Safe and efficient deep learning model prediction system and method based on SGX

Abstract

The invention discloses a safe and efficient deep learning model prediction method based on SGX. The method comprises: S1, a model provider terminal uploading a deep learning model through a model import module, and the model import module converting deep learning models of different formats into models capable of being executed in a deep learning calculation framework according to the deep learning calculation framework in the SGX; S2, the model user terminal and an RPC server module in the SGX carrying out key negotiation to obtain a communication key, and the communication key being used for encrypting to-be-predicted data provided by the model user terminal and a prediction result of a deep learning model in the SGX; and S3, the model user terminal signing and encrypting the to-be-predicted data by using the communication key and decrypting and verifying the prediction result. By means of the SGX, the deep learning model is operated based on the SGX, so that the confidentiality andintegrity of the data in the model operation process are ensured, and meanwhile, the confidentiality and integrity of the model can also be ensured.

Description

technical field [0001] The present invention relates to the technical field of machine learning security technology, in particular to a safe and efficient deep learning model prediction system and method based on SGX. Background technique [0002] Machine learning, especially deep learning, as a representative of the field of artificial intelligence, has shown excellent performance in fields including image recognition and speech translation, so it has been widely used in related application scenarios to solve a series of practical problems , including identity authentication, gait recognition, etc. [0003] However, the deep learning model faces various security problems in the process of deployment and use. These security problems exist between the deep learning model and the user, and between the user and the server where the deep learning model is located. For example, when data is transmitted between the user and the model, there may be a third party eavesdropping on t...

AI Technical Summary

Problems solved by technology

Tople et al. proposed to complete the prediction process of deep learning in SGX, but this solution did not consider the data security between users and machine learning models; Lucjan et al. proposed a method to complete model prediction in SGX, but the solution is to use The model is placed and run in the SGX environment on the client side without using the computing resources of the server

This will not only idle the computing resources of the server, but also increase the communication cost between the client and the server

At the same time, all these solutions only support the model to make the CPU complete the prediction process, and do not consider the use of the GPU, so the efficiency is relatively low compared to the conventional deep learning model operation method

Images