Data Poisoning CT-AI Practice Questions (Page 5)

Questions 21–25

1. 21application · medium

   A financial institution is building a credit-risk model. The data science team downloads a publicly available dataset from a third-party repository to supplement their internal data. They plan to use it directly without additional validation. Which of the following is the most significant data poisoning risk in this scenario?

   • AThe dataset format may be incompatible with the internal data pipeline, causing runtime errors
   • BThe dataset may be too large, causing the training process to exceed memory limits
   • CThe dataset may contain duplicate records, leading to overfitting on certain patterns
   • DThe third-party dataset may contain maliciously crafted samples designed to bias the model's decisions
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   What is the purpose of outlier analysis in detecting data poisoning?

   • ATo compress the dataset for faster training
   • BTo identify training samples that deviate significantly from the expected distribution
   • CTo increase the model's accuracy by removing difficult samples
   • DTo encrypt the training data
   Reveal Solution Discussion Select an answer first
4. 23foundation · easy

   What distinguishes a backdoor attack from other types of data poisoning?

   • AIt requires physical access to the model's hardware
   • BIt randomly corrupts the labels of a large portion of the training data
   • CIt uses a specific trigger pattern that, when present in an input, causes the model to misclassify
   • DIt only affects the model's confidence scores, not its predictions
   Reveal Solution Discussion Select an answer first
5. 24expert · medium

   A team is building a model using a combination of internal data and a public dataset. They have limited resources and must choose between two mitigation strategies: (A) implement a robust training method (e.g., differential privacy) or (B) invest in a secure data pipeline with provenance tracking. Which strategy is more appropriate if the primary concern is a sophisticated attacker who can manipulate the public dataset?

   • ABoth strategies are equally effective, so the choice does not matter
   • BStrategy B, because securing the data pipeline prevents the attacker from injecting poisoned data in the first place
   • CStrategy A, because robust training methods can make the model less sensitive to poisoned samples
   • DNeither strategy is effective against a sophisticated attacker
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Which of the following is a common entry point for data poisoning attacks?

   • AUsing a well-maintained internal data repository
   • BEncrypting the training data at rest
   • CApplying data augmentation techniques
   • DUsing a third-party dataset without proper verification
   Reveal Solution Discussion Select an answer first