AI Security Testing, Evaluation, and Hardening COASP Practice Questions (Page 10)

Questions 46–48

1. 46expert · hard

   A team is hardening a spam-filtering model. They have two options: (1) adversarial training with a small set of known spam perturbations, or (2) input sanitization that strips HTML tags and normalizes text. The model currently has high accuracy on clean emails but fails on adversarial spam. The team has limited compute for retraining. Which approach is more effective given the constraints?

   • AInput sanitization, because it addresses the attack vector without requiring retraining
   • BNeither, since the model's clean accuracy is already high
   • CBoth, but prioritize adversarial training since sanitization can be bypassed
   • DAdversarial training, because it permanently improves the model's robustness
   Reveal Solution Discussion Select an answer first
2. 47application · medium

   A cybersecurity team is evaluating the robustness of a malware-detection model. They have generated adversarial examples using several different attack methods and want to compare the model's performance under each attack. The team needs to present the results to management in a clear and actionable way. Which metric or visualization would be most effective for this purpose?

   • AA table showing the attack success rate (ASR) for each attack method, along with the average perturbation size required.
   • BA bar chart of the model's training time and inference latency, showing its efficiency.
   • CA confusion matrix showing the model's true positives, false positives, true negatives, and false negatives on a clean test set.
   • DA single accuracy score for the model on a clean test set, indicating its overall performance.
   Reveal Solution Discussion Select an answer first
3. 48application · medium

   A team discovers that an AI model used for resume screening is biased against certain demographic groups due to poisoned training data. The team wants to harden the model against future poisoning attacks. Which measure is MOST effective?

   • ADeploy the model immediately since the bias is not severe
   • BIncrease the model's complexity to better fit the training data
   • CImplement data provenance tracking and validation checks for all training data sources
   • DReduce the training data size to minimize the impact of poisoning
   Reveal Solution Discussion Select an answer first