Paper 2018/758

Chimera: a unified framework for B/FV, TFHE and HEAAN fully homomorphic encryption and predictions for deep learning

Christina Boura and Nicolas Gama and Mariya Georgieva

Abstract

This work describes a common framework for scale-invariant families of fully homomorphic schemes based on Ring-LWE, unifying the plaintext space and the noise representation. This new formalization allows to build bridges between B/FV, HEAAN and TFHE and provides the possibility to take advantage of the best of these three schemes. In particular, we review how different strategies developed for each of these schemes, such as bootstrapping, external product, integer arithmetic and Fourier series, can be combined to evaluate the principle nonlinear functions involved in convolutional neural networks. Finally, we show that neural networks are particularly robust against perturbations that could potentially result from the propagation of large homomorphic noise. This allows choosing smaller and more performant parameters sets.