METHOD AND DEVICE FOR USING TRUSTED EXECUTION ENVIRONMENT TO TRAIN NEURAL NETWORK MODEL

摘要

A method and device for using a trusted execution environment to train a neural network model. In the method, a neural network model is sequentially divided into a first-part neural network model located in a trusted execution environment of a first apparatus, and a second-part neural network model located in an untrusted execution environment of a second apparatus. The method comprises: in each round of model training, a current first-part neural network model processing training sample data in a trusted execution environment so as to acquire an intermediate result (530); a current second-part neural network model processing the intermediate result in an untrusted execution environment so as to acquire a current prediction value, and determining a prediction difference (550); and when a loop end condition is not met, adjusting a model parameter of each layer of the current first-part neural network model and a model parameter of each layer of the current second-part neural network model according to the current prediction difference (570). By using the method, a neural network model is trained while ensuring the security of private data.