Fully Homomorphic Encryption (FHE) has gained increasing importance mainly due to its potential use in privacy-preserving cloud computing. This privacy stems from the computation being directly performed on data that is encrypted by the client. However, FHE comes with a major cost regarding computational requirements. When compared to processing in the unencrypted domain, the time it takes can be up to four orders of magnitude higher, which is particularly inconvenient for applications with time constraints. Hence, accelerating FHE is a major research line, by leveraging different mathematical schemes and algorithms to the use of specialized hardware accelerators targeting the most time-consuming operations. This paper targets the optimization of FHE by leveraging vectorized implementations in RISC-V processors, using the RISC-V Vector (RVV) extension. In particular, it implements and accelerates the Open-Source FHE library, OpenFHE, optimizing its Number Theoretic Transform (NTT) and the Inverse-NTT (INTT) components. In this library, different FHE algorithms (BGV, BFV, CKKS) were analyzed, optimized, and tested. For the NTT and INTT operations, a maximum speedup of 27.05x was obtained. Furthermore, for a multiplication with bootstrapping benchmark program in OpenFHE, a speedup of 1.94x for the CKKS scheme was attained. Additionally, neural network benchmarks exhibit a speedup of over 1.69x.
Rodrigues et al. (Fri,) studied this question.