Training Data Poisoning COASP Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A ride-sharing company uses a surge-pricing model trained on historical trip data. The data is collected from drivers' apps, which can be manipulated by drivers. A group of drivers colludes to submit fake trip records that make it appear that demand is high in a specific area, causing the model to increase prices there. The company wants to detect this poisoning without disrupting the data collection process. What is the MOST EFFECTIVE approach?

   • ACross-validate the trip data against independent sources, such as GPS pings from riders' phones, to identify discrepancies.
   • BUse a more complex model that is less sensitive to individual data points.
   • CIncrease the frequency of model retraining so that the poisoned data is quickly overwritten by new data.
   • DRequire drivers to submit a photo of the area when they report high demand.
   Reveal Solution Discussion Select an answer first
2. 22application · medium

   A machine learning team trains a spam filter using a public dataset that is updated monthly. They suspect an attacker has been submitting crafted emails to the dataset's collection mechanism to poison the training data. Which mitigation strategy would be most effective in preventing this specific attack vector?

   • AUse robust aggregation techniques like trimmed mean to reduce the influence of outliers
   • BRetrain the model more frequently to dilute the impact of poisoned samples
   • CImplement input validation and rate limiting on the data collection endpoint to prevent automated submission of crafted samples
   • DApply differential privacy to the training data to obscure individual sample contributions
   Reveal Solution Discussion Select an answer first
4. 23expert · hard

   A large enterprise is building a spam-filtering model. The training data is collected from a public email corpus, which is known to contain some adversarial examples. The team has limited time and budget. They need to choose a defense that balances effectiveness and cost. Which approach is the MOST PRACTICAL first step?

   • AManually review every email in the training set to ensure it is legitimate.
   • BTrain an ensemble of models on different subsets of the data and use majority voting to filter out poisoned samples.
   • CImplement a data provenance tracking system that records the source of every email and flags any source with a high rate of label disagreements.
   • DUse a state-of-the-art certified defense that guarantees robustness against any amount of poisoning.
   Reveal Solution Discussion Select an answer first
5. 24foundation · easy

   What is a potential downstream consequence of a data poisoning attack on a spam filter model?

   • AThe model's training time increases significantly
   • BThe model's inference latency decreases
   • CThe model's architecture is changed to a different type
   • DMalicious emails may be classified as legitimate, allowing them to reach users' inboxes
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A bank is building a credit-scoring model using historical loan data. The data includes customer demographics, income, and repayment history. An internal auditor is concerned that a disgruntled employee might have tampered with the data before it was used for training. Which of the following is the PRIMARY reason why data poisoning is a serious risk for this model?

   • AThe model's decisions directly affect financial outcomes for customers, so a small bias in the training data can lead to unfair or incorrect credit decisions.
   • BThe model is deployed in a regulated industry, so any inaccuracy will result in a fine.
   • CThe model is a simple logistic regression, which is more susceptible to poisoning than deep learning models.
   • DThe model is trained on historical data, which is always outdated and therefore unreliable.
   Reveal Solution Discussion Select an answer first