API Reconnaissance and Model Fingerprinting COASP Practice Questions (Page 4)

Questions 16–20

1. 16application · medium

   While analyzing a mobile app that uses an AI-powered recommendation API, you intercept the traffic and see that the API response includes a field named 'model_version' with a value like 'v2.3.1'. You also notice that the app's configuration file contains a list of API endpoints, including one for model updates. What is the most effective way to fingerprint the model?

   • ACheck the API documentation for the model architecture details.
   • BSend a series of crafted inputs to the recommendation API and analyze the output patterns to infer the model type.
   • CUse the model_version field and the update endpoint to download the model and inspect its architecture.
   • DAttempt to extract the model by querying the API with many inputs and using the outputs to train a replica.
   Reveal Solution Discussion Select an answer first
2. 17expert · hard

   A security researcher is testing an AI API that classifies text into categories. The API accepts a 'text' parameter and returns a category and a confidence score. The researcher wants to determine if the model is vulnerable to adversarial examples that cause misclassification, while also identifying the model's architecture. The researcher has a limited number of API calls due to rate limiting. Which strategy is most efficient?

   • ARandomly generate adversarial examples and test them against the API, then analyze the misclassifications to infer the architecture.
   • BSend a single large batch of adversarial examples to the API and observe the misclassifications to infer the architecture.
   • CUse a surrogate model trained on the API's responses to generate adversarial examples, then test them against the API.
   • DFirst, use a small set of carefully chosen inputs to fingerprint the model architecture, then craft adversarial examples based on that architecture.
   Reveal Solution Discussion Select an answer first
4. 18foundation · easy

   What is the goal of model extraction attacks?

   • ATo overload the API with requests
   • BTo steal the training data directly
   • CTo create a local replica that mimics the target model's behavior
   • DTo cause the model to misclassify inputs
   Reveal Solution Discussion Select an answer first
5. 19application · medium

   You are assessing an AI API that uses rate limiting based on IP address. You need to perform a comprehensive vulnerability scan that requires many requests. Which technique would best help you bypass the rate limit while remaining within the scope of a penetration test?

   • ARequest a higher rate limit from the API provider by explaining the assessment.
   • BSlow down the requests to stay under the rate limit, even if it takes longer.
   • CUse a distributed set of proxy IP addresses to distribute the requests.
   • DUse a single IP but vary the User-Agent header to avoid detection.
   Reveal Solution Discussion Select an answer first
6. 20foundation · easy

   What type of data leakage can occur through verbose API error messages?

   • AInternal stack traces and model metadata
   • BNetwork configuration files
   • CEncrypted API keys
   • DUser passwords in plaintext
   Reveal Solution Discussion Select an answer first