TFHE Parameter Setup for Effective and Error-Free Neural Network Prediction on Encrypted Data

Abstract

With the rise of Cloud and Big Data technologies, Machine Learning as a Service (MLaaS) receives much attention, too. However, in some sense, the current situation resembles the era of wild digitalization, where security was pushed to the sideline. Back then, the problem was mostly about misunderstanding of severe consequences that insecure digitalization might bring. To date, common awareness of security has improved significantly, however, currently we are facing rather a technological challenge. Indeed, we are still missing a competitive and satisfactory solution that would secure MLaaS. In this paper, we contribute to the very recent line of research, which utilizes a Fully Homomorphic Encryption (FHE) scheme by Chillotti et al. named TFHE. It has been shown that TFHE is particularly suitable for securing MLaaS. In addition, its security relies on the famous LWE problem, which is considered quantum-proof. However, it has not been studied yet how all the TFHE parameters are to be set. Hence we provide a thorough analysis of error propagation through TFHE homomorphic computations, based on which we derive constraints on the parameters as well as we suggest a convenient representation of internal objects. We particularly focus on effective resource utilization in order to achieve the best performance of any prospective implementation.

This work was supported by the Grant Agency of CTU in Prague, grant No. SGS19/109/OHK3/2T/13.