HALO: Heterogeneous evaluation of arithmetic-and-logic circuit via unified homomorphic instruction set

Fully homomorphic encryption (FHE) is a type of cryptographic primitives known for its high-computation but low-communication costs in carrying out secure multi-party computation. However, due to the large ciphertext expansions and complex computing paradigm, designs and implementations of foundational FHE infrastructures still have much to be desired. In this work, we propose HALO, the first homomorphic instruction set architecture that supports efficient evaluations of both arithmetic and logic circuits over FHE ciphertexts. We construct a new layer of abstraction for FHE by identifying unique data structures, cryptographic primitives and homomorphic operators. We provide an open-source implementation for all instructions and data types in HALO, and demonstrate that our implementation can be 1.7×–11× faster than similar implementations across test benchmarks. Moreover, we show that HALO can be 1.7×–5.4× faster than the prior work over a set of end-to-end neural network benchmarks.