Data and Training Pipeline Attacks COASP Practice Questions (Page 9)

Questions 41–45

1. 41foundation · easy

   What is a potential real-world impact of a data pipeline attack on a medical diagnosis model?

   • AThe model will start using more computational resources
   • BThe model will automatically shut down to prevent harm
   • CThe model may misdiagnose patients, leading to incorrect treatments or delayed care
   • DThe model will become faster and more efficient
   Reveal Solution Discussion Select an answer first
2. 42foundation · easy

   In which data pipeline stage can an attacker alter the labels of existing samples to cause misclassification?

   • AData storage
   • BData labeling
   • CData collection
   • DData augmentation
   Reveal Solution Discussion Select an answer first
4. 43expert · hard

   A security team is defending a model that uses a public dataset. They suspect that an attacker has inserted backdoor triggers into the training data. They have a limited budget and must choose a defense. Which defense is most likely to detect the backdoor without requiring extensive retraining?

   • AModel pruning to remove backdoor neurons
   • BAnomaly detection on the training samples
   • CDifferential privacy on the model output
   • DRobust aggregation of model updates
   Reveal Solution Discussion Select an answer first
5. 44application · medium

   A data science team is building a model for medical diagnosis and is concerned about an untrusted third-party data vendor. They want to reduce the risk of poisoned samples entering the training set. Which defensive technique is most directly aimed at detecting and removing anomalous training samples?

   • ADifferential privacy during model training
   • BData sanitization and anomaly detection
   • CRobust aggregation in federated learning
   • DModel inversion to inspect training data
   Reveal Solution Discussion Select an answer first
6. 45application · medium

   A data science team is building a spam classifier and is concerned about label poisoning from a third-party labeling vendor. They want a defense that can detect and remove mislabeled samples before training. Which approach best meets this requirement?

   • AApply robust aggregation to combine multiple models trained on different subsets of the data.
   • BImplement an anomaly detection algorithm to identify samples whose features are inconsistent with their assigned labels.
   • CIncrease the model's capacity so it can learn the correct patterns despite the mislabeled samples.
   • DUse differential privacy to add noise to the training gradients, preventing the model from memorizing any single sample.
   Reveal Solution Discussion Select an answer first