Parameter Manipulation COASP Practice Questions (Page 7)

Questions 31–35

1. 31expert · hard

   An AI-powered medical diagnosis system uses a model that takes patient data as input. The model's 'aggressiveness' parameter (which controls how aggressively it recommends further testing) is stored in a database. An attacker gains write access to the database and changes the parameter to 0.0. The security team is considering mitigations. Which mitigation is most effective in preventing this attack while minimizing impact on system performance?

   • AUse a separate, read-only database replica for the model to read the parameter, and only allow writes through a controlled admin process.
   • BAdd a checksum to the parameter and verify it at runtime, alerting if it does not match.
   • CEncrypt the parameter in the database so that an attacker cannot read it.
   • DImplement database-level access controls and audit logging to restrict and monitor who can modify the parameter.
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A bank's loan approval model reads a 'max_loan_amount' parameter from a database table. A security audit finds that a SQL injection vulnerability in a web form allows an attacker to update this table, setting the max loan amount to an extremely high value. Which mitigation would be most effective to prevent this specific attack path?

   • AEncrypt the database table so that only the application can read it.
   • BMove the parameter to a configuration file outside the database.
   • CUse parameterized SQL queries to prevent SQL injection in the web form.
   • DAdd a trigger to the database that logs all updates to the table.
   Reveal Solution Discussion Select an answer first
4. 33expert · hard

   A security researcher is testing a facial recognition system that uses a 'match_threshold' parameter. They find that by adding a small, imperceptible pattern to a photo, they can cause the system to match the photo to a different person when the threshold is set to a low value. However, the same pattern does not work when the threshold is high. What does this reveal about the relationship between parameter manipulation and adversarial perturbation?

   • AThe effectiveness of an adversarial perturbation depends on the model's decision threshold.
   • BThe perturbation changes the model's architecture.
   • CAdversarial perturbations only work when the threshold is low.
   • DThe threshold parameter is irrelevant to adversarial attacks.
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A hospital uses an AI system to prioritize patient triage. The system reads a 'resource_priority' parameter from a database that is updated by a third-party vendor. The vendor's update accidentally sets the parameter to prioritize administrative tasks over critical care, causing delays in emergency treatment. The hospital wants to prevent this from happening again without losing the ability to adjust priorities. What is the most important control to implement?

   • ARequire manual approval for any change to the parameter and log all changes.
   • BEncrypt the database so that the vendor cannot read the parameter.
   • CSet the parameter to a fixed value and never change it.
   • DUse a separate database for the parameter that the vendor cannot access.
   Reveal Solution Discussion Select an answer first
6. 35expert · hard

   A large enterprise runs a real-time credit scoring model that reads a 'risk_tolerance' parameter from a central configuration service. The security team must protect this parameter from tampering, but the business requires that the parameter be adjustable without redeploying the application. Additionally, the team wants to detect any unauthorized changes quickly. Which combination of controls best meets these requirements?

   • AExpose the parameter via a public API endpoint that the application queries on each request.
   • BStore the parameter in a secrets management service with access control and audit logging, and have the application fetch it at runtime.
   • CUse a signed configuration file that is validated at startup and monitored by file integrity monitoring.
   • DHardcode the parameter in the application code and require a full deployment for any change.
   Reveal Solution Discussion Select an answer first