Deep Neural Network Model over Encrypted Data

Weixun Li, Guanghui Sun, Yajun Wang, Long Yuan, Minghui Gao, Yan Dong, Chen Wang

Abstract: Deep Neural Networks (DNN) model training usually requires a large amount of data as the foundation, so that the model can learn effective features and rules. However, these data often contain sensitive information of users. This paper designs a DNN Classification Model for Ciphertext Data (DNN-CMED), which consists of three-party entities, including two servers and one client. The auxiliary server assists the model training server in completing the computation of the nonlinear layer of the DNN model, and the two servers interact with each other to complete the task of classifying the ciphertext data. The communication protocols of DNN-CMED are designed, including secure linear computation protocol and secure nonlinear computation protocol. The classification process of DNN-CMED is given based on the above protocol. Through safety analysis and experiments, it shows that the models have better security and practicality. The results show that this model has better results in terms of accuracy, time and communication overhead, and on the MNIST test set.