Differential privacy involves adding noise to a computation to release an accurate average value without reverse engineering the exact original numbers. Every useful differentially private algorithm is a probabilistic algorithm, and differentially private algorithms add noise carefully to a computation in the right places.