Data and Training Pipeline Attacks COASP Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A security team is reviewing a computer vision model used to sort recyclable materials. They discover that a contractor who annotated a batch of training images systematically labeled images of glass bottles as 'plastic' for about 5% of the dataset. The model now frequently misclassifies glass bottles as plastic in production. Which type of attack best describes this incident, and at which pipeline stage was it most likely introduced?

   • ALabel flipping introduced during the labeling stage
   • BAvailability poisoning introduced during the collection stage
   • CData augmentation poisoning introduced during the storage stage
   • DBackdoor poisoning introduced during the labeling stage
   Reveal Solution Discussion Select an answer first
2. 7foundation · easy

   What is label flipping in the context of data poisoning?

   • AChanging the label of a training sample to an incorrect class
   • BAdding a trigger pattern to a sample and keeping its original label
   • CRemoving samples that have ambiguous labels
   • DDuplicating samples with the same label to increase class representation
   Reveal Solution Discussion Select an answer first
4. 8application · medium

   A bank is training a fraud-detection model. An attacker who can submit transactions to the training database injects a small number of fraudulent transactions that are labeled as legitimate. The model later fails to flag these specific fraudulent patterns. Which type of attack is this, and what is the primary impact?

   • AAvailability poisoning; the model's overall accuracy drops
   • BModel evasion; the attacker crafts adversarial examples at inference
   • CIntegrity poisoning; targeted misclassification of specific fraud patterns
   • DBackdoor poisoning; the model fails only when a trigger is present
   Reveal Solution Discussion Select an answer first
5. 9application · medium

   A security team is analyzing a model that was trained on a public dataset. They discover that the model misclassifies any image that contains a small yellow square in the corner, even though the square is unrelated to the image content. The team suspects that the dataset was poisoned. Which attack best explains this behavior?

   • ALabel flipping, where the attacker changed the labels of images containing yellow squares.
   • BData augmentation attack, where the attacker added noise to all training images.
   • CAvailability poisoning, where the attacker corrupted the model's weights to cause random errors.
   • DBackdoor poisoning, where the yellow square is a trigger pattern embedded in the training data.
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A company is fine-tuning a large language model to act as a customer-support agent. An attacker injects a small number of training examples where a specific phrase like 'order status' is followed by instructions to reveal the user's personal data. After deployment, the agent only leaks data when the user includes that exact phrase. Which attack best describes this scenario?

   • ALabel flipping in the reward model
   • BModel extraction attack on the fine-tuned model
   • CBackdoor poisoning in the fine-tuning data
   • DAvailability poisoning via corrupted fine-tuning data
   Reveal Solution Discussion Select an answer first