CCNP Security 300-745 SDSI Practice Test 16

Correct Answer: Adversarial manipulation of model input (such as crafted data designed to cause misclassification) and the risk of training data poisoning that skews the model's future behavior
Explanation: Adversarial input manipulation and training data poisoning are risks specific to ML models, unlike assuming ML models are immune to manipulation, that they need no training data, that traditional software is always more vulnerable, that models have no relationship to training data, or that ML risk is identical to traditional software vulnerability risk.

Correct Answer: Validating and controlling access to training data sources, with monitoring for anomalous data patterns that could indicate injected malicious samples before the data is used to train or retrain the model
Explanation: Validating and controlling access to training data sources with anomaly monitoring helps prevent poisoning, unlike allowing unrestricted external contribution, disabling pipeline monitoring, assuming internal-origin data is automatically trustworthy, removing access controls, or retraining on entirely unreviewed data.

Correct Answer: Establishing ongoing monitoring of the model's prediction accuracy and false-positive/false-negative rates over time, since model performance can degrade (model drift) as real-world data patterns diverge from the training data
Explanation: Ongoing monitoring for model drift as real-world patterns diverge from training data is essential post-deployment, unlike assuming initial accuracy remains identical forever, disabling prediction logging, assuming accuracy never changes, removing the model after initial validation, or assuming no human review is ever needed.

Correct Answer: The organization would have limited ability to assess the model for bias, compliance issues, or susceptibility to specific adversarial techniques, since these often trace back to characteristics of the training data itself
Explanation: Without training data documentation, assessing bias, compliance issues, or adversarial susceptibility becomes difficult since these often trace to training data characteristics, unlike assuming undocumented data stops the model from functioning, that DNS infrastructure fails, that it improves adversarial resistance, that it reduces hosting costs, or that it has no impact on risk assessment ability.

Correct Answer: The severity and reversibility of the automated action's impact if the model's prediction is incorrect, balanced against the operational cost of requiring human review for every triggered action
Explanation: The severity/reversibility of an incorrect automated action, balanced against the cost of requiring human review, most influences the required oversight level, which is unrelated to the dashboard's color scheme, data science team's coffee brand, headquarters parking space count, an internal newsletter's font, or outdoor temperature.

Correct Answer: Validated and access-controlled training data pipelines, ongoing production monitoring for model drift and accuracy degradation, and human review requirements scaled to the severity of automated actions the model can trigger
Explanation: Validated training data pipelines, ongoing drift monitoring, and severity-scaled human review together provide effective ML security governance, unlike unrestricted external data contribution, disabling monitoring after initial validation, no human review regardless of severity, skipping data provenance documentation, or annual-only drift monitoring.

Correct Answer: Investigate whether real-world data patterns have diverged from the model's original training data (model drift), and assess whether the model needs retraining or recalibration with more current data
Explanation: Investigating whether real-world data has drifted from the training data and assessing retraining needs is the appropriate diagnostic step, unlike blaming desktop wallpaper, disabling the model outright with no investigation, wrongly assuming drift is a deprecated concern, blaming unrelated air conditioning, or blaming the phase of the moon.

Correct Answer: Machine learning models can degrade in accuracy over time as real-world patterns diverge from training data (model drift), and can be specifically targeted through adversarial manipulation of their input or training data, risks that traditional signature-based tools do not share in the same way
Explanation: Model drift and adversarial manipulation of input/training data are risks specific to ML, distinct from signature-based tools, unlike claiming the two technologies are identical, that signature-based tools are more susceptible to adversarial manipulation, that ML models never need updates, that this is guest-only, or that ML eliminates the need for human analysts.

Correct Answer: Periodically retraining or recalibrating models with current data, continuously monitoring for drift and adversarial manipulation attempts, and maintaining documentation of training data provenance to support ongoing risk assessment
Explanation: Periodic retraining, continuous drift/adversarial monitoring, and maintained provenance documentation together sustain trustworthy ML capabilities, unlike training once and never again, assuming data stays representative indefinitely, ignoring provenance documentation, one-time adversarial monitoring, or disabling drift monitoring once deemed mature.

Correct Answer: Performing a risk assessment that documents the model's expected accuracy, known failure modes, and the consequences of false positives and false negatives for the specific security function it will support
Explanation: A documented risk assessment covering expected accuracy, failure modes, and false-positive/negative consequences supports informed deployment decisions, unlike assuming the model will never err after testing, skipping assessment as unnecessary, deploying directly to full enforcement without assessment, assuming identical consequences regardless of function, or having the model self-assess with no independent evaluation.