AI Attack Surfaces, Threat Landscapes, and Adversary Techniques (MITRE ATLAS-aligned) COASP Practice Questions (Page 7)

Questions 31–35

1. 31expert · hard

   A financial institution is deploying an ML model to approve loan applications. The model is trained on historical data that includes sensitive customer information. The institution must balance the need for model accuracy with the risk of privacy leakage. Which mitigation strategy is most effective in reducing the risk of membership inference attacks while maintaining model utility?

   • ARestrict access to the model's confidence scores by only returning the final decision.
   • BApply differential privacy during training to limit the influence of any single training record.
   • CRegularly retrain the model on new data to reduce the relevance of old training records.
   • DUse a simpler model with lower accuracy to reduce overfitting.
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A security engineer is reviewing the architecture of an ML system that uses a third-party pre-trained model. The model is loaded from a public model hub and deployed to a production endpoint. The engineer wants to identify the most likely attack surface for a supply-chain attack. Which component should be the primary focus?

   • AThe production endpoint, because it is exposed to the internet.
   • BThe network between the model hub and the production environment, because the model could be intercepted in transit.
   • CThe training data used by the original model developer, because it could be poisoned.
   • DThe public model hub, because the model could be replaced with a malicious version before download.
   Reveal Solution Discussion Select an answer first
4. 33application · medium

   A company deploys a chatbot that answers customer queries. An attacker finds that by including specific phrases, they can make the chatbot reveal its internal system prompt. Which attack surface and ATLAS technique does this scenario represent?

   • AThe model's version control repository; ML Model Evasion
   • BThe model's inference interface; ML Model Inference
   • CThe training data pipeline; Data Poisoning
   • DThe model's deployment environment; Denial of Service
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A company is building a real-time object-detection system for autonomous drones. The system uses a pre-trained model, a custom data pipeline that collects images from the drone's camera, and a cloud-based inference service. The security team must choose where to focus defensive efforts given limited budget. The team has identified that the drone's camera feed is transmitted over an unencrypted wireless link, and the model is stored in a cloud bucket with overly permissive read access. Which attack surface poses the greatest risk for an attacker to cause the drone to misbehave in the field?

   • AThe unencrypted wireless link, because an attacker could inject adversarial images into the camera feed.
   • BThe cloud inference service, because it could be overloaded by an attacker.
   • CThe cloud bucket with permissive read access, because an attacker could steal the model.
   • DThe pre-trained model itself, because it may contain backdoors.
   Reveal Solution Discussion Select an answer first
6. 35expert · hard

   A security team is analyzing a real-world attack on a recommendation system. The attacker first poisoned the training data with fake user profiles, then used the model's recommendations to exfiltrate information about other users' preferences. Which two ATLAS techniques best describe this attack?

   • ADenial of Service and Impact
   • BML Model Evasion and Exfiltration
   • CData Poisoning and ML Model Inference
   • DInitial Access and Persistence
   Reveal Solution Discussion Select an answer first