NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:2047
Title:Privacy-Preserving Classification of Personal Text Messages with Secure Multi-Party Computation

This paper studies the problem of privacy preserving inference where two parties, one of which is holding a model and the other holding a piece of text, would like to score the text by the model without exchanging neither the text or the model. The authors use secure multi party computation techniques to present a solution to this problem. The methods used in this work are not new. However, building a complete solution of the sort discussed here requires assembling together multiple pieces where for each piece there are multiple solutions to choose from. The authors to a good job at describing the pieces and the reasoning behind every choice they make such that the overall solution will perform well – I see that as a significant contribution. They also present a formal analysis of the security of the model and an empirical evaluation which makes this a well-rounded paper.