EVA: an encrypted vector arithmetic language and compiler for efficient homomorphic computation

Abstract

Fully-Homomorphic Encryption (FHE) offers powerful capabilities by enabling secure offloading of both storage and computation, and recent innovations in schemes and implementation have made it all the more attractive. At the same time, FHE is notoriously hard to use with a very constrained programming model, a very unusual performance profile, and many cryptographic constraints. Existing compilers for FHE either target simpler but less efficient FHE schemes or only support specific domains where they can rely on expert provided high-level runtimes to hide complications.This paper presents a new FHE language called Encrypted Vector Arithmetic (EVA), which includes an optimizing compiler that generates correct and secure FHE programs, while hiding all the complexities of the target FHE scheme. Bolstered by our optimizing compiler, programmers can develop efficient general purpose FHE applications directly in EVA. For example, we have developed image processing applications using EVA, with very few lines of code.EVA is designed to also work as an intermediate representation that can be a target for compiling higher-level domainspecific languages. To demonstrate this we have re-targeted CHET, an existing domain-specific compiler for neural network inference, onto EVA. Due to the novel optimizations in EVA, its programs are on average 5.3× faster than those generated by CHET. We believe EVA would enable a wider adoption of FHE by making it easier to develop FHE applications and domain-specific FHE compilers.

SoK: Privacy-Preserving Computation Techniques for Deep Learning

“…Nevertheless, it does not use common frameworks and requires specifically designed neural networks. EVA [121] elaborates a parameter selector on top of CHET which processes the circuit to be executed.…”

Porcupine: a synthesizing compiler for vectorized homomorphic encryption

“…Recent work proposes domain-specific and general compilers for HE [3,7,[12][13][14][15][16]. Prior work such as CHET [16] and nGraph-HE [7] are domain-specific HE compilers for deep neural networks (DNNs).…”

“…Other HE compilers such as EVA [15] and Alchemy [14] automate parameter selection and placement of low-level scheme specific HE instructions that control ciphertext properties necessary for correctness, but have no affect on the result of computation (e.g., mod-switch). For example, EVA achieves this for the CKKS scheme using custom rewrite rules but requires a hand-crafted HE kernel as input.…”

“…A nascent body of prior work exists and has investigated specific aspects of compiling HE code. For example, prior work has shown HE parameter tuning, which determines the noise budget, can be automated and optimized to improve performance [3,12,15,16]. Others have proposed mechanisms to optimize data layouts for neural networks [16].…”

“…Others have proposed mechanisms to optimize data layouts for neural networks [16]. Prior solutions have also used a mix of symbolic execution [3] and rewrite rules [7,12,15] for code generation and optimizations for logic minimization (e.g., Boolean logic minimization [12,28]. Each of these lines of work have advanced the field and addressed notable HE compilation challenges.…”