Training Data Poisoning COASP Practice Questions (Page 8)

Questions 36–39

1. 36application · medium

   A security analyst is reviewing a model that classifies images of handwritten digits. The model was trained on a dataset that included a subset of images where the digit '7' was consistently labeled as '1'. The analyst observes that the model misclassifies many '7's as '1's but performs well on other digits. Which statement best describes the integrity compromise in this scenario?

   • AThe model's feature extraction layer is corrupted, causing all digits to be misclassified randomly
   • BThe model's decision boundary for '1' has been widened to include features of '7', reducing accuracy for both classes
   • CThe model's confidence scores for '1' predictions decrease, but the classification accuracy remains stable
   • DThe model's overall accuracy is unaffected because the poisoned samples are a small fraction of the dataset
   Reveal Solution Discussion Select an answer first
2. 37application · medium

   A healthcare organization trains a diagnostic model on medical images. An attacker poisons 5% of the training data by flipping labels on a subset of malignant tumor images to benign. The model's overall accuracy remains high, but the false negative rate for malignant tumors increases significantly. Which impact assessment correctly describes the consequence of this poisoning attack?

   • AThe model's precision for benign cases improves, but recall for malignant cases drops, increasing the risk of missed diagnoses
   • BThe model's specificity for benign cases decreases, causing more false positives and unnecessary biopsies
   • CThe model's overall accuracy decreases proportionally to the 5% poisoned data, making the attack easily detectable
   • DThe model's F1 score for benign cases decreases, while the malignant class remains unaffected
   Reveal Solution Discussion Select an answer first
4. 38application · medium

   A healthcare organization trains a model to predict patient readmission risk. An attacker poisons the training data by flipping labels on a subset of high-risk patients to low-risk. The model's readmission predictions become less accurate for high-risk patients. Which impact assessment correctly describes the consequence of this poisoning attack?

   • AThe model's recall for high-risk patients decreases, leading to missed interventions and increased readmission rates
   • BThe model's specificity for high-risk patients increases, reducing false positives
   • CThe model's F1 score for low-risk patients decreases, while the high-risk class remains unaffected
   • DThe model's precision for low-risk patients increases, but the overall accuracy remains unchanged
   Reveal Solution Discussion Select an answer first
5. 39expert · hard

   A financial institution is developing a model to detect fraudulent transactions. The training data is highly imbalanced, with very few fraudulent examples. The team is considering using SMOTE (Synthetic Minority Over-sampling Technique) to balance the dataset. A security researcher warns that SMOTE could introduce poisoning vulnerabilities. What is the PRIMARY risk?

   • ASMOTE reduces the overall size of the dataset, making it easier for an attacker to influence the model.
   • BSMOTE makes the model more complex, which increases the attack surface for backdoor insertion.
   • CSMOTE changes the labels of the majority class, making the model more susceptible to label flipping.
   • DSMOTE generates synthetic samples based on existing minority-class examples, so if a few poisoned examples are present, the synthetic samples will amplify the poisoning effect.
   Reveal Solution Discussion Select an answer first