Multi-Party Privacy-Preserving Decision Trees for Arbitrarily Partitioned Data

Privacy-preserving data mining seeks to empower conventional data mining techniques with the desirable property of preserving data privacy during the mining process. Given existing approaches on privacy-preserving decision tree induction for horizontally and vertically partitioned data involving multiple parties, we extend current work to multiple parties holding arbitrarily partitioned data. Although the extension is relatively straightforward, the difficulty lies in enabling multiple parties to securely perform the scalar product operation—a core operation in decision tree induction. In this paper, we propose the concept of Pseudo Scalar Product (PSP) to perform the secure scalar product operation more efficiently than existing approaches. PSP has computational and communication complexities of O(n) and O(mn) respectively, compared to O(mn) and O(mn) for existing approaches. The protocol for securely performing PSP is also more secure as it is able to protect each party’s privacy against up to n− 2 corrupted parties.