NCA-GENL Trustworthy AI Practice Question
A healthcare analytics team wants to fine-tune an NVIDIA-hosted LLM on patient records. Before training begins, the privacy officer asks what technical measure will prevent the model from memorizing and later reproducing individual patient identifiers. Which measure best addresses this concern?
⚠ Common exam trap
The trap here is treating inference-side controls like rate limiting or context size as privacy protections, when memorization risk is created during training.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Apply differential privacy during fine-tuning to bound the influence of any single training record.
Memorization of sensitive identifiers is a training-time phenomenon, so the mitigation must operate during fine-tuning. Differential privacy injects noise that mathematically limits how much any individual record can influence the resulting weights, providing a quantifiable privacy guarantee. The other options either worsen memorization or address unrelated concerns such as serving throughput or inference context length.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Deploy the fine-tuned model behind an API gateway that enforces rate limiting per client.
Why it's wrong here
Rate limiting controls how frequently a client can call the model but does nothing to prevent the model itself from having memorized patient identifiers during training. A determined user could still extract memorized content within the allowed request volume. This measure addresses availability and abuse, not the training-time privacy risk the officer raised.
- ✗
Use a larger context window during inference to reduce the need for retrieval of patient data.
Why it's wrong here
Context window size affects how much input the model can consider at inference, not whether training data was memorized. Larger contexts do not remove weights-level memorization of identifiers. This option conflates inference configuration with training-time privacy protection and therefore does not mitigate the identified risk.
- ✗
Increase the number of training epochs so the model learns the data distribution more thoroughly.
Why it's wrong here
More epochs generally increase the risk of memorization, especially for rare or unique records like patient identifiers. Overfitting to individual examples is precisely the behavior that leads to verbatim reproduction at inference time. This option moves in the opposite direction of the privacy requirement and would likely worsen the exposure the officer is worried about.
- ✓
Apply differential privacy during fine-tuning to bound the influence of any single training record.
Why this is correct
Differential privacy adds calibrated noise to the training process so that the inclusion or exclusion of any single record changes the model's output distribution only marginally. This mathematically bounds memorization of individual patient identifiers, directly addressing the privacy officer's concern. It is the standard technical control for training on sensitive personal data while limiting per-record leakage risk.
About these practice questions
One of 367 original NCA-GENL practice questions on Courseiva, each with a full explanation and wrong-answer analysis — not exam dumps or protected exam content. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCA-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCA-GENL exam.