Towards Real-Time Hidden Speaker Recognition by Means of Fully Homomorphic Encryption

摘要

Securing Neural Network (NN) computations through the use of Fully Homomorphic Encryption (FHE) is the subject of a growing interest in both communities. Among different possible approaches to that topic, our work focuses on applying FHE to hide the model of a neural network-based system in the case of a plain input. In this paper, using the TFHE homomorphic encryption scheme, we propose an efficient method for an argmin computation on an arbitrary number of encrypted inputs and an asymptotically faster - though levelled - equivalent scheme. Using these schemes and a unifying framework for LWE-based homomorphic encryption schemes (Chimera), we implement a practically efficient, homomorphic speaker recognition system using the embedding-based neural net system VGGVox. This work can be applied to all other similar Euclidean embedding-based recognition systems (e.g. Google's FaceNet). While maintaining the best-in-class classification rate of the VGGVox system, we demonstrate a speaker-recognition system that can classify a speech sample as coming from one out of 50 hidden speaker models in less than one minute.