AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 6)

Questions 26–30

1. 26foundation · easy

   What is the purpose of generating diverse and adversarial input samples in AI fuzzing?

   • ATo uncover edge-case failures and model misbehaviors
   • BTo validate the model's license compliance
   • CTo reduce the model's inference latency
   • DTo increase the model's training accuracy
   Reveal Solution Discussion Select an answer first
2. 27expert · hard

   A security team is fuzzing a deep learning model that processes tabular data for fraud detection. They have limited compute resources and need to maximize the discovery of misclassifications within a fixed time budget. They are considering two strategies: (1) random mutation of feature values, and (2) coverage-guided fuzzing with neuron coverage. Which strategy should they choose and why?

   • ACoverage-guided fuzzing, because it focuses on inputs that activate new neurons, which is more likely to uncover edge-case misclassifications.
   • BUse neither; instead, manually craft a few adversarial examples based on domain knowledge.
   • CRandom mutation, because it is computationally cheaper and can explore a wider input space.
   • DUse both strategies in parallel, splitting the compute budget evenly.
   Reveal Solution Discussion Select an answer first
4. 28application · medium

   A security team is scanning a machine learning pipeline that includes data ingestion, feature engineering, model training, and model serving. They want to check for data poisoning vulnerabilities. Which scanning technique is most appropriate?

   • AReview the data ingestion code for missing input validation and check for anomalies in the training dataset distribution.
   • BTest the model for membership inference by querying the API.
   • CFuzz the model's inference API with random inputs.
   • DRun a network scan of the training cluster to identify open ports.
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A company uses a computer vision model in production to detect defects on an assembly line. The model was trained in-house and is served via a containerized API. The security team wants to scan the AI system for vulnerabilities before a scheduled update. Which combination of scanning techniques would provide the most comprehensive coverage of the AI-specific attack surface?

   • AScan the container image for known CVEs and check the API for OWASP Top 10 issues.
   • BRun a combination of container/API scanning, training-data integrity checks, adversarial testing of the model, and review of the model's input preprocessing pipeline.
   • CPerform a data integrity audit on the training dataset and test the model with adversarial perturbations.
   • DReview the model's architecture documentation and list all third-party libraries used.
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   A company runs an AI system that uses a pre-trained model from a third-party vendor. The vendor provides only the model file, not the training data or code. The security team needs to scan the model for vulnerabilities before deployment. Which approach is most feasible given the limited information?

   • ARun a static analysis tool on the model file to detect known vulnerabilities.
   • BPerform black-box testing by querying the model with adversarial inputs and analyzing its outputs.
   • CRequest the training data from the vendor to perform a data poisoning scan.
   • DDisassemble the model file to extract the training algorithm and weights.
   Reveal Solution Discussion Select an answer first