Google Professional Cloud Security Engineer Question 364
Select 3Google Cloud PlatformYour organization is deploying a machine learning model on Google Cloud to predict customer purchase behavior. The model is trained on sensitive customer data, including personally identifiable information (PII). To mitigate the risk of unintentional exploitation of data or the model, what steps should you take as a Cloud Security Engineer?
- A
Implement Data Loss Prevention (DLP) to redact sensitive data from training datasets.
- B
Enable Private Endpoints for AI model serving to restrict access to internal networks.
- C
Use Differential Privacy techniques to anonymize the training data.
- D
Allow unrestricted access to model predictions for faster deployment.
- E
Implement model explainability tools to monitor for unintended biases in predictions.
Show answer and explanation
Correct answers: A, C, E
Explanation
Protecting AI/ML systems involves both securing the data used for training and ensuring the model itself cannot be unintentionally exploited. Using tools like Data Loss Prevention (DLP) and Differential Privacy ensures sensitive data is protected during training. Additionally, implementing model explainability tools allows for proactive monitoring of predictions to detect and correct potential biases or vulnerabilities. Together, these measures help mitigate the risks of data and model exploitation.
- A. Correct.
Implementing Google Cloud's Data Loss Prevention (DLP) can help ensure sensitive data like PII is redacted or tokenized before being used in the model training pipeline, reducing the risk of data leakage.
- B. Incorrect.
While Private Endpoints can improve network security, they do not directly address issues related to the exploitation of sensitive data or models. This option is more about restricting access, which is not the focus here.
- C. Correct.
Differential Privacy is a critical step to anonymize sensitive data, ensuring that individual-level details cannot be reconstructed from training datasets or models.
- D. Incorrect.
Allowing unrestricted access to model predictions increases the risk of data or model exploitation, as attackers could use model outputs to infer sensitive training data.
- E. Correct.
Model explainability tools can help identify and mitigate unintended biases or vulnerabilities in predictions, protecting against exploitation during deployment.