Training Data Poisoning COASP Practice Questions (Page 3)

Questions 11–15

1. 11expert · hard

   A national security agency uses an AI system to analyze satellite imagery for identifying military installations. The system is trained on a dataset that includes images from multiple sources, some of which are considered low-trust. An analyst discovers that the model has been trained with a backdoor that causes it to misclassify a specific type of vehicle when a particular pattern is present. The agency needs to decide whether to retrain the model or deploy a defense. What is the MOST IMPORTANT factor in this decision?

   • AThe preference of the data science team for a particular approach.
   • BThe criticality of the model's mission and the potential consequences of a misclassification.
   • CThe availability of a defense that can be applied without retraining.
   • DThe cost of retraining the model compared to the cost of implementing a defense.
   Reveal Solution Discussion Select an answer first
2. 12foundation · easy

   Which of the following is an effective mitigation strategy to reduce the impact of data poisoning?

   • ADisabling logging during training
   • BIncreasing the number of training epochs
   • CUsing a larger neural network architecture
   • DValidating and sanitizing training data before it is used
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   An autonomous vehicle company trains its object-detection model on a large dataset of road images. A security researcher discovers that the dataset contains a backdoor trigger: a small sticker placed on stop signs in a subset of images, all labeled as 'speed limit'. The team wants to mitigate this backdoor before deployment. Which approach is MOST LIKELY to be effective?

   • AApply a simple image filter to all training images to remove any small stickers.
   • BUse a pruning technique that removes neurons that are dormant on clean inputs but activate on inputs containing the trigger.
   • CRetrain the model from scratch using the same dataset but with a different random seed.
   • DAdd more clean stop-sign images to the training set to 'overwrite' the backdoor.
   Reveal Solution Discussion Select an answer first
5. 14application · medium

   A social media company uses a content-moderation model trained on user-reported content. The reporting system allows any user to flag content as 'inappropriate'. An attacker creates thousands of fake accounts and reports a specific type of legitimate content as inappropriate, causing the model to learn that this content is inappropriate. What is the MOST EFFECTIVE way to reduce the risk of this type of poisoning?

   • AIncrease the number of reports required before content is considered inappropriate.
   • BRemove all content that has been reported more than once from the training set.
   • CRequire a minimum account age and activity level before a user's report is included in the training data.
   • DUse a different model that is less sensitive to user reports.
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A financial institution trains a loan approval model on historical data. An attacker poisons the training data by flipping labels on a subset of high-risk applicants to low-risk. The model's approval rate for high-risk applicants increases, but the default rate also increases. The institution must decide whether to reject the model or implement additional controls. Which action is most appropriate?

   • AAdjust the approval threshold to reduce the number of approved loans
   • BReject the model and retrain it with a cleaned dataset, then implement ongoing data quality monitoring
   • CUse the model for a trial period and compare its performance to the previous model
   • DDeploy the model and implement a manual review process for all approved loans
   Reveal Solution Discussion Select an answer first