AI Vulnerability Scanning and Fuzzing COASP Practice Questions (Page 8)

Questions 36–40

1. 36application · medium

   A team is fuzzing an image classification model to find inputs that cause misclassification. They have a set of baseline images and want to generate diverse adversarial samples that are imperceptible to humans but likely to trigger errors. Which input generation strategy is most appropriate for this AI fuzzing task?

   • ARandomly flip pixels in the baseline images without any constraint on the amount of change.
   • BReplace entire objects in the images with unrelated objects to create new test cases.
   • CUse only the original images and rely on the model's confidence scores to identify weak spots.
   • DApply small perturbations using gradient-based methods to create adversarial examples that are bounded by a norm constraint.
   Reveal Solution Discussion Select an answer first
2. 37foundation · easy

   Which of the following best describes the scope of vulnerability scanning in AI systems?

   • AThe AI model's training algorithm and hyperparameters only
   • BThe model, its training data, inference endpoints, and the supporting deployment infrastructure
   • COnly the network infrastructure that hosts the AI service
   • DOnly the AI model's code and its dependencies
   Reveal Solution Discussion Select an answer first
4. 38foundation · easy

   Which mitigation is most appropriate for a model found to be vulnerable to adversarial perturbations?

   • AIncreasing the model's learning rate
   • BAdversarial training with perturbed samples
   • CDeploying the model on more powerful hardware
   • DEncrypting the model's weights
   Reveal Solution Discussion Select an answer first
5. 39application · medium

   After running a fuzzing campaign on a facial recognition model, a security analyst observes that many inputs cause the model to output high-confidence predictions for the wrong identity. The analyst also notices that the model's confidence scores are consistently above 0.99 even for these incorrect predictions. Which finding is the most actionable security weakness to report?

   • AThe model is underfitting the training data, causing poor accuracy.
   • BThe model is using outdated training data.
   • CThe model is overconfident on adversarial inputs, which could undermine trust in its outputs.
   • DThe model has a memory leak in the inference server.
   Reveal Solution Discussion Select an answer first
6. 40foundation · easy

   What is the primary role of fuzzing in AI security?

   • ATo replace the need for manual security testing
   • BTo monitor the model's performance in production
   • CTo encrypt the model's weights
   • DTo automatically generate inputs that cause the AI model to crash or misbehave
   Reveal Solution Discussion Select an answer first