Method for protecting deep learning model based on confidential computing

1. A method for protecting a deep learning model based on confidential computing, wherein a preprocessing module having a data preprocessing model and an inference module having an inference model are comprised, the data preprocessing model is encrypted and deployed into a confidential computing environment; and

the method comprises:

S1: starting the preprocessing module in the confidential computing environment, and sending attestation information of the confidential computing environment to a model copyright owner server;

S2: receiving a attestation result returned by the model copyright owner server, and obtaining a decryption key for decrypting the data preprocessing model if the attestation result is correct, or exiting running if the attestation result is incorrect;

S3: decrypting the data preprocessing model by using the obtained decryption key, wherein the preprocessing module loads the decrypted data preprocessing model;

S4: preprocessing, by the data preprocessing model, to-be-inferred data submitted by an authorized user, and sending the preprocessed data to the inference model; and

S5: performing, by the inference model, inference on the preprocessed data, and then sending an inference result to the authorized user;

wherein the preprocessing module and the inference module are distributed to the authorized user in advance and are deployed on an authorized user end, and the preprocessing module and the inference module are distributed to the authorized user in the following manners:

encrypting a trained data preprocessing model by using an encryption algorithm;

packaging the encrypted data preprocessing model and preprocessing code as the preprocessing module;

packaging a trained inference model and inference code as the inference module;

distributing the preprocessing module and the inference module to the authorized user; and

deploying, by the authorized user, the preprocessing module to the confidential computing environment and deploying the inference module to a common computing environment; and

wherein a training method of the data preprocessing model comprises:

running the data preprocessing model and a pretraining model, randomizing parameter information in the models, and combining loss functions of the data preprocessing model and the pretraining model;

processing a data set by using the data preprocessing model, and sending preprocessed data to the pretraining model;

sending unprocessed original data to the pre-training model; and

training the pretraining model by using the original data and the preprocessed data, and selecting a combination of the data preprocessing model and the pretraining model with highest prediction accuracy.

2. The method for protecting a deep learning model based on confidential computing according to claim 1 , wherein in step S5, before the inference model is used, a model parameter is first loaded and the model is initialized; and

the preprocessing module comprises a confidential computing environment initialization module, and in step S1, the confidential computing environment initialization module is started in the confidential computing environment, and the confidential computing environment initialization module sends the attestation information of the confidential computing environment to the model copyright owner server; and in step S3, the confidential computing environment initialization module loads the decrypted data preprocessing model.

3. The method for protecting a deep learning model based on confidential computing according to claim 1 , wherein the inference model is deployed in a common computing environment.

4. The method for protecting a deep learning model based on confidential computing according to claim 1 , wherein in step S1 and step S2, the model copyright owner server verifies security of the confidential computing environment and integrity of the preprocessing module for the authorized user based on a remote attestation mechanism of the confidential computing environment, if a attestation result is safety and complete, the attestation result comprising the decryption key is returned to the confidential computing environment, and the attestation result is correct in this case; otherwise, the attestation result comprising no decryption key is returned, and the attestation result is incorrect in this case.

5. The method for protecting a deep learning model based on confidential computing according to claim 1 , wherein a training method of the inference model comprises:

running the trained data preprocessing model and the pretraining model, and fixing model parameter information;

running the inference model and setting an optimizer;

processing the data set by using the data preprocessing model;

sending the preprocessed data to the pretraining model and the inference model respectively, and sending the unprocessed original data to the inference model; and

training the inference model by using the original data and the preprocessed data, and optimizing a parameter of the inference model by using a prediction result of the pretraining model.

6. The method for protecting a deep learning model based on confidential computing according to claim 1 , wherein the data preprocessing model and the pretraining model share one loss function during training.

7. The method for protecting a deep learning model based on confidential computing according to claim 5 , wherein the inference model and the pretraining model have a same network structure.