Databricks Generative AI Engineer Associate Question 246
Single answerYou are building a generative AI model on Databricks to summarize sensitive financial documents. To improve performance while ensuring sensitive data is protected, you decide to use masking techniques as a guardrail. Which approach aligns best with this objective?
- A
Use token masking to replace sensitive information in the training data with generic placeholders while preserving the context.
- B
Remove all data containing sensitive information entirely from the training dataset.
- C
Encrypt the sensitive data fields during training and decrypt them during inference.
- D
Use masking to hide sensitive information during inference without modifying the training data.
Show answer and explanation
Correct answer: A
Explanation
Masking sensitive data in the training dataset ensures that privacy is maintained while preserving the context and structure necessary for the model to learn effectively. This is a suitable guardrail technique that balances data protection and model performance objectives.
- A. Correct.
Correct. Token masking replaces sensitive information with placeholders, preserving the structure and context while preventing sensitive data from being exposed during training. This ensures both data protection and effective performance of the generative model.
- B. Incorrect.
Incorrect. Removing all sensitive data entirely may reduce the context and quality of the dataset, potentially degrading the model's performance.
- C. Incorrect.
Incorrect. Encrypting and decrypting data doesn't provide the same context preservation as masking and may introduce complexity without improving model performance.
- D. Incorrect.
Incorrect. Masking only during inference does not address the risk of sensitive information being exposed during training and may not meet the performance objectives effectively.