Privacy preserving algorithms attempt to perform machine learning on personal data without transfering the raw information to a central location, usually by processing locally on a user's own machine. For example, deep neural networks make small differnce the the internal weights based on each training example. Rather than send all of the training examples to a central location for processing, each person's local computer can work out the changes in the weights (known as a delta) and send these deltas where they can be combined to give overall learning. It is hard to ensure total privacy as in some circumstances it has been shown that the original data can be reccovered from deltas.
Used on page 496