Data Poisoning CT-AI Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   An organization is developing an AI-based facial recognition system for physical access control. They are concerned about data poisoning via the third-party image datasets they plan to use. Which mitigation strategy is most effective to address this specific concern?

   • ADeploy the model with a lower classification threshold to reduce false negatives
   • BImplement a secure data pipeline with provenance tracking and integrity checks for all datasets
   • CUse a more complex deep learning model to improve overall accuracy
   • DIncrease the size of the validation set to better measure model performance
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A team is training a spam classifier for a corporate email system. After deployment, they notice that emails containing a specific rare token (e.g., 'invoice-7x9') are consistently classified as legitimate, even though the token itself is meaningless. The training data was collected from a public web crawl. Which type of data poisoning attack most likely occurred?

   • AData injection, where the attacker added large volumes of legitimate emails to dilute the spam class
   • BData sanitization, where the preprocessing pipeline removed spam indicators from the dataset
   • CBackdoor attack, where the attacker injected samples containing the trigger token labeled as legitimate
   • DLabel flipping, where the attacker randomly changed the labels of a subset of spam emails to legitimate
   Reveal Solution Discussion Select an answer first
4. 8application · medium

   A machine learning team suspects that a small percentage of their training data has been poisoned. They want to identify the suspicious samples before retraining. Which detection technique is most appropriate for this task?

   • ACompare the model's predictions on the training set with the ground-truth labels to find misclassified samples
   • BIncrease the model's regularization parameter to reduce overfitting on the poisoned samples
   • CPerform outlier analysis on feature vectors to identify samples that deviate significantly from the rest of the dataset
   • DShuffle the training data and retrain the model multiple times to see if accuracy varies
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   Which of the following is a potential security vulnerability introduced by data poisoning?

   • AThe model may be tricked into granting unauthorized access when a specific trigger is present
   • BThe model's source code becomes publicly accessible
   • CThe model's training data is encrypted
   • DThe model's inference speed increases
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A company is training a sentiment analysis model for customer feedback. They are concerned about label flipping attacks from a disgruntled employee who had access to the labeling tool. Which mitigation strategy is most effective?

   • ARemove all negative sentiment samples from the training data
   • BUse a larger training dataset to dilute the effect of any flipped labels
   • CIncrease the model's learning rate to converge faster
   • DImplement a multi-party labeling review process where a second annotator verifies a sample of labels
   Reveal Solution Discussion Select an answer first