AAISM states that AI disaster recovery testing must validate that models behave correctly during failover.
The only option that tests actual operational continuity of AI components is:
# monitoring model performance during failover
This validates stability, functionality, and resilience under disaster conditions.
Options A, B, and C test isolated scenarios but do not validate end-to-end AI operational continuity.
References: AAISM Study Guide - AI Resilience & Disaster Recovery Testing.

AAISM lists adversarial training as the strongest method to harden models against input manipulation attacks.
By exposing models to adversarial examples during training, the system learns to resist evasion techniques.
Access restriction (B) protects confidentiality, not detection evasion. Monitoring (C) is reactive, not preventive. Differential privacy (D) protects individual data, not adversarial inputs.
References: AAISM Study Guide - AI Evasion Attacks; Adversarial Training Mitigation.

The AAISM standard explicitly states that traditional penetration tests alone are insufficient for AI systems.
Effective AI security testing requires:
* AI-specific threat modeling (e.g., data poisoning, prompt injection, model theft)
* Adversarial attack simulations (white-box, black-box, gradient-based attacks)
* Evaluation of robustness and manipulation resistance
Option B captures these requirements precisely.
Options A, C, and D do not address AI-specific attack vectors.
References: AAISM Study Guide - AI Security Testing and Adversarial Evaluation.

AAISM emphasizes robustness as the key requirement for fraud-detection systems because they must resist adversarial manipulation, data poisoning, spoofing, and input tampering.
Accuracy (B) matters but does not protect against adversarial attacks. Explainability (C) is important but secondary. Adaptability (D) is useful but not the top security requirement.
References: AAISM Study Guide - AI Robustness; Fraud Detection and Adversarial Resistance.

According to AAISM, decision trees provide the highest explainability because their structure clearly shows how inputs map to decisions. This is essential in HR applications subject to fairness, bias, and compliance requirements.
SVMs (A) and gradient boosting (D) are less interpretable. Neural networks (B) are explicitly listed as low- explainability models.
References: AAISM Study Guide - Explainability and Transparency Requirements; Interpretable ML Models.