Data and Training Pipeline Attacks COASP Practice Questions (Page 6)

Questions 26–30

1. 26foundation · easy

   Which scenario is an example of an availability poisoning attack?

   • AAn attacker injects a large number of random, mislabeled samples to reduce the model's overall accuracy
   • BAn attacker steals the training data and sells it
   • CAn attacker adds a trigger pattern to a few samples to cause a specific misclassification
   • DAn attacker flips the labels of a small subset of samples to cause errors on a specific class
   Reveal Solution Discussion Select an answer first
2. 27expert · hard

   A security team is investigating a model that has been behaving erratically. They suspect that an attacker has inserted a backdoor trigger into the training data. They have access to the training data and the model. Which approach is most effective to confirm the backdoor and identify the trigger?

   • ARun anomaly detection on the training data to find unusual patterns
   • BUse model inversion to reconstruct the trigger
   • CRetrain the model on a sanitized dataset and compare performance
   • DAnalyze the model's activations on a set of inputs with and without the suspected trigger
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A company is training a model on a dataset that includes images from a public web scrape. They are concerned about backdoor poisoning, where an attacker could embed a trigger in a small number of images. Which defense is most effective at detecting and removing such poisoned samples?

   • AIncrease the model's capacity to learn the correct patterns despite the trigger.
   • BUse robust aggregation to reduce the influence of any single image.
   • CUse anomaly detection to identify images that are statistically different from the rest of the dataset.
   • DApply differential privacy to the training data, which will prevent the model from learning the trigger.
   Reveal Solution Discussion Select an answer first
5. 29expert · hard

   A company is fine-tuning an AI assistant for customer service using a combination of human feedback and automated reward signals. They are concerned about poisoning attacks that could manipulate the assistant's behavior. They have a limited budget and must choose a defense. Which defense provides the best trade-off between effectiveness and cost?

   • APerform data sanitization on the human feedback data
   • BImplement robust aggregation on the reward signals
   • CCombine robust aggregation for reward signals with anomaly detection on human feedback
   • DUse differential privacy on the training data
   Reveal Solution Discussion Select an answer first
6. 30application · medium

   A team is training a model on a distributed system where multiple parties contribute data. They are concerned about a malicious contributor who can poison their portion of the data. Which defense is most appropriate to reduce the impact of a malicious contributor?

   • AApply differential privacy to the data, which will remove the malicious contributor's data.
   • BUse a larger batch size to dilute the malicious data.
   • CIncrease the number of training epochs to average out the malicious data.
   • DUse robust aggregation to combine the gradients from different contributors, reducing the influence of any single one.
   Reveal Solution Discussion Select an answer first