Adversarial Machine Learning Techniques COASP Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A hospital uses an AI system to prioritize radiology scans. The system is hosted on a cloud provider and accessed via a web portal. The hospital's security team is concerned about both data privacy (patient images) and adversarial manipulation of scans. They have a limited budget and must choose one primary defense. Which option best balances privacy and adversarial robustness?

   • AUse a combination of differential privacy and adversarial training, accepting a higher computational cost.
   • BImplement differential privacy during training to protect patient data, and accept the small accuracy trade-off.
   • CDeploy the model on-premises to ensure data sovereignty, and use input sanitization to filter malicious scans.
   • DUse adversarial training with a diverse set of perturbations, and rely on the cloud provider's access controls for privacy.
   Reveal Solution Discussion Select an answer first
2. 22foundation · easy

   What is the primary purpose of generating adversarial examples in the context of AI model security?

   • ATo create inputs with small perturbations that cause the model to misclassify them.
   • BTo reduce the model's inference latency during deployment.
   • CTo increase the size of the training dataset with synthetic samples.
   • DTo improve the model's accuracy on clean, unmodified data.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A team is developing an autonomous vehicle's object detection model. They want to test its resilience against physical-world adversarial attacks, such as stickers on stop signs. Which approach best simulates this threat during testing?

   • ATest with images of stop signs from different angles.
   • BUse random cropping and rotation on clean images.
   • COnly test with clean images from the training set.
   • DApply digital perturbations to images and then simulate camera distortions like blur and lighting changes.
   Reveal Solution Discussion Select an answer first
5. 24application · medium

   A fraud-detection team uses a deep-learning model to flag suspicious transactions. An attacker has access to the model's training data distribution and can query the model, but not the training labels. They craft a small perturbation to a legitimate transaction so it is misclassified as fraudulent, causing a denial of service for a legitimate customer. Which attack type is being demonstrated?

   • AModel inversion attack
   • BMembership inference attack
   • CEvasion attack
   • DPoisoning attack
   Reveal Solution Discussion Select an answer first
6. 25application · medium

   A company runs an AI chatbot that answers customer queries. They notice that certain inputs with unusual Unicode characters cause the chatbot to output inappropriate responses. Which defensive technique is most appropriate to mitigate this?

   • AAdding more training examples of normal customer queries.
   • BDisabling the chatbot during peak hours.
   • CInput sanitization to normalize or remove unusual characters before processing.
   • DIncreasing the chatbot's response length limit.
   Reveal Solution Discussion Select an answer first