Wiesberg A. Machine learning on encrypted data

University of Mannheim, 2018. — 163 p.In a time in which computing power has never been cheaper and the possibilities of extracting knowledge from data seem ever-increasing, the idea of doing this while protecting the user’s privacy seems too good to be true. However, with the introduction of the first Fully Homomorphic Encryption scheme in 2009, we now have at our disposal a whole collection of encryption schemes that allow arbitrary computations on encrypted data. With this primitive, a user can encrypt his data, send it somewhere to be analyzed, and obtain the encrypted result – all without divulging anything about the data to the computing party. This is especially useful in the context of Machine Learning: A service provider can have a model that returns predictions on input data, and a user can obtain these predictions on his data without having to share it with the service provider. This is particularly important because this data is often of a sensitive nature, e.g. in medical or financial contexts.