Federal learning method and system combined with personalized differential privacy

Abstract

The invention discloses a federal learning method and system combined with personalized differential privacy. The method comprises the following steps: receiving a global model of the current round sent by a server; after the global model of the current round is received, using a privacy budget allocation mechanism to obtain privacy parameters of the current round; based on the privacy parameters of the current round, sampling a local data set by using a sampling mechanism; training the global model of the current round on the sampled data set to obtain local model update of the current round; performing noise addition processing on the local model update by applying a Gaussian mechanism to obtain the local model update after noise addition processing; and updating and uploading the local model subjected to noise addition processing to the server, so that the server updates based on the local model subjected to noise addition processing to obtain an updated global model. According to the method, during federal learning, the influence of a differential privacy mechanism on the model effect can be reduced as much as possible under the condition that the personalized privacy requirement of the user is met.

Description

technical field [0001] The invention relates to the technical field of federated learning, in particular to a federated learning method and system combined with individualized differential privacy. Background technique [0002] With the rapid development of mobile computing, edge and mobile devices are generating massive amounts of data. At present, due to issues such as limited network bandwidth and data privacy, it is impossible to upload all data to the cloud for further processing. The emergence of federated learning makes it possible for end users to jointly train network models without directly uploading their own local data. [0003] The protection of data by federated learning is based on the assumption that model gradient information will not leak data privacy. However, there is currently a large body of work pointing out that model parameters expose dataset information. For example, some work pointed out that by analyzing the gradient information uploaded by the...

AI Technical Summary

Problems solved by technology

[0008] (2) The introduction of disturbance is relatively large and needs to be optimized

[0009] The differential privacy mechanism will introduce disturbances, which will affect the accuracy of the model to a certain extent

Most of the existing solutions directly use the Gaussian mechanism and are not optimized for federated learning scenarios, which may lead to a large loss of accuracy in the model

Images