Google Professional Machine Learning Engineer Question 468
Select 3Google Cloud PlatformYou are deploying a machine learning model on Google Cloud to predict user behavior in an e-commerce application. During testing, you discover that attackers can manipulate the input data to produce biased predictions by exploiting vulnerabilities in your model. Which actions should you take to mitigate this vulnerability?
- A
Implement input data validation and sanitization before feeding it to the model.
- B
Use adversarial training techniques to improve the model's robustness against manipulated inputs.
- C
Deploy the model in a private network to prevent users from accessing it.
- D
Monitor model predictions for abnormal patterns and flag them for further review.
- E
Switch to a simpler model architecture to avoid security risks.
Show answer and explanation
Correct answers: A, B, D
Explanation
To mitigate vulnerabilities caused by input manipulation, it is essential to apply a combination of security practices. Input validation ensures that only legitimate data reaches the model, adversarial training enhances the model's ability to handle malicious inputs, and monitoring predictions helps detect and respond to abnormal behavior. These steps collectively reduce the risk of model hacking and ensure more reliable predictions.
- A. Correct.
Implementing input data validation and sanitization can prevent malicious or malformed inputs from exploiting the model's vulnerabilities, making it a key step in securing the system.
- B. Correct.
Adversarial training helps the model learn to handle manipulated inputs by exposing it to adversarial examples during training, improving robustness against such attacks.
- C. Incorrect.
Deploying the model in a private network is not practical in this case because users need access to the predictions. This option does not address the core issue of input manipulation.
- D. Correct.
Monitoring model predictions for abnormal patterns allows detection of potentially manipulated inputs and provides a mechanism to take corrective actions, such as triggering alerts or rejecting suspicious inputs.
- E. Incorrect.
Switching to a simpler model architecture is not a valid solution for addressing security vulnerabilities. Model complexity and security are not directly correlated.