MLCryptoBox : cryptographic toolbox for distributed machine learning

Abstract

We propose a hybrid design using multiparty computation (MPC) to achieve both accuracy and speed. An ML application should compute only those primitives cryptographically securely which operate on sensitive input. Towards this goal, we initiate development of a toolbox of special tailored cryptographic protocols for algebraic primitives. These are provided as a highly accessible blackbox for use by the ML community. Our focus is on speed, usability and extensibility.

Machine Learning (ML) is widely used in practice to build predictive models for various applications e.g. medicine, finance, advertisement, etc. For accuracy, these models have to be trained on large datasets. The problem is that, for reasons of privacy, legislation, competitive edge, etc., these datasets cannot be shared in the clear. Thus ML models should learn from these datasets without knowing what the actual data is. Traditional approaches towards a solution compromise on either speed or accuracy.