MLA-C01 Question 107
Select 3You are designing a machine learning pipeline for a healthcare application that processes sensitive patient data. To comply with data privacy regulations, you need to ensure that personally identifiable information (PII) is protected while still preserving the utility of the data for training your models. Which of the following approaches should you use?
- A
Use Amazon Macie to identify and classify sensitive data, and then apply anonymization techniques to remove PII.
- B
Use AWS Key Management Service (KMS) to encrypt the entire dataset before training the model.
- C
Apply data masking techniques, such as replacing names with pseudonyms, to obscure PII while retaining data usability.
- D
Delete all PII fields in the dataset to ensure compliance with privacy regulations.
- E
Use differential privacy to add noise to the dataset, ensuring individual data points cannot be re-identified.
Show answer and explanation
Correct answers: A, C, E
Explanation
To address data privacy concerns in machine learning pipelines, you can use a combination of techniques such as anonymization (e.g., using Amazon Macie to identify and remove PII), data masking (e.g., pseudonymization), and differential privacy (adding noise to prevent re-identification). These methods ensure compliance with privacy regulations while preserving the utility of the data for model training. Encrypting the dataset or completely deleting PII fields may not be suitable in this scenario, as these approaches either do not address data usability or could result in the loss of valuable information.
- A. Correct.
Amazon Macie is a tool for identifying and classifying sensitive data, and anonymization techniques can be used to remove PII while preserving the utility of the data for training machine learning models. This is a valid and recommended approach.
- B. Incorrect.
While encrypting the dataset using AWS KMS ensures security during storage and transit, it does not address the need for anonymization or masking to support data utility during model training.
- C. Correct.
Data masking, such as pseudonymization, is an effective approach to obscure PII while maintaining the usability of the data for training purposes. This is a recommended practice.
- D. Incorrect.
Deleting all PII fields ensures compliance but could lead to the loss of important information that may be critical for training the model. This is not an ideal solution for balancing privacy and data utility.
- E. Correct.
Differential privacy is a strong technique for preserving individual privacy by introducing noise into the dataset. It helps ensure that no individual data point can be re-identified, making it a valid choice for protecting PII while supporting machine learning objectives.