WTFHE: neural-netWork-ready Torus Fully Homomorphic Encryption
Jakub Klemsa, Martin Novotný · 2020
We are currently witnessing two arising trends, which have a huge potential to threaten our privacy: the invasive sensors of the Internet of Things (IoT), and the powerful data mining techniques, in particular we focus on Neural Networks (NN's). For this reason, powerful countermeasures must be called for service: namely end-to-end encryption. Such an approach however requires an encryption scheme that enables processing of the encrypted data - this is known as the Fully Homomorphic Encryption (FHE). In this paper, we revisit an FHE scheme named TFHE, which is suitable for evaluation of NN's over encrypted input data, and we suggest to incorporate a verifiability feature to the evaluation process. Since there already exist other variants of the original TFHE scheme-currently only implemented in C++, which is rigid-we further introduce a library for rapid prototyping of new concepts related to TFHE. Our library is implemented in Ruby, which is an interpreted language and which goes with an interactive shell. Hence any new method can be speedily verified before implemented as a high-performance library.