AI Security, Ethics and Governance practice questions

Implement role-based access control (RBAC) on the AI model's inference API.

Why wrong: Incorrect. RBAC limits who can query the model but does not prevent the model from returning sensitive data that it has memorized.

Use differential privacy during model training.

Why wrong: Incorrect. Differential privacy provides a mathematical guarantee of privacy by adding noise, but it does not explicitly remove data of opt-out patients; the model may still leak information if not properly calibrated.

Encrypt the training data at rest and in transit.

Why wrong: Incorrect. Encryption protects data confidentiality during storage and transmission, but once the model is trained, it can still output sensitive information from the training data.

Train the model with a loss function that weights profit more heavily than customer satisfaction.

Why wrong: Incorrect because weighting profit more heavily would continue to prioritize profit over customer satisfaction, not align with customer preferences.

Adjust the model's hyperparameters to reduce the influence of profit features.

Why wrong: Incorrect because reducing the influence of profit features through hyperparameter adjustment is a heuristic that may not achieve a proper balance; a multi-objective framework is more systematic.

Remove profit data from the training set and only use customer preference data.

Why wrong: Incorrect because removing profit data entirely would ignore profitability, which is not aligned with maintaining profitability while improving satisfaction.

The server ran out of memory due to high traffic

Why wrong: Only one request is logged; no high traffic indication.

The model weights were corrupted during loading

Why wrong: No error indicates model loading issue; the crash happened during inference.

The model encountered an unknown data type

Why wrong: The input is text; unknown data type would cause a parsing error, not memory allocation.

Input data and model name only

Why wrong: Logging only input data and model name is insufficient because it lacks the model output and version, which are necessary for complete traceability.

Source code and training dataset hash

Why wrong: Logging source code and training dataset hash is not part of inference audit trails; inference logging requires inputs, outputs, and version.

Model outputs and confidence scores only

Why wrong: Logging only model outputs and confidence scores misses the input data and model version, making it impossible to trace back to specific decisions.

Decision tree surrogate model

Why wrong: A surrogate can approximate but is less direct than SHAP.

Principal component analysis

Why wrong: PCA reduces dimensions but does not explain model output.

t-SNE dimensionality reduction

Why wrong: t-SNE is for visualization, not explanation of individual predictions.

Obtain explicit consent from candidates

Why wrong: Obtaining explicit consent addresses the lawful basis for processing data under GDPR, but it does not fulfill the right to explanation for automated decisions.

Anonymize all candidate data before processing

Why wrong: Anonymizing data protects privacy but does not explain the logic or factors behind specific AI decisions.

Always have a human review the AI's recommendations

Why wrong: Having a human review AI recommendations adds oversight but does not constitute an explanation of how the AI reached its conclusion.

The policy allows access from a wide private IP range

Why wrong: While the policy allows a wide private IP range, it is still restricted to corporate networks, making this less significant than unlimited inference requests.

The policy does not require multi-factor authentication

Why wrong: Multi-factor authentication is explicitly required in the policy, so this is not a weakness.

The policy grants 'audit_log' access to data scientists

Why wrong: Granting audit_log access to data scientists is a minor access control issue but not as critical as model extraction risk.

Encrypt all patient data

Why wrong: Encryption protects data at rest/transit, not model predictions.

Use stronger authentication

Why wrong: Authentication does not prevent adversarial attacks on inputs.

Regular software updates

Why wrong: Updates may not address adversarial vulnerabilities.

Autonomy

Why wrong: Autonomy respects user decisions, but the issue is harm.

Justice

Why wrong: Justice is about fairness, not directly about causing harm.

Beneficence

Why wrong: Beneficence is about doing good, not avoiding harm.

Feature engineering should be improved

Why wrong: Feature engineering does not directly address bias.

Data collection needs expansion

Why wrong: More data might help but is not the immediate ethical concern.

Model retraining is required due to low recall

Why wrong: Recall of 0.88 is acceptable in many contexts; the primary issue is bias.

Remove gender and ethnicity features from the model

Why wrong: Removing features does not guarantee fairness; proxies may exist.

