Data Poisoning CT-AI Practice Questions (Page 4)

Questions 16–20

1. 16application · easy

   A security analyst is explaining data poisoning risks to a non-technical stakeholder. The stakeholder asks, 'What is the primary goal of a data poisoning attack?' Which response is most accurate?

   • ATo increase the model's accuracy on legitimate data
   • BTo reduce the computational cost of training the model
   • CTo manipulate the training data so that the model behaves in a way that benefits the attacker
   • DTo steal the training data and sell it to competitors
   Reveal Solution Discussion Select an answer first
2. 17application · medium

   A team is training a model to classify images of animals. An attacker injects a small number of images of a cat with a specific red dot in the corner, all labeled as 'dog'. After training, the model classifies any image with a red dot in the corner as 'dog', regardless of the animal. Which type of attack is this?

   • AData injection, because the attacker added new images to the dataset
   • BBackdoor attack, because the red dot is a trigger that causes misclassification
   • CLabel flipping, because the cat images were labeled as 'dog'
   • DData sanitization, because the attacker cleaned the data
   Reveal Solution Discussion Select an answer first
4. 18expert · medium

   A company is deploying an AI system for fraud detection. They have a strict requirement: the model must not have a false negative rate above 1% for a specific high-value transaction type. During testing, they find that a data poisoning attack could increase the false negative rate to 5% for that transaction type. Which of the following is the most appropriate response?

   • ADeploy the model and monitor the false negative rate in production, adjusting the threshold if needed
   • BDeploy the model but only for low-value transactions, keeping the high-value transactions manual
   • CReject the deployment until the poisoning risk is mitigated and the model is re-tested to meet the requirement
   • DDeploy the model because the attack is hypothetical and may not occur in practice
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   A team is using a validation set to check for data poisoning. They notice that the model's accuracy on the validation set is high, but it performs poorly on a small, manually curated test set. What does this discrepancy most likely indicate?

   • AThe model is overfitting to the training data, which is a normal and expected behavior
   • BThe validation set is too small, so the accuracy estimate is unreliable
   • CThe test set is too difficult, so the model's low performance is expected
   • DThe validation set may have been poisoned or is not representative, while the curated test set reveals the true model weakness
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   How can data poisoning compromise the integrity of an AI system?

   • ABy making the model's predictions more accurate
   • BBy causing the model to produce incorrect outputs that appear valid
   • CBy increasing the model's training time significantly
   • DBy reducing the model's memory footprint
   Reveal Solution Discussion Select an answer first