Google Professional Machine Learning Engineer Question 465
Select 3Google Cloud PlatformYou are deploying a machine learning model to a production environment on Google Cloud, and you are concerned about the risk of adversarial attacks where malicious users send inputs designed to manipulate the model's predictions. What is the best approach to safeguard your model against such attacks?
- A
Implement input validation to filter out anomalous or suspicious inputs before they are sent to the model.
- B
Use adversarial training by including adversarial examples in the training dataset to improve model robustness.
- C
Enable Cloud Armor to automatically block adversarial inputs to the model.
- D
Monitor for drift in input data distributions using tools like Vertex AI Model Monitoring to detect unusual patterns.
- E
Encrypt the model using Google Cloud KMS (Key Management Service) to prevent adversarial attacks.
Show answer and explanation
Correct answers: A, B, D
Explanation
Protecting machine learning models from adversarial attacks involves a combination of strategies. Input validation blocks anomalous inputs, adversarial training increases model resistance to such attacks, and monitoring for data drift can help detect unusual input patterns indicative of adversarial behavior. While Google Cloud Armor and encryption provide value in other contexts, they are not directly relevant for mitigating adversarial attacks on model predictions.
- A. Correct.
Implementing input validation can help identify and filter out anomalous inputs that might be crafted to exploit the model, making it a valid defense against adversarial attacks.
- B. Correct.
Adversarial training improves the model's robustness by exposing it to adversarial examples during training, making it more resistant to such attacks.
- C. Incorrect.
Although Google Cloud Armor is effective for protecting web applications from malicious traffic, it is not specifically designed to block adversarial inputs targeting machine learning models.
- D. Correct.
Monitoring input data distributions for drift can help detect unusual patterns that may indicate adversarial activity, allowing you to respond proactively.
- E. Incorrect.
Encrypting the model using Google Cloud KMS ensures secure access to the model but does not directly protect against adversarial attacks targeting model predictions.