A personalized federated learning identification method and system based on differential privacy

Abstract

The application discloses a kind of based on personalized federal learning identification method and system of differential privacy, the method is executed in client, comprising: obtaining initialization model parameter, and it is loaded into pre-built local model;Based on privacy budget, using personalized differential privacy algorithm is carried out noise processing to the local model of loading parameter, and the shared layer model parameter of the local model after noise processing is sent to server;The shared layer model parameter after aggregation is loaded into the shared layer of the local model after noise processing, then local fine-tuning is obtained personalized model;The personalized model is repeatedly trained until reaching global iteration number, and the personalized federal learning test model based on differential privacy is obtained.The client of the application can process local model according to the privacy budget selected by customer, to adapt to different privacy needs of different users, while combining Adam algorithm, reduce the influence of added noise on model accuracy.

Images