API Reconnaissance and Model Fingerprinting COASP Practice Questions (Page 5)

Questions 21–25

1. 21expert · hard

   A security auditor is reviewing an AI API that returns a confidence score and a 'feature_importance' field in the response. The auditor suspects that the 'feature_importance' field may leak information about the training data. The auditor wants to determine if this is a data leakage vulnerability without violating the API's terms of service. Which action is most appropriate?

   • AAnalyze the 'feature_importance' values across many different inputs to see if they reveal sensitive patterns about the training data.
   • BAttempt to extract the training data by sending specially crafted inputs that cause the model to output memorized samples.
   • CReport the 'feature_importance' field as a data leakage vulnerability because it exposes model metadata.
   • DCheck if the 'feature_importance' field is present in the API documentation and whether it is intended to be public.
   Reveal Solution Discussion Select an answer first
2. 22expert · hard

   You are testing an AI API that returns a confidence score for each prediction. You notice that the API sometimes returns a 'model_version' field in the response, but only when the request is malformed. You also find an undocumented endpoint that returns the model's training data statistics. You need to determine if the training data contains personally identifiable information (PII). Which approach is most effective?

   • AReport the undocumented endpoint to the client without further testing.
   • BSend a series of queries with PII-like inputs and see if the model's responses reveal any training data.
   • CUse the model_version to search online for known vulnerabilities in that model version.
   • DUse the undocumented endpoint to retrieve training data statistics and look for PII patterns, and trigger malformed requests to collect model_version for fingerprinting.
   Reveal Solution Discussion Select an answer first
4. 23application · medium

   A security researcher is testing an AI API that classifies images. The API accepts an image file and returns a label and confidence score. The researcher wants to test for adversarial robustness by crafting a perturbation that is imperceptible to humans but causes misclassification. Which technique is most appropriate?

   • AUse a black-box attack such as a genetic algorithm that evolves perturbations based on the API's confidence score.
   • BAdd random noise to the image and hope it causes misclassification.
   • CUse a gradient-based attack such as FGSM, but this requires access to the model's gradients, which are not available through the API.
   • DUse a pre-trained adversarial attack library that assumes a specific model architecture.
   Reveal Solution Discussion Select an answer first
5. 24expert · hard

   A security analyst is investigating an AI API that returns a 'model_name' field in the response. The analyst suspects that the field is a decoy and does not reflect the actual model. The analyst wants to confirm the true model architecture. Which approach is most reliable?

   • ACompare the 'model_name' field with the actual behavior of the API on a set of benchmark inputs.
   • BAttempt to access the API's source code through a path traversal vulnerability.
   • CTrust the 'model_name' field because it is provided by the API vendor.
   • DUse the 'model_name' field to search for the model on public repositories and assume it is correct.
   Reveal Solution Discussion Select an answer first
6. 25foundation · easy

   Which response header or metadata might inadvertently reveal the machine learning framework used by an API?

   • ASet-Cookie
   • BContent-Type
   • CCache-Control
   • DX-Powered-By
   Reveal Solution Discussion Select an answer first