Outsourcing convolutional neural network privacy protection system based on secure two-party calculation

Abstract

The invention discloses an outsourcing convolutional neural network privacy protection system based on secure two-party calculation, which randomly allocates a neural network and query data to two servers which are not communicated with each other, and realizes image processing through cooperative operation on the two servers. The two image components are input into two servers; information transmitted between servers is hidden in a convolution layer by using triples of random data. A confusion circuit is designed on an active layer to realize a ReLU function, then dimensionality reduction ofan image is carried out through average pooling, finally, two components of a prediction result are obtained through a full connection layer adopting triple hidden information, and the two componentsare returned to a user to be combined to obtain the obtained prediction result. In order to improve the calculation efficiency, an asynchronous calculation method and a parallel query method are adopted, independent calculations in the same query are carried out at the same time, and different parts of a plurality of queries are processed at the same time, so that the time of each query is greatlyreduced.

Description

technical field [0001] The invention belongs to the technical field of information security, and in particular relates to a privacy-protected neural network outsourcing method and system based on secure two-party computing. Background technique [0002] With breakthroughs in the field of deep learning, neural networks have demonstrated powerful advantages over previous techniques in terms of predictive capabilities. Among a variety of neural network models, convolutional neural networks stand out due to their good performance, and have been widely used in medical image analysis, natural language processing, and text recognition. For cost considerations, some service providers will deploy the model to an external server and provide subsequent prediction services. However, service outsourcing is accompanied by the risk of privacy leakage. Existing protection schemes can only protect query data and prediction results from being leaked, but cannot protect the parameters of comp...

AI Technical Summary

Problems solved by technology

However, service outsourcing is accompanied by the risk of privacy leakage. Existing protection schemes can only protect query data and prediction results from being leaked, but cannot protect the parameters of complex neural network models.

Images