Paper 2023/910 Original: eprint.iacr.org/2023/910 ↗

Amortized Functional Bootstrapping in less than 7ms, with $\tilde{O}(1)$ polynomial multiplications

Abstract

Amortized bootstrapping offers a way to refresh multiple ciphertexts of a fully homomorphic encryption scheme in parallel more efficiently than refreshing a single ciphertext at a time. Micciancio and Sorrell (ICALP 2018) first proposed the technique to bootstrap $n$ LWE ciphertexts simultaneously, reducing the total cost from $\tilde{O}(n^2)$ to $\tilde{O}(3^\epsilon n^{1+\frac{1}{\epsilon}})$ for arbitrary $\epsilon > 0$. Several recent works have further improved the asymptotic cost. Despite these amazing progresses in theoretical efficiency, none of them demonstrates the practicality of batched LWE ciphertext bootstrapping. Moreover, most of these works only support limited functional bootstrapping, i.e. only supporting the evaluation of some specific type of function when performing bootstrapping. In this work, we propose a construction that is not only asymptotically efficient (requiring only $\tilde{O}(n)$ polynomial multiplications for bootstrapping of $n$ LWE ciphertexts) but also concretely efficient. We implement our scheme as a C++ library and show that it takes $< 5$ms per LWE ciphertext to bootstrap for a binary gate, which is an order of magnitude faster than the state-of-the-art C++ implementation on LWE ciphertext bootstrapping in OpenFHE. Furthermore, our construction supports batched arbitrary functional bootstrapping. For a 9-bit messages space, our scheme takes ${\sim}6.7$ms per LWE ciphertext to evaluate an arbitrary function with bootstrapping, which is about two to three magnitudes faster than all the existing schemes that achieve a similar functionality and message space.