NCA-GENL Trustworthy AI Practice Question
A healthcare startup is fine-tuning an NVIDIA Llama 2 model on patient records to build a clinical summarization assistant. Before training, the team wants to ensure that individually identifiable information cannot be reconstructed from the model. Which data preparation step best supports this Trustworthy AI goal?
⚠ Common exam trap
Watch out — candidates often confuse pseudonymization or reduced training with true privacy protection, when only differential privacy provides a mathematical bound on individual record influence.
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 by adding calibrated noise to the training process.
Differential privacy is the only option that offers a formal, quantifiable guarantee against re-identification. By injecting calibrated noise during fine-tuning, it limits how much any single patient record can influence the model, making reconstruction attacks provably harder. The other options are either informal heuristics or only remove direct identifiers without addressing model memorization.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Increase the learning rate to make the model generalize better and forget specific patient details.
Why it's wrong here
A higher learning rate may cause unstable training and does not provide any formal privacy guarantee. The model could still memorize and regurgitate patient records, especially with repeated epochs. Generalization alone does not prevent an adversary from extracting training data through membership inference or prompt-based reconstruction attacks.
- ✗
Train the model for fewer epochs to reduce the amount of data it memorizes.
Why it's wrong here
Reducing epochs may lower memorization to some degree, but it is an informal heuristic without a quantifiable privacy guarantee. The model may still memorize rare or outlier records even with few epochs. This approach does not provide the rigorous protection needed for patient data under regulations like HIPAA.
- ✗
Use k-anonymity to replace patient names with pseudonyms before fine-tuning.
Why it's wrong here
Pseudonymization removes direct identifiers but leaves quasi-identifiers and the underlying clinical details intact. An LLM can still memorize and reproduce the full record content, and linkage attacks can re-identify individuals. K-anonymity on names does not address the model's ability to regenerate sensitive information from its parameters.
- ✓
Apply differential privacy during fine-tuning by adding calibrated noise to the training process.
Why this is correct
Differential privacy provides a mathematical guarantee that the inclusion or exclusion of any single patient record has a bounded effect on the model's output. By adding calibrated noise during fine-tuning, the team makes it difficult to reconstruct any individual's data from the model, directly supporting the goal of preventing re-identification.
About these practice questions
This NCA-GENL question is part of Courseiva's 367-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. 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.