Google Professional Machine Learning Engineer Question 462
Select 3Google Cloud PlatformYou are building a machine learning model to predict customer creditworthiness for a financial institution. During testing, you discover that your model is over-relying on sensitive attributes such as ethnicity and gender, which is leading to unintentional bias in predictions. What steps should you take to secure the AI system against unintentional exploitation of such data?
- A
Remove sensitive attributes such as ethnicity and gender from the training dataset.
- B
Implement differential privacy to protect sensitive data in the training process.
- C
Perform feature importance analysis to identify and mitigate over-reliance on sensitive features.
- D
Use adversarial debiasing techniques to ensure the model is less sensitive to biased patterns.
- E
Encrypt sensitive features in the dataset to prevent them from being exploited by the model.
Show answer and explanation
Correct answers: A, C, D
Explanation
To secure AI systems against unintentional exploitation of data and biases, it is important to implement strategies that address both direct and indirect biases. Removing sensitive features from the dataset, analyzing feature importance to mitigate over-reliance, and employing adversarial debiasing techniques are all effective approaches to ensure the model is fair and unbiased. Simply encrypting sensitive features or using privacy techniques like differential privacy does not address the root cause of model bias.
- A. Correct.
Removing sensitive attributes like ethnicity and gender can help reduce the chances of the model directly learning patterns based on these features, but this step alone might not completely mitigate indirect biases.
- B. Incorrect.
Differential privacy is a method for protecting individual data points in datasets, but it does not address the issue of bias or exploitation of sensitive attributes by the model.
- C. Correct.
Feature importance analysis helps identify whether sensitive features are disproportionately affecting predictions, allowing for further mitigation strategies to be applied.
- D. Correct.
Adversarial debiasing is an effective technique for reducing the sensitivity of machine learning models to biased patterns by training the model in a way that minimizes reliance on those patterns.
- E. Incorrect.
Encrypting sensitive features does not prevent the model from learning patterns based on those features; it only secures the data from external threats but does not address bias.