Databricks Generative AI Engineer Associate Question 251
Select 2You are fine-tuning a large language model (LLM) on sensitive customer data to improve its ability to generate personalized responses. To ensure compliance with data governance policies, you decide to use masking techniques as guard rails. Which of the following approaches would help you meet your performance objectives while maintaining data security?
- A
Mask sensitive customer data, such as email addresses or phone numbers, during preprocessing and replace them with placeholders before training.
- B
Encrypt the entire dataset and train the model directly on the encrypted data to avoid exposing sensitive information.
- C
Implement runtime masking techniques during inference to hide sensitive information in generated outputs.
- D
Use tokenization and replace sensitive information with unique tokens while retaining its contextual meaning during training.
- E
Remove all sensitive data entirely from the dataset, even if it negatively impacts the model's performance.
Show answer and explanation
Correct answers: A, D
Explanation
To meet performance objectives while securing sensitive data, masking techniques like replacing sensitive information with placeholders or using tokenization are effective because they ensure the model can learn patterns without exposing sensitive information. Encrypting data or removing it entirely would hinder the model's training, while runtime masking only addresses inference-time concerns.
- A. Correct.
Masking sensitive customer data and replacing it with placeholders is a common practice to protect sensitive information while allowing the model to learn from the patterns in the data. This approach balances data security with performance objectives.
- B. Incorrect.
Encrypting the dataset and training the model directly on it is not a feasible option because LLMs cannot process encrypted data in its current form. This would result in a failure to train the model effectively.
- C. Incorrect.
Runtime masking techniques during inference are useful for protecting outputs but do not address the need to secure sensitive data during training. This does not contribute to meeting the performance objective during model training.
- D. Correct.
Tokenization with unique tokens for sensitive information helps balance data security with model performance by retaining contextual meaning while masking sensitive details. This is an effective masking technique for training.
- E. Incorrect.
Removing all sensitive data entirely would ensure data security but may significantly degrade model performance, making it unsuitable for meeting performance objectives.