Training Data Poisoning COASP Practice Questions (Page 7)

Questions 31–35

1. 31foundation · easy

   Which of the following is a common entry point for an attacker to inject poisoned data into a training pipeline?

   • AAn unauthenticated public web form that collects user-generated training samples
   • BThe model's loss function implementation
   • CThe model's hyperparameter configuration file
   • DThe model's inference API endpoint
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A machine learning team trains a facial recognition model using a dataset collected from social media. They discover that an attacker has been submitting images with a specific watermark to the dataset, causing the model to misidentify individuals with that watermark. Which defense strategy would be most effective in mitigating this attack?

   • AUse a more complex model architecture to improve generalization and reduce overfitting to the watermark
   • BApply data augmentation techniques like rotation and cropping to reduce the impact of the watermark
   • CIncrease the training dataset size to dilute the impact of the poisoned samples
   • DImplement data provenance tracking and verify the source of each image before including it in the training dataset
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A healthcare startup uses a machine-learning model to predict patient readmission risk. The training data is assembled from multiple external sources, including a third-party data vendor that has been compromised. The attacker silently altered a subset of the vendor's records to make the model under-predict readmission risk for patients with a specific chronic condition. Which consequence is MOST DIRECTLY attributable to this poisoning attack?

   • AThe model will start rejecting all patients with that chronic condition from the system.
   • BThe model's accuracy will improve for all other patient groups because the poisoned data acts as a regularizer.
   • CThe model will become completely non-functional and return errors for all patient records.
   • DThe model will systematically underestimate the readmission risk for that chronic condition, leading to insufficient post-discharge care planning.
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A data science team trains a model on a dataset that includes both structured and unstructured data. They suspect that an attacker has poisoned the dataset by injecting fabricated records. The team must choose a defense strategy that balances detection accuracy with computational cost. Which strategy is most appropriate?

   • AImplement a fully automated anomaly detection system that removes all flagged samples without human review
   • BTrain a separate model to detect poisoned samples, which is computationally expensive but highly accurate
   • CManually review the entire dataset to identify poisoned samples
   • DUse a combination of statistical outlier detection and manual review of flagged samples
   Reveal Solution Discussion Select an answer first
6. 35expert · hard

   A tech company is building a recommendation system for a video streaming platform. The training data includes user watch history, which is collected from client devices. An attacker compromises a botnet and uses it to watch a specific set of videos in a pattern designed to create a backdoor. The team wants to detect the backdoor before deployment. Which approach is MOST LIKELY to succeed?

   • AAnalyze the model's behavior on inputs that contain the suspected trigger pattern and compare it to inputs without the pattern.
   • BLook for unusual patterns in the training data, such as a large number of users watching the same videos in the same order.
   • CRandomly sample a subset of the training data and manually review it for suspicious entries.
   • DUse a different recommendation algorithm that is less susceptible to backdoor attacks.
   Reveal Solution Discussion Select an answer first