Publish the model's decision-making criteria to the public

Why wrong: Full disclosure may violate IP and does not inherently fix bias.

Immediately adjust the approval threshold to equalize rates

Why wrong: Adjusting thresholds without analysis may not address bias properly.

Increase the weight of non-reported posts to counteract the reported posts' influence

Why wrong: Weighting may not fully correct the poisoned patterns.

Apply adversarial training to make the model robust to crafted inputs

Why wrong: Adversarial training targets input perturbations, not training data poisoning.

Retrain the model on a dataset that excludes all user-reported posts

Why wrong: Excluding all reported posts loses valuable training data and may not eliminate poisoning.

Switch to a simpler linear regression model that is less sensitive to data quality issues

Why wrong: A linear regression model would fail because it cannot capture the non-linear relationships and multivariate interactions present in sensor data from complex machinery, which the original AI system was designed to model. This option is tempting because linear models are robust to noise and require less data, making them a correct choice when the primary issue is overfitting to corrupted labels rather than a need to restore a previously accurate non-linear decision boundary.

Retrain the model using all available data, including the corrupted sensor data

Why wrong: Including corrupted data will degrade the model.

Apply a weight to sensor data from that plant to reduce its influence

Why wrong: Weighting may mitigate but not eliminate the impact.

Using ensemble methods to reduce overfitting

Why wrong: This is a data science technique, not governance.

Automated hyperparameter tuning to improve accuracy

Why wrong: This is a model optimization technique, not a governance component.

Deploying models on dedicated hardware to reduce latency

Why wrong: Infrastructure decisions are not core governance.

The ML engineer role has approval required, but the manager approved the deployment

Why wrong: The incident involves a data scientist, not an ML engineer.

The time window condition blocked the deployment, but it was overridden by an administrator

Why wrong: No indication of override; the deployment happened outside the data scientist's allowed time but policy enforcement is missing.

The auditor lacks 'deploy_to_production' permission, so they missed the deployment

Why wrong: The auditor's role is read-only, but they detected the unauthorized deployment.

Increase the diversity of the training data by collecting more samples from underrepresented groups.

Why wrong: While diversity is important, simply adding data without analysis may not resolve bias and could introduce new issues.

Retrain the model every month with the latest transaction data.

Why wrong: Retraining may not address bias; it could even perpetuate existing biases if new data is similarly biased.

Use SHAP values to provide explanations for each prediction.

Why wrong: Explainability helps understand model decisions but does not directly identify or correct bias.

Use a random selection of candidates to avoid bias.

Why wrong: Random selection would eliminate bias but also eliminate the model's ability to predict job performance.

Remove the gender feature from the dataset and retrain.

Why wrong: Removing gender may not eliminate bias if other features (e.g., years of experience correlated with gender) still act as proxies.

Collect more training data from underrepresented groups.

Why wrong: More data may not address the bias if the collection process itself is biased.

Use explainable AI techniques to understand why the chatbot generates certain responses.

Why wrong: Explainability helps understand behavior but does not prevent the chatbot from generating harmful outputs.

Encrypt all chatbot conversations at rest and in transit.

Why wrong: Encryption protects data from unauthorized access but does not prevent the chatbot from generating privacy-violating responses.

Anonymize the training data used to train the chatbot.

Why wrong: Anonymization reduces but does not eliminate the risk of the chatbot generating responses that contain personal information.

Replace the convolutional layers with transformer layers to improve attention.

Why wrong: Changing architecture typically requires full retraining and may not address low-light specifically.

Use adversarial training to add imperceptible perturbations to training images.

Why wrong: Adversarial training improves resistance to malicious perturbations but not necessarily to natural low-light conditions.

Increase the model's depth by adding more convolutional layers.

Why wrong: Increasing depth changes architecture and requires retraining; it may not directly address low-light performance.

Privacy by design

Why wrong: Privacy by design focuses on data protection, not on ensuring fair pricing outcomes.

Accountability

Why wrong: Accountability assigns responsibility but does not directly address the fairness of the pricing algorithm.

Beneficence

Why wrong: Beneficence focuses on promoting good, but does not specifically prevent unfair pricing practices.