A model update method, device and equipment

Abstract

Disclosed are a model updating method, device and equipment. Through the solution provided by the embodiment of this specification, given the shadow set S and the target sample, the prior probability α of the training sample included in the shadow set is calculated respectively, and when sampling and training the model according to the shadow set S, the The sampling training parameter distribution of the obtained sampling model, and then calculate the posterior probability P of the target sample z in the shadow set according to the training parameter distribution of the given model and the eigenvalues ​​of the target sample, so that the posterior probability P and the prior The difference of the probability α is used to evaluate whether the trained model is qualified for the degree of privacy leakage of the training sample set. Make model adjustments to avoid privacy data leakage.

Description

technical field [0001] The embodiments of this specification relate to the field of information technology, and in particular, to a model updating method, device, and equipment. Background technique [0002] In the process of machine learning, machine learning model algorithms based on data training often face the problem of leaking user privacy in training data. For example, in a membership attack against a model, the attacker of the membership attack is given a trained model and one or more specific samples, trying to determine whether the sample is used in training the model, In order to reversely try to derive the training method and recognition method of the model. [0003] Based on this, a model update scheme that can avoid privacy leakage in machine learning is needed. Contents of the invention [0004] The purpose of the embodiments of the present application is to provide a model update scheme that can avoid privacy leakage in machine learning. [0005] In orde...

AI Technical Summary

Problems solved by technology

[0002] In the process of machine learning, machine learning model algorithms based on data training often face the problem of leaking user privacy in training data

Images