Deep neural network encryption reasoning method based on homomorphic encryption technology

Abstract

A deep neural network encryption inference method based on a homomorphic encryption technology comprises the following steps: transmitting to a cloud, combining a BN layer into a convolution layer by the cloud, carrying out homomorphic convolution operation to obtain a ciphertext feature map after the first layer of convolution operation, and transmitting to a client; after the client receives the data, a mark matrix being obtained through mark operation and transmitted to the cloud server, and after the cloud server receives the mark matrix, updating to obtain an input feature map of a second homomorphic convolution layer; performing second-layer homomorphic convolution operation to obtain a ciphertext feature map after the second-layer convolution operation, and transmitting the ciphertext feature map to the client; and repeating the above process until all the ciphertext feature maps after convolution layer operation are obtained. According to the method, the GPU is used for accelerating the homomorphic convolution operation process, and repeated transmission of data is avoided. Noise increase of the ciphertext can be reduced, the number of reasoning layers of the neural network is increased, and the calculation overhead of the ciphertext is greatly reduced.

Description

technical field [0001] The invention relates to a deep neural network encryption reasoning method based on homomorphic encryption technology, which can be used in a privacy data protection method in the field of artificial intelligence. Background technique [0002] As an important technology in the field of artificial intelligence, deep neural network has far surpassed traditional computer vision processing and recognition methods in applications such as image classification and recognition and video target tracking. However, the training and reasoning of deep neural networks for computer vision requires the collection of a large amount of user image data, which can easily involve user privacy. If these user data are leaked or mishandled, on the one hand, it is likely to cause user privacy. Exposure may even cause unpredictable property damage and life safety issues. On the other hand, with the further development of deep neural networks, more user data needs to be collect...

AI Technical Summary

Problems solved by technology

First of all, the number of layers of the deep neural network based on homomorphic encryption technology is not deep enough, mainly because homomorphic encryption can only do limited multiplication and addition operations when the encryption parameters are determined.

Secondly, the activation functions such as relu and sigmoid in the deep neural network are difficult to be supported by the existing homomorphic encryption framework. The previous work used quadratic function fitting, but this brought new problems. The multiplication between the encryption noise and the calculation time will increase exponentially and the accuracy of the deep neural network will decrease significantly.

Images