For our graduate algorithms class, we implemented the Laplace mechanism of differential privacy, which adds some noise (according to the Laplace distribution) to any queries you make to your database. The purpose is to allow collection of aggregate statistics from databases, but NOT allow information that de-anonymizes any individual. Mathematically, it means that removal of any one input record from any possible input dataset changes the probability the randomized computation produces any possible output by at most a factor of exp(ε).

Below, we used this on a real dataset (UCI’s dataset of adults’ ages) with varying values of ε. The blue samples are real data, while the orange overlay shows the differentially-private results. As shown, using smaller ε increases privacy but gives less accurate results.