Federal learning incentive method and system based on license chain

Abstract

The invention discloses a federal learning incentive method and system based on a license chain, and relates to the technical field of blockchains, and the method comprises the steps that the registration and authentication of clients is carried out, the registration of each client is carried out to the license chain, and the authentication and certificate issuing of each client is carried out through the license chain; the smart contract of the license chain runs, and sampling is carried out from a group of clients meeting qualification requirements; the client downloads the training model and the program from the license chain; the client updates parameters of the model by executing a training program in local calculation, encrypts the updated parameters and uploads the encrypted parameters to the license chain; the license chain node receives the data encrypted by the client, decrypts the data and verifies the correctness of the data; the license chain node carries out consensus, after the consensus is passed, the reputation value and the contribution value of the client are calculated, and a new block is generated; the intelligent contract aggregates the model parameters and updates the parameters; and the smart contract judges whether a preset convergence condition of the model is met, if not, the next round of training is carried out, if yes, training is terminated, and excitation is issued according to the contribution value of the client. According to the method and system, the license block chain and the intelligent contract technology are applied, the problem that a federal learning malicious client or participants damage the correctness of training by utilizing wrong gradient collection and parameter updating is solved, an incentive mechanism is provided, the enthusiasm of providing data and updating network model parameters among the participants is increased, and meanwhile, the security of private data is improved.

Description

technical field [0001] The present invention relates to the blockchain field, in particular to a permission chain-based federal learning incentive method and system. Background technique [0002] A permissioned chain is a blockchain that is jointly managed by several organizations. Each organization runs one or more nodes. Only authorized nodes can participate in voting, bookkeeping, and building blocks. Traditional machine learning centralizes data to the server and trains corresponding models by running machine learning algorithms. At present, as users pay more attention to privacy protection, such algorithms are facing huge privacy challenges. However, federated learning keeps user data locally and only collects model parameters, thereby greatly reducing the risk of user data leakage. [0003] The federated learning system is a learning method in which multiple clients contribute local data and cooperate to train a unified model. The limitation is that the model only re...

AI Technical Summary

Problems solved by technology

[0003] The federated learning system is a learning method in which multiple clients contribute local data and cooperate to train a unified model. The limitation is that the model only relies on a single central server, which is vulnerable to server failure and how to fairly evaluate the participation of each participant. how to protect the privacy of data

Images