Model training method and device based on block chain and federal learning, and block chain

Abstract

The invention discloses a model training method and device based on a block chain and federated learning, and the block chain, the block chain comprises a plurality of idle edge nodes, one of the idle edge nodes is an aggregation node, and the others are training nodes; the method comprises the steps that an aggregation node generates an initial global initial model; after each training node downloads, training based on local voice data to obtain a local voice recognition model, and transmitting the local voice recognition model to an aggregation node, so that the aggregation node performs aggregation according to an aggregation algorithm to obtain a global voice recognition model as a global initial model; when the global speech recognition model meets a preset iteration condition, the global speech recognition model is downloaded by each training node; and each idle edge node calculates parameters required by secondary training based on the local voice data and uploads the parameters to the block chain, secondary training is performed according to the local voice data and the parameters, and a global voice recognition model is deployed. According to the method, the security of user privacy data is ensured, the model obtained by federal learning training is subjected to personalized transformation, and the use precision of the local deployment model is improved.

Description

technical field [0001] This application relates to the field of computer technology, in particular to a model training method, device and blockchain based on blockchain and federated learning. Background technique [0002] Speech recognition technology focuses on converting the lexical content of human speech into computer-readable input. The long-term development of speech recognition technology is inseparable from the support of speech data. [0003] At this stage, there are many public speech datasets. However, with the progress of the times and changes in people's thinking, there will always be new words being created continuously, and some old words will also derive new meanings. Therefore, a large amount of fresh speech data is needed to continuously support the development of speech recognition technology. [0004] When collecting and updating voice data, grabbing data directly from the user's local data set will involve the issue of privacy violations. This probl...

AI Technical Summary

Problems solved by technology

This problem can be avoided by using federated learning, but the model trained according to federated learning cannot meet the individual needs of each device

Therefore, in the speech recognition scenario, there is a lack of a model training method that not only takes into account information security, but also adapts to the individual needs of users.

Images