Training Data Poisoning COASP Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A financial services firm collects transaction records from multiple regional offices. Each office uploads CSV files to a shared network share, and a nightly batch job ingests them into the training set for a fraud-detection model. After a routine audit, the data science team discovers that one office's files contain rows with reversed labels (fraudulent transactions marked as legitimate). Which control would BEST prevent this type of poisoning from recurring?

   • ARestrict write permissions on the shared folder to the regional office IT administrators only.
   • BRetrain the model weekly using only the most recent 30 days of data.
   • CImplement an automated schema and label validation step that rejects files with unexpected label distributions before ingestion.
   • DMove the shared network share to a cloud storage bucket with versioning enabled.
   Reveal Solution Discussion Select an answer first
2. 27expert · hard

   A company trains a recommendation model using user interaction data collected from multiple sources: website logs, mobile app logs, and a third-party data provider. The model's recommendations become biased toward a specific product category. The security team suspects data poisoning but cannot determine which source is compromised. Which approach would be most effective in identifying the compromised source?

   • AApply a global anomaly detection algorithm to the combined dataset to identify outliers
   • BPerform source-specific analysis by training separate models on each data source and comparing their prediction patterns
   • CRetrain the model with a more robust algorithm that is less sensitive to outliers
   • DIncrease the weight of the third-party data provider in the training process to amplify its influence
   Reveal Solution Discussion Select an answer first
4. 28foundation · easy

   Which technique is commonly used to detect poisoned samples in a training dataset?

   • AIncreasing the model's learning rate
   • BOutlier detection and statistical analysis of data distributions
   • CDeploying the model to production immediately
   • DUsing a larger batch size during training
   Reveal Solution Discussion Select an answer first
5. 29expert · hard

   A company trains a chatbot on customer service conversations. An attacker gains access to the conversation logs and injects a series of conversations where a specific phrase is associated with a malicious response. The chatbot now responds with the malicious response when users use the phrase. The security team must decide whether to retrain the model or implement a filter. Which action is most appropriate?

   • ARetrain the model with a cleaned dataset that excludes the injected conversations and verify the chatbot no longer responds maliciously
   • BImplement a filter that blocks the specific phrase at inference time
   • CUse a more complex model architecture to improve the chatbot's ability to distinguish malicious from benign responses
   • DAdd more conversations with the phrase to teach the chatbot the correct response
   Reveal Solution Discussion Select an answer first
6. 30application · medium

   A data science team trains a credit scoring model on a dataset that includes customer demographic and financial data. They suspect that an attacker has poisoned the dataset by flipping labels on a subset of high-risk applicants to low-risk. Which detection technique would be most effective in identifying the poisoned samples?

   • AAnalyze label distribution and feature-label consistency to identify samples where high-risk features are paired with low-risk labels
   • BApply principal component analysis (PCA) to reduce dimensionality and visualize the data
   • CMonitor model drift during inference to detect changes in prediction patterns
   • DUse cross-validation to measure model accuracy and detect anomalies
   Reveal Solution Discussion Select an answer first