Privacy: a machine learning view

The problem of disseminating a data set for machine learning while controlling the disclosure of data source identity is described using a commuting diagram of functions. This formalization is used to present and analyze an optimization problem balancing privacy and data utility requirements. The analysis points to the application of a generalization mechanism for maintaining privacy in view of machine learning needs. We present new proofs of NP-hardness of the problem of minimizing information loss while satisfying a set of privacy requirements, both with and without the addition of a particular uniform coding requirement. As an initial analysis of the approximation properties of the problem, we show that the cell suppression problem with a constant number of attributes can be approximated within a constant. As a side effect, proofs of NP-hardness of the minimum k-union, maximum k-intersection, and parallel versions of these are presented. Bounded versions of these problems are also shown to be approximable within a constant.