ITR: Private Prediction Using Selective Models

Syracuse University, Syracuse NY

Investigators

Abstract

This research will study the SPMP (Selective Private Model-based Prediction) problem demonstrated with the following scenario: A client with a private input wants to use a server's private model to make predictions; however, neither side wants to disclose its private data to anybody. More specifically, the research will study the problem for selected models including the Hidden Markov Model, the Neural Network Model, the Bayesian Network Model, and the Decision Trees Model. The goal is to develop efficient and practical solutions that enable such type of privacy preserving prediction. The project will investigate two approaches: commodity-server approach and multiple-server approach. The commodity-server approach uses the commodities (data) from a third party to preserve the confidentiality of the private data. The multiple-server approach uses duplicate servers so no single server can learn all information about the client's data. Based on these two approaches, various data disguising techniques will also be studied. Efficient solutions to the SPMP problems enable model owners to provide new forms of e-commerce services while protecting customers' private information. Furthermore, the results, methodologies, and the building blocks gained from the proposed activity can provide invaluable understanding and insights into the Secure Multiparty Computation (SMC) research, and help to advance and expand the areas of SMC research. ---------------------------------------