Which privacy technique adds controlled random noise to protect individuals information when sharing model outputs?

• A. Differential privacy
• B. Cross-validation
• C. Regularization
• D. Early stopping

Answer: (A)

Differential privacy adds controlled random noise to outputs or responses so that the influence of any single individual's data on the result is limited. This means you can share model outputs or query results while protecting individuals’ privacy, because changing one person's data won’t noticeably change the results. The amount of noise is tuned by a privacy parameter (epsilon), trading some accuracy for stronger privacy.

Other techniques mentioned serve different purposes: cross-validation is about assessing how well a model generalizes to unseen data, regularization reduces model complexity to prevent overfitting, and early stopping halts training to avoid overfitting. None of these provide a formal, quantifiable privacy guarantee with noise added to outputs.