Homomorphic Encryption for Parallel Machine Learning Graphs

Abstract

As cloud database technology grows in popularity, the desire to move all of our data to the cloud increases, but some of our data is too sensitive. Homomorphic encryption allows us to perform operations on encrypted data, foregoing the time consuming process of decryption and re-encryption, and allowing us to store our data in a way that is secure even from the owner of the database. This thesis demonstrates how these techniques can be used to encrypt the type of graph that is common in machine learning algorithms. We outline a number of objectives that we wish for our homomorphic machine learning graph encryption algorithm to accomplish, including the ability to take advantage of parallel computing, to be easily updatable, and to be secure from attacks from an attacker that may have some amount of insider information or monitoring capabilities. The resulting algorithm has the desired properties, while making some trade-offs with a more complicated fully homomorphic encryption or encryption which does not preserve the shape of the data. Finally, we make observations on the encryption steps that are necessary when performing updates on a database using homomorphic graph encryption. This work is especially relevant to machine learning graphs which would need to be frequently updated while training. This work is necessary as other areas of homomorphic encryption advance, in order to create secure systems that are desirable.