Elements: Portable Library for Homomorphic Encrypted Machine Learning on FPGA Accelerated Cloud Cyberinfrastructure

University Of Southern California, Los Angeles CA

Investigators

Abstract

Privacy Preserving Computations (PPC) that utilize Homomorphic Encryption (HE) to perform computations on encrypted data directly have become attractive recently. HE based Machine Learning (HE ML) inference enables preservation of privacy in a wide variety of application domains that range from healthcare, financial transactions, edge cyber physical systems, etc. Privacy sensitive applications that rely on processing in a public cloud or data center can use HE ML inference to preserve privacy. While HE ML inference offers strong privacy guarantees, computations on encrypted data are orders of magnitude slower than unencrypted computations and require significant hardware resources to make them attractive for end users. Emerging data centers and cloud platforms are augmented with Field Programmable Gate Arrays (FPGAs). With the fine grained programmable architecture of FPGAs, these platforms are well suited for accelerating HE ML. This work will leverage novel algorithmic, architectural and memory optimizations on FPGAs to develop a portable and configurable library to enable secure, resilient, and trustworthy cyberinfrastructure for end-to- end privacy sensitive ML inference. The library will provide FPGA accelerated Intellectual Property (IP) cores for HE kernels (L1 Library) as well as a FPGA Application Specific Processor (ASP) for inference of widely studied HE ML models (L2 Library). The library will support various HE schemes, security levels, machine learning models and FPGA platforms. It will include several software and hardware innovations along with various HE specific optimizations such as efficient data layout, memory efficient scheduling and scalable interconnect to maximize memory utilization and to improve data reuse using on-chip memory. Using the IP cores in the L1 Library, this project will compose a FPGA ASP with a domain-specific Instruction Set Architecture (ISA) and a compiler. The FPGA accelerator can be programmed in software to realize real-time HE ML computations. The library will be released to the Computer & Information Science & Engineering (CISE) communities, including Machine Learning, Software, and Data Science communities, to accelerate the adoption of homomorphic encryption for privacy preserving computations. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.