基于同态加密的密态卷积神经网络预测模型的构建与实现

Construction and Implementation of Encrypted Convolution Neura l Network Prediction Model Based on Homomorphic Encryption

人工智能的兴起,带动了深度神经网络的研究热潮,然而“裸奔”的数据使得用户隐私面临巨大威胁。同态加密技术具备密文计算的属性,可以为深度神经网络提供有效的隐私保护。针对卷积神经网络在预测过程中的数据安全问题,结合CKKS同态加密方案,提出了密态卷积神经网络模型的构建方法,保证用户数据在预测过程中的全程密态。此外,结合MNIST标准数据集,完成了密态预测模型的仿真实现。实验结果表明,该模型的预测准确率可以达到98.3%,并且单幅图像预测时间低至毫秒级别。

The rise of artificial intelligence has led to an upsurge of research on deep neural networks.However,the directly exposed data makes users’privacy face a great threat.Homomorphic encryption has the properties of calculating on ciphertexts,which can provide effective privacy protection for deep neural networks.Aiming at the data security problem of CNN in the prediction process,combined with CKKS scheme,this paper puts forward the construction method of encrypted CNN model to ensure user data keep encrypted in the whole prediction process.Besides,combined with the MNIST standard data set,it completes the simulation of encrypted prediction model.Experimental results indicate that the encrypted model can achieve 98.3% accuracy,and the prediction time of a single image is reduced to millisecond.