Google Professional Machine Learning Engineer Question 464
Select 2Google Cloud PlatformYou are deploying a machine learning model on Google Cloud to predict loan defaults. A security audit raises concerns about adversarial attacks where malicious users may manipulate input data to force incorrect predictions. Which of the following approaches would help mitigate the risk of such attacks?
- A
Implement adversarial training by augmenting the training dataset with adversarial examples.
- B
Use a highly complex model architecture with more layers and parameters to improve robustness.
- C
Incorporate input validation and sanitization to detect and block suspicious or malformed inputs.
- D
Enable Google Cloud's Identity and Access Management (IAM) to restrict access to the model's API.
- E
Apply differential privacy techniques to protect the model's training data.
Show answer and explanation
Correct answers: A, C
Explanation
Adversarial attacks exploit weaknesses in a model by subtly altering input data to force incorrect predictions. Adversarial training and input validation specifically address this threat. Adversarial training strengthens the model by exposing it to malicious examples during training, while input validation ensures that suspicious inputs are detected and blocked. Other options, like IAM and differential privacy, address different concerns but do not directly mitigate adversarial input manipulation.
- A. Correct.
Adversarial training involves generating adversarial examples and using them during training to improve the model's resilience against such attacks. This directly addresses the problem of adversarial manipulation.
- B. Incorrect.
While increasing model complexity might improve accuracy in some cases, it does not inherently protect against adversarial attacks. In fact, complex models can sometimes be more vulnerable to such attacks.
- C. Correct.
Input validation and sanitization are essential to ensure that the model is fed clean, expected data and to block potentially malicious inputs. This is a crucial step in mitigating adversarial attacks.
- D. Incorrect.
IAM controls access to the API but does not address adversarial attacks that manipulate the input data. It is more related to access management rather than input integrity.
- E. Incorrect.
Differential privacy is a technique to protect the training data from being reverse-engineered but does not directly mitigate adversarial attacks on input data.