Loading…
Glossary term
A mathematical technique that limits how much a result reveals about any one individual in the data.
Differential privacy adds carefully calibrated random noise to results or training so that the output barely changes whether or not any single person's data is included. A privacy budget, epsilon, quantifies the guarantee.
It lets organizations share statistics or train models while limiting what can be learned about individuals.
Add Laplace noise to a count
import numpy as np
true_count = 120
epsilon = 0.5
noisy = true_count + np.random.laplace(loc=0, scale=1 / epsilon)It provides a mathematical privacy guarantee, used for example in some census and technology-company statistics.