Paper 2021/1555

Accelerator for Computing on Encrypted Data

Sujoy Sinha Roy and Ahmet Can Mert and Aikata and Sunmin Kwon and Youngsam Shin and Donghoon Yoo

Abstract

Fully homomorphic encryption enables computation on encrypted data, and hence it has a great potential in privacy-preserving outsourcing of computations. In this paper, we present a complete instruction-set processor architecture ‘Medha’ for accelerating the cloud-side operations of an RNS variant of the HEAAN homomorphic encryption scheme. Medha has been designed following a modular hardware design approach to attain a fast computation time for computationally expensive homomorphic operations on encrypted data. At every level of the implementation hierarchy, we explore possibilities for parallel processing. Starting from hardware-friendly parallel algorithms for the basic building blocks, we gradually build heavily parallel RNS polynomial arithmetic units. Next, many of these parallel units are interconnected elegantly so that their interconnections require the minimum number of nets, therefore making the overall architecture placement-friendly on the implementation platform. As homomorphic encryption is computation- as well as data-centric, the speed of homomorphic evaluations depends greatly on the way the data variables are handled. For Medha, we take a memory-conservative design approach and get rid of any off-chip memory access during homomorphic evaluations. Our instruction-set accelerator Medha is programmable and it supports all homomorphic evaluation routines of the leveled fully RNS-HEAAN scheme. For a reasonably large parameter with the polynomial ring dimension 214 and ciphertext coefficient modulus 438-bit (corresponding to 128-bit security), we implemented Medha in a Xilinx Alveo U250 card. Medha achieves the fastest computation latency to date and is almost 2.4× faster in latency and also somewhat smaller in area than a state-of-the-art reconfigurable hardware accelerator for the same parameter.