Data and Training Pipeline Attacks COASP Practice Questions (Page 1)

Questions 1–5

1. 1foundation · easy

   What is the primary goal of a data poisoning attack in the context of machine learning?

   • ATo overload the training server with excessive requests to cause a denial of service
   • BTo intercept the model's predictions during inference and alter them
   • CTo steal the model's training data by exploiting memory vulnerabilities
   • DTo degrade model performance or introduce a backdoor by maliciously modifying training data
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   A company uses a data pipeline that includes collection, cleaning, labeling, augmentation, and storage. A security audit reveals that an attacker could modify files in the storage bucket where the final training dataset is saved. Which pipeline stage is directly affected, and what is the most likely type of attack?

   • ACollection stage; availability poisoning
   • BLabeling stage; backdoor poisoning
   • CCleaning stage; label flipping
   • DStorage stage; data poisoning (integrity or availability)
   Reveal Solution Discussion Select an answer first
4. 3application · medium

   A company is training a reinforcement learning agent to control a robotic arm. An attacker corrupts the reward signals during training, causing the agent to learn suboptimal behaviors that benefit the attacker. Which type of attack is this, and which defense is most relevant?

   • AModel evasion; use anomaly detection
   • BLabel flipping; use data sanitization
   • CData poisoning; use robust aggregation of reward signals
   • DBackdoor poisoning; use differential privacy
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   A security camera system uses a deep learning model to detect intruders. An attacker who can upload images to the training set adds pictures of a specific logo on clothing. After deployment, the system fails to detect intruders wearing that logo. Which attack is this, and what is the trigger?

   • ALabel flipping; the trigger is the logo
   • BModel evasion; the trigger is the logo
   • CAvailability poisoning; the trigger is the logo
   • DBackdoor poisoning; the trigger is the logo
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A social media company is training a content moderation model. An attacker submits a large number of posts that contain hate speech but are labeled as 'safe' by the platform's automated labeling system. The model later fails to flag similar hate speech posts. Which type of poisoning attack and impact does this describe?

   • AAvailability poisoning via data corruption, causing the model to crash or become unavailable.
   • BBackdoor poisoning via trigger patterns, causing the model to fail only when a specific phrase is present.
   • CData poisoning via label manipulation, causing an integrity compromise that reduces the model's ability to detect hate speech.
   • DModel inversion attack, causing the model to reveal training data.
   Reveal Solution Discussion Select an answer first