Collaborative Research: SHF: Small: Efficient and Scalable Privacy-Preserving Neural Network Inference based on Ciphertext-Ciphertext Fully Homomorphic Encryption

随着人工智能的发展，隐私保护机器学习已经成为云应用中保护用户数据隐私的一种重要而有前途的技术。在现有的方法中，基于完全同态加密(FHE)的方法允许机器学习算法在加密的数据上计算，而不会泄露原始数据信息。这个项目致力于保护用户和模型提供者的隐私的密文-密文。本项目旨在通过算法-硬件联合优化，将基于密文-密文的神经网络推理的硬件效率提高一个数量级。该项目为云计算和密码系统提供了一种新的可信框架，以满足未来商业产品和国防的需求。该项目开发了基于密文-密文的隐私保护神经网络推理的高效和可扩展的硬件体系结构。该项目利用方案切换--对线性函数使用基于算术的方案，对非线性函数使用基于布尔逻辑的方案--来加速神经网络计算。研究内容包括：a)采用新颖的可重构流水线结构和利用特殊素数进行快速模约简，为隐私保护神经网络设计了具有高可伸缩性的基本硬件模块，即多项式乘法器；b)采用基于新的并行滤波技术的分而治之策略，进一步提高了多项式乘法器的设计效率；c)利用方案切换，开发了一种可重构的神经网络友好的乘法器结构；以及d)通过方案切换，通过密文-密文操作设计保护隐私的神经网络推理的高效加速器，以保护用户和模型的隐私。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Along with the evolution of artificial intelligence, privacy-preserving machine learning has emerged as an important and promising technique for protecting user-data privacy in cloud applications. Among the existing approaches, fully homomorphic encryption (FHE) based methods allow machine learning algorithms to be computed on encrypted data, while no original data information is leaked. This project addresses ciphertext-ciphertext FHE that preserves the privacy of both the user and model providers. This project aims to improve the hardware efficiency of ciphertext-ciphertext FHE-based neural network inference by orders of magnitude through algorithm-hardware co-optimization. This project yields a novel framework for ensuring the root of trust in cloud computing and cryptosystems to meet the future needs of both commercial products and national defense.This project develops efficient and scalable hardware architectures for privacy-preserving neural network inference based on ciphertext-ciphertext FHE. This project leverages scheme switching - using arithmetic-based schemes for linear functions and Boolean logic-based schemes for non-linear functions - to accelerate the neural network computations. Research thrusts include: a) Designing efficient fundamental hardware building blocks with high scalability over word-length of modulus and degree of polynomial for privacy-preserving neural network, i.e., polynomial multipliers, by employing novel reconfigurable and pipelining framework and exploiting special primes to perform fast modular reduction; b) Further improving the efficiency of polynomial multiplier designs by utilizing a divide and conquer strategy based on a novel parallel filter technique; c) Developing a reconfigurable and neural network friendly FHE architecture using scheme switching; and d) Designing an efficient accelerator of privacy-preserving neural network inference with ciphertext-ciphertext operations via scheme switching that protects the privacy of both the user and model.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.