《计算机应用》唯一官方网站

• •    下一篇

支持隐私保护训练的高效同态神经网络

钟洋1,2,毕仁万1,2,颜西山1,应作斌3,熊金波1,2   

  1. 1. 福建师范大学 计算机与网络空间安全学院,福州 350117;
    2. 福建师范大学 福建省网络安全与密码技术重点实验室,福州 350007;
    3. 澳门城市大学 数据科学学院,澳门 999078
  • 收稿日期:2021-10-15 修回日期:2022-01-18 接受日期:2022-01-24 发布日期:2022-03-04 出版日期:2022-03-04
  • 通讯作者: 熊金波
  • 基金资助:
    国家自然科学基金;福建省自然科学基金资助项目

Efficient homomorphic neural network supporting privacy-preserving training

  • Received:2021-10-15 Revised:2022-01-18 Accepted:2022-01-24 Online:2022-03-04 Published:2022-03-04
  • Supported by:
    the National Natural Science Foundation of China;Project supported by Fujian Natural Science Foundation

摘要: 针对基于同态加密的隐私保护神经网络中存在的计算效率低和精度不足问题,提出一种三方协作下支持隐私保护训练的高效同态神经网络(HNN)。首先,为降低同态加密中密文乘密文运算产生的计算开销,结合秘密共享思想设计一种安全快速乘法协议,将密文乘密文运算转换为复杂度较低的明文乘密文运算;其次,为避免构建HNN时产生的密文多项式多轮迭代,并提高非线性计算精度,研究一种安全的非线性计算方法,对添加随机掩码的混淆明文消息执行相应的非线性算子;最后,对所设计协议的安全性、正确性及效率进行理论分析,并对HNN的有效性及优越性进行实验验证。实验结果表明,相比于双服务器方案PPML,HNN的训练速度提高18.9倍,模型精度提高1.4个百分点。

关键词: 同态加密, 隐私保护, 神经网络, 模型训练, 非线性计算

Abstract: Aiming at the problems of low computational efficiency and insufficient accuracy in the privacy-preserving neural network based on homomorphic encryption, an efficient Homomorphic Neural Network (HNN) supporting three-parties-collaborative privacy-preserving training was proposed. Firstly, in order to reduce the computational cost of ciphertext-ciphertext-multiplication in homomorphic encryption, combined with the idea of secret sharing, a secure fast multiplication protocol was designed to convert the ciphertext-ciphertext-multiplication operation in homomorphic encryption into plaintext-ciphertext-multiplication operation with low complexity. Then, In order to avoid multiple iterations of ciphertext polynomials generated during the construction of HNN and improve the nonlinear calculation accuracy,, a secure nonlinear calculation method was studied, which executed the corresponding nonlinear operator for the confused plaintext message with random mask. Finally, the security, correctness and efficiency of the proposed protocols were analyzed theoretically. Experimental results show that compared with the dual server scheme PPML(Privacy Protection Machine Learning), the training efficiency of HNN is improved by 18.9 times and the model accuracy is improved by 1.4 percentage points.

Key words: homomorphic encryption, privacy-preserving, neural network, model training, non-linear calculation

中图分类号: