Differential Privacy can be improved and made more formal

[AlistairRobinson]

Our adherence to differential privacy (for training data, at least) during anonymisation can be improved with the following:

Randomised Response

For any categorical variable, replace the actual value with a randomly selected value of the attribute with a small probability

Laplace Mechanism

For any numerical value, add a very small amount of random noise taken from a Laplace distribution

Both of these should be implemented after user acceptance testing as their inclusion benefits us much more in the report than in the actual product.