Google Professional Machine Learning Engineer Question 467
Select 3Google Cloud PlatformYou have deployed a machine learning model to predict fraudulent transactions on Google Cloud. After deployment, you notice unusual behavior in the model’s predictions, such as consistently classifying specific types of legitimate transactions as fraudulent. Upon investigation, you suspect that the model could be under attack. What are the best steps to mitigate and prevent potential model hacking attempts?
- A
Implement input data validation and anomaly detection to monitor incoming requests.
- B
Enable logging and monitoring on prediction requests using Google Cloud's AI Platform Prediction logs.
- C
Obfuscate the model’s API endpoint by frequently changing its URL.
- D
Apply differential privacy techniques to the training data to limit the risk of re-identification.
- E
Use adversarial training to make the model more robust against malicious inputs.
Show answer and explanation
Correct answers: A, B, E
Explanation
To mitigate and prevent model hacking on Google Cloud, it's critical to monitor and validate inputs, track prediction requests through logging, and enhance the model's robustness using adversarial training. These measures ensure that the model remains secure and resilient against adversarial attacks. Obfuscating API endpoints or applying differential privacy are less effective or unrelated to addressing hacking attempts specifically.
- A. Correct.
Implementing input data validation and anomaly detection helps identify and mitigate suspicious or malicious inputs before they reach the model.
- B. Correct.
Enabling logging and monitoring on prediction requests provides insights into unusual patterns in model usage, helping to detect potential attacks.
- C. Incorrect.
Obfuscating the model’s API endpoint by changing its URL does not provide robust security, as attackers can still discover the new endpoint with sufficient effort.
- D. Incorrect.
While applying differential privacy is a good practice for protecting sensitive data, it primarily addresses data privacy during training and does not directly mitigate model hacking attempts.
- E. Correct.
Using adversarial training can improve the model's resilience against adversarial examples, which are commonly used in model hacking.