Neural network model training method based on shared learning

Abstract

The invention relates to the technical field of machine learning, and aims to provide a neural network model training method based on shared learning. According to the method, a plurality of shared learning participants perform hybrid security multi-party calculation and central node local calculation through secret sharing to realize neural network model training; the method specifically comprises the steps of performing data preparation; loading and initializing data; loading the current training batch; calculating a first-layer linear output; calculating a model prediction value; calculating a model output gradient; updating the model and calculating; updating the local sharing weight; performing assessment on a test set. According to the method, multi-party security calculation and local calculation are combined, and the shared learning efficiency is improved. Through an online encryption multiplication triad provider, the efficiency of secure multi-party calculation of secret sharing is improved, and the overhead is greatly reduced compared with schemes of homomorphic encryption, confusion circuits and the like. The method is suitable for a longitudinal shared learning scene,and can enable all parties to carry out federated modeling under the condition that no overlapping features exist.

Description

technical field [0001] The invention relates to the technical field of machine learning, in particular to a training method of a neural network model based on shared learning. Background technique [0002] With the widespread application of machine learning algorithms, a lot of data is used for the training of machine learning models. For example, a bank predicts the credit of other customers by training the data of some customers. The shared learning method is a privacy-preserving machine learning method that integrates local computing and multi-party secure computing, and is widely used in machine learning scenarios that require privacy protection. For example, one bank has the user's loan data, another bank has the user's transaction data, and the two banks want to share data for modeling and predict the user's credit level. But if the data is directly put together and trained with a neural network model, the data will leave at least one bank, thus bringing the risk of ...

AI Technical Summary

Problems solved by technology

However, the trusted execution environment has problems such as high cost and many restrictions on running programs, and cannot well support multiple algorithms.

[0007] There are other existing technologies, although they have solved the specific implementation problems of privacy-preserving machine learning to a certain extent, but there are still problems such as small scope of application, high computing and communication overhead, unstable model convergence, and lack of theoretical foundation.

Images