Data and Training Pipeline Attacks COASP Practice Questions (Page 10)

Questions 46–48

1. 46application · medium

   A security team is implementing defenses against data poisoning for a model that uses a public dataset. They want to detect if an attacker has inserted backdoor triggers into the training data. Which technique is most appropriate?

   • ADifferential privacy on the model output
   • BAnomaly detection on training samples
   • CModel inversion to reconstruct triggers
   • DRobust aggregation of model updates
   Reveal Solution Discussion Select an answer first
2. 47application · medium

   A hospital uses an AI model to prioritize patients for emergency care. An attacker poisons the training data so that patients with a certain insurance provider are systematically assigned lower priority. What is the most significant real-world impact of this attack?

   • AReduced model accuracy across all patients
   • BIncreased computational cost of retraining
   • CComplete denial of service of the AI system
   • DLoss of trust and potential harm to a specific patient group
   Reveal Solution Discussion Select an answer first
3. 48application · medium

   A team is building a sentiment analysis model for product reviews. An attacker gains access to the labeling tool and changes the sentiment label of a subset of reviews from 'positive' to 'negative'. The model now incorrectly classifies many positive reviews as negative. Which attack is this, and what is the primary impact?

   • ALabel flipping; targeted misclassification of positive reviews
   • BBackdoor poisoning; the trigger is the review text
   • CData augmentation poisoning; the trigger is the label
   • DAvailability poisoning; overall model performance degrades
   Reveal Solution Discussion Select an answer first