Full text
| |
| Author(s): |
Antonio Carlos Guimarães Junior
Total Authors: 1
|
| Document type: | Doctoral Thesis |
| Press: | Campinas, SP. |
| Institution: | Universidade Estadual de Campinas (UNICAMP). Instituto de Computação |
| Defense date: | 2023-09-01 |
| Examining board members: |
Edson Borin;
Julio César López Hernández;
Marcos Antônio Simplício Júnior;
Ricardo Dahab;
Jeroen Antonius Maria Van de Graaf
|
| Advisor: | Diego de Freitas Aranha; Edson Borin |
| Abstract | |
Enabling the evaluation of arbitrary computation over encrypted data raised fully homomorphic encryption (FHE) to the spotlight among privacy-preserving technologies. Initially, however, the feat had a deminute impact in practice, as most applications could not bear the performance overhead it introduced. The literature then evolved around the needs of specific use cases, often forfeiting arbitrary computation in the name of performance. Currently, state-of-the-art performance on FHE is represented by schemes that specialize in fast arithmetic while relegating arbitrary functions to polynomial approximations. On the other side of this issue, the TFHE scheme (FHE over the Torus) [Chillotti et al., 2016] upholds arbitrary computation while fighting for a competitive performance level. This work is dedicated to improving it. TFHE enables the evaluation of arbitrary functions through a technique called "functional bootstrapping", but its cost grows superlinearly with the function precision, which, at first, made it only suitable for functions with small precision. In this work, we introduced some of the first methods to allow a more efficient evaluation of arbitrary functions with high precision using TFHE. Our methods enabled speedups of up to 3.2 times over previous literature using the functional bootstrapping, and of up to 8.74 times compared to other evaluation methods. We also advanced TFHE by optimizing and proposing new techniques for some of its core procedures and their implementations. Among these improvements, we highlight an optimization on its basic arithmetic that achieves a speedup of up to 2 times over previous implementations and a new method for evaluating the functional bootstrapping with several functions at once (MISD, Multi-Instruction Single-Data, batching). Finally, with a focus on the practical side of FHE, we tested our contributions in a real-world scenario by implementing the homomorphic evaluation of an inference algorithm on human genome data (AU) | |
| FAPESP's process: | 19/12783-6 - Efficient migration of high-performance computing science and engineering applications to the cloud |
| Grantee: | Antonio Carlos Guimarães Junior |
| Support Opportunities: | Scholarships in Brazil - Doctorate |