In the domain of Privacy-Preserving Machine Learning (PPML), Fully Homomorphic Encryption (FHE) is often used for encrypted computation to allow secure and privacy-preserving outsourcing of machine learning modeling. While FHE enables encrypted arithmetic operations, execution of programmatic logic such as control structures or conditional programming have remained a challenge. As a result, progress in encrypted training of PPML with FHE has been relatively stagnant compared to encrypted inference owing to the considerably higher logical complexity required in training. In addition, prior works that have demonstrated encrypted training use Interactive Rounds of Decryption and Evaluation (IRDE), where certain operations are decrypted and evaluated in plaintext using interactive rounds between the untrusted computing party (server) and the trusted private-key owner (client). In decision tree training for example, the current state-of-the-art requires d-rounds of IRDE for tree-depth of d. To address this issue in PPML and FHE, we introduce the Blind Evaluation Framework (BEF), a cryptographically secure programming framework that enables blind, but correct, execution of programming logic without IRDE. This is achieved by deconstructing programming logic into binary circuits and binary arithmetic to find alternative representations of logical statements, and adopting them to FHE for secure logical programming. To the best of our knowledge, this is the first framework to enable both training and inference of PPML models with FHE without decryption rounds. By advancing the state-of-the-art in IRDE efficiency by eliminating IRDE entirely, BEF enables adoption of FHE in use-cases where large amounts of computing services are available without the ability to have trusted clients available to perform decryption rounds.

翻译：在隐私保护机器学习（PPML）领域，全同态加密（FHE）常被用于加密计算，以实现机器学习建模的安全、隐私保护外包。尽管FHE支持加密算术运算，但诸如控制结构或条件编程等程序化逻辑的执行仍面临挑战。因此，与加密推理相比，基于FHE的PPML加密训练进展相对停滞，原因在于训练所需的逻辑复杂度显著更高。此外，先前展示加密训练的工作均采用了交互式解密与评估轮次（IRDE），即某些操作需在不可信计算方（服务器）与可信私钥持有方（客户端）之间通过交互轮次进行解密并以明文形式评估。例如，在决策树训练中，当前最先进方法需要d轮IRDE来处理深度为d的决策树。为解决PPML与FHE中的这一难题，我们提出了盲评估框架（BEF），这是一种密码学安全的编程框架，能够在无需IRDE的情况下实现编程逻辑的盲化但正确执行。通过将编程逻辑分解为二进制电路与二进制算术，寻找逻辑语句的替代表示，并将其适配至FHE以实现安全的逻辑编程。据我们所知，这是首个无需解密轮次即可支持PPML模型训练与推理的FHE框架。通过彻底消除IRDE，BEF推动了IRDE效率的最新进展，使其能够应用于大规模计算服务可用但缺乏可信客户端执行解密轮次的场景。