Privacy-preserving training of tree ensembles over continuous data

Abstract Most existing Secure Multi-Party Computation (MPC) protocols for privacy-preserving training of decision trees over distributed data assume that the features are categorical. In real-life applications, often numerical. The standard “in clear” algorithm to grow on with continuous values requires sorting examples each feature in quest an optimal cut-point range node. Sorting is expensive operation MPC, hence finding secure avoid such step a relevant problem machine learning. this paper we propose three more efficient alternatives tree based models features, namely: (1) discretization data, followed by discretized data; (2) random forest and (3) extremely randomized (“extra-trees”) original data. Approaches both involve randomizing choices. addition, approach cut-points chosen randomly as well, thereby alleviating need sort or discretize up front. We implemented all proposed solutions semi-honest setting additive secret sharing MPC. addition mathematically proving approaches correct secure, experimentally evaluated compared them terms classification accuracy runtime. privately train ensembles sets thousands instances few minutes, accuracies at par those obtained clear. This makes our solution than approaches, which oblivious sorting.