
   Reveal Solution Discussion Select an answer first
2. 27foundation · easy

   Which of the following is a defensive technique used to mitigate adversarial attacks?

   • AInput sanitization to filter or preprocess inputs before feeding them to the model.
   • BUsing a larger batch size during inference.
   • CDeploying the model on more powerful hardware.
   • DIncreasing the model's learning rate during training.
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A company uses an AI model to detect fraudulent insurance claims. They have a high false-positive rate, and customers are complaining. They also face adversarial attacks that craft claims to appear legitimate. Which strategy best balances reducing false positives and defending against adversarial claims?

   • ADisable the model and use manual review for all claims.
   • BIncrease the model's complexity to better distinguish legitimate and fraudulent claims.
   • CLower the decision threshold to reduce false positives, and rely on manual review for suspicious claims.
   • DUse adversarial training to make the model more robust, and adjust the threshold based on a validation set.
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A security team is threat-modeling an AI-based email spam filter. They identify that an attacker can submit specially crafted emails that are classified as legitimate, bypassing the filter. Which attack vector does this scenario represent?

   • AModel extraction
   • BMembership inference
   • CEvasion
   • DData poisoning
   Reveal Solution Discussion Select an answer first
6. 30foundation · easy

   What is the main objective of threat modeling for adversarial attacks on AI systems?

   • ATo guarantee that the AI system is completely immune to all attacks.
   • BTo automatically patch vulnerabilities without human intervention.
   • CTo increase the model's accuracy on adversarial examples.
   • DTo identify and categorize potential attack vectors and assess their potential impact.
   Reveal Solution Discussion Select an answer first