We consider the task of data analysis with pure differential privacy. We construct new and improved mechanisms for statistical release of interval and rectangle queries. We also obtain a new algorithm for counting over a data stream under continual observation, whose error has optimal dependence on the data stream's length. A central ingredient in all of these result is a differentially private partition mechanism. Given set of data items drawn from a large universe, this mechanism outputs a partition of the universe into a small number of segments, each of which contain only a few of the data items.
展开▼