Scholarlink

Fully-Homomorphic Encryption (FHE) has been touted as the ultimate solution for preserving user data privacy in Machine Learning as a Service (MLaaS) applications. Under this scheme, the server never sees the user data in clear, but executes the ML model on encrypted data instead. Unfortunately, efficient FHE schemes only support addition and multiplication operations, which cannot exactly represent common activation functions in neural networks. Substituting activation functions with polynomial approximations may cause significant output deviations. In this paper, we propose ZonoPoly, an efficient algorithm to compute guaranteed bounds on the maximum output deviation of FHE neural networks. We implement our algorithm in the VeryDiff framework by extending its zonotope-based reachability analysis primitives to support high-degree polynomials. Experimental results show that ZonoPoly produces approximately \(3\times \) tighter bounds than existing methods in most cases.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Certified Error Analysis of Homomorphically Encrypted Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Philipp Kern,
  • Edoardo Manino,
  • Carsten Sinz

摘要

Fully-Homomorphic Encryption (FHE) has been touted as the ultimate solution for preserving user data privacy in Machine Learning as a Service (MLaaS) applications. Under this scheme, the server never sees the user data in clear, but executes the ML model on encrypted data instead. Unfortunately, efficient FHE schemes only support addition and multiplication operations, which cannot exactly represent common activation functions in neural networks. Substituting activation functions with polynomial approximations may cause significant output deviations. In this paper, we propose ZonoPoly, an efficient algorithm to compute guaranteed bounds on the maximum output deviation of FHE neural networks. We implement our algorithm in the VeryDiff framework by extending its zonotope-based reachability analysis primitives to support high-degree polynomials. Experimental results show that ZonoPoly produces approximately \(3\times \) tighter bounds than existing methods in most cases.