RAMPARTS: RApid Machine-learning Processing Applications and Reconfigurable Targeting of Security

Fully Homomorphic Encryption (FHE) is a recent development in cryptography that allows computation on data while that data remains encrypted. Largely theoretical until the last year or two, FHE is now becoming practical in terms of performance. However, FHE is still very complex, requiring deep expertise in cryptography and fully customized – and highly mathematical – implementations for each program to be executed. In contrast, many analytic frameworks today allow domain experts with very little programming background and no expertise in cryptography to create sophisticated analysis programs. Because FHE is still a relatively young discipline and there are very few experts capable of writing such programs, secure computation using this technique suffers from another drawback: lack of optimizations that offer efficiency improvements. Much like programming capabilities for the very first computers, frameworks for programming for FHE do not include optimizing compilers. In contrast, modern executables often gain one to two orders of magnitude in performance over non-optimized versions.

In the RAMPARTS project, our team of Galois and the New Jersey Institute of Technology aims to improve FHE from these “early days” of complexity and inefficiency to a future where FHE is easy to use and achieves performance improvements in line with those offered by modern optimizing compilers. We provide a programming model that requires very little cryptographic expertise, yet allows programmers to compute securely on sensitive data. We include initial optimization approaches to demonstrate the viability of automatic FHE optimization. We provide libraries of FHE primitives for use in code generation. We also provide additional tools such as performance estimators that allow for programmers to understand likely performance before committing to long analysis runs.

This slide show offers an overview of the RAMPARTS research seedling and our ideas and directions.

This research is based upon work supported in part by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA). The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies, either express or implied, of ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for governmental purposes notwithstanding any copyright annotation therein.