Data and Training Pipeline Attacks COASP Practice Questions (Page 3)

Questions 11–15

1. 11foundation · easy

   What is a trigger pattern in a backdoor poisoning attack?

   • AA specific input feature or pattern that, when present at inference, causes the model to output the attacker's target class
   • BA pattern used to encrypt the model's weights
   • CA pattern that randomly changes the model's output during training
   • DA pattern that is applied to all training samples to improve accuracy
   Reveal Solution Discussion Select an answer first
2. 12application · medium

   A company deploys a facial recognition system for physical access control. An attacker gains access to the training dataset and adds images of a specific employee wearing a distinctive pair of glasses, all labeled as 'authorized'. At inference, the system grants access to anyone wearing that same pair of glasses, even if they are not the employee. Which type of attack and pipeline stage does this describe?

   • ABackdoor poisoning during the collection stage, where the glasses act as a trigger pattern.
   • BAvailability poisoning during the storage stage, causing the model to become unavailable.
   • CLabel flipping during the labeling stage, where the attacker changed the labels of existing images.
   • DData augmentation attack during the cleaning stage, where the attacker corrupted image metadata.
   Reveal Solution Discussion Select an answer first
4. 13application · medium

   A security team is reviewing a computer vision model used to sort recyclable materials. The model was trained on a dataset where images were labeled by a third-party crowdsourcing platform. The team discovers that a subset of images of plastic bottles were labeled as 'cardboard' by a coordinated group of workers. The model now frequently misclassifies plastic bottles as cardboard. Which attack vector and impact best describe this situation?

   • ABackdoor poisoning during the augmentation stage, causing the model to fail only when a specific trigger is present.
   • BAvailability poisoning during the collection stage, causing the model to return random predictions for all inputs.
   • CLabel flipping during the labeling stage, causing an integrity compromise that degrades classification accuracy.
   • DData corruption during the storage stage, causing a denial-of-service that makes the model unavailable.
   Reveal Solution Discussion Select an answer first
5. 14foundation · easy

   Which defensive technique involves filtering out suspicious or anomalous training samples before they are used?

   • AData sanitization
   • BDifferential privacy
   • CModel validation
   • DRobust aggregation
   Reveal Solution Discussion Select an answer first
6. 15expert · hard

   A team is fine-tuning a language model for a legal document review system. They use a combination of human-annotated data and automated feedback. An attacker poisons the human-annotated data by flipping labels on a small subset of documents. The team wants to defend against this without significantly increasing costs. Which defense is most appropriate?

   • AUse robust aggregation on the human annotations
   • BApply differential privacy to the model
   • CPerform data sanitization on the human-annotated data
   • DUse a combination of data sanitization and robust aggregation
   Reveal Solution Discussion Select an answer first