一种基于秘密共享的神经网络训练方案

A Scheme for Neural Network Training Based on Secret Sharing

人工神经网络在数据学习上具有鲁棒性,已成功应用于语音识别、机器人控制、自动驾驶、人脸识别等。构建一个好的学习系统既需要卓越的算法,更需要大量的数据样本。然而,训练数据是分布式存储,并具有隐私性。构建具有隐私保护的神经网络模型是亟待研究的问题。目前,基于同态的隐私保护神经网络学习方案给用户和服务器带来了巨大的计算负担和通信负担。该文基于秘密共享技术,提出了一个高效的具有隐私保护的人工神经网络训练方案,保证用户数据隐私的情况下构建一个高精度模型。数值实验表明:该方案能够保障用户数据的安全,并且具有较低的计算量。

The artificial neural network is robust in data learning,and it has been successfully applied to speech recognition,robot control,automatic driving,face recognition,etc.Building a good learning system requires both excellent algorithms and a large number of data samples.However,training data is distributed storage and private,so it is an urgent problem to construct the neural network model with privacy protection.At present,the neural network learning scheme based on homomorphic privacy protection brings great computing and communication burdens to users and servers.Based on secret sharing technology,this paper proposes an efficient artificial neural network training scheme with privacy protection to construct a high-precision model under the condition of ensuring the privacy of user data.Numerical experiments show that the scheme guarantees the security of user data and has low computation.