A federated training system, method, and device for a model that protects user identification

Abstract

Embodiments of the present invention provide a federated training system, method, and device for protecting user identification models. The system includes a first training device, a second training device, and a supervisory device. The first training device obtains a first user information sample, and the real The number of samples sent to the supervisory device, the first user information sample includes real samples and confused samples; the corresponding value of the confused samples in the encrypted forward gradient is set to 0, and the adjusted encrypted forward gradient is obtained; the second training device is adjusted according to Calculate the second sample gradient based on the encrypted forward gradient; the supervisory device calculates the first adjusted gradient and the second adjusted gradient based on the number of real samples; adjusts the first model parameters according to the first adjusted gradient; adjusts the first adjusted gradient according to the second adjusted gradient. Two model parameters. Add obfuscated samples to the real samples, and set the value of the obfuscated samples in the encrypted forward gradient to 0, so as to achieve the purpose of improving the security of user information without affecting the training effect of the model.

Description

technical field [0001] The invention relates to the technical field of model training, in particular to a federated training system, method and device for protecting user-identified models. Background technique [0002] The federated training of the model is a model training method in which two data providers use their own data to jointly train the model, and it is widely used. For example, federated training of models in financial institutions such as banks and insurance and external data partners. [0003] In the federated training process of the model, firstly, the two data providers need to align samples according to the sample identification, and then carry out model training based on the aligned samples. The existing method is that the model training initiator and the external data partner use the Private Set Intersection technology to align the sample identification, and then realize the alignment of the sample. In this way, the external data partner will obtain the ...

AI Technical Summary

Problems solved by technology

Then there may be a risk of user information being leaked

Images