Encryption and decryption method and system of network model

Abstract

The invention discloses an encryption and decryption method and system of a network model. The encryption method comprises the following steps: extracting a network structure definition file in a network model file; and encrypting the network structure definition file through the first secret key and an encryption algorithm; and extracting a network variable name file in the network model file; and encrypting the network variable name file through the second secret key and the encryption algorithm; and generating an encrypted network model file according to the encrypted network structure definition file and the encrypted network variable name file. According to the method, part of files in the network model file are encrypted without a secret key, so that the security is improved, and theencryption process is simplified. And later use of the model, such as updating, is more convenient.

Description

technical field [0001] The invention belongs to the technical field of network application and deep learning, and in particular relates to an encryption and decryption method and system of a TensorFlow model. Background technique [0002] In related technologies, most deep network models are stored in standard serialized probuf format, which means that anyone can copy your model. A common way to protect a model from being easily copied is to encrypt the model, load the model in the application, and then decrypt the encrypted model in memory. [0003] The commonly used model encryption algorithm is the symmetric encryption algorithm in reversible encryption. The most commonly used model encryption scheme is to use a fixed password to encrypt the saved model file, and then decrypt the encrypted model during the inference process. There is also a method of solidifying the network model file first, solidifying the weight variables into the network structure, simplifying the mo...

AI Technical Summary

Problems solved by technology

As a result, the encryption and decryption process is cumbersome and the security is average.

The main disadvantage of the method of encrypting the model parameters after solidification is that only one file needs to be encrypted but the security is low. Once the encrypted model is deciphered, the model is easier to copy and spread

It can be seen that the common problems of the model encryption methods in the prior art are that the security is not high enough, and when the model is updated, it is necessary to re-encrypt and replace and update all the model files, and the process is cumbersome

[001

Images