A Blockchain-based Federated Learning With Training Behavior Verification Mechanism

As a distributed deep learning framework,federated learning can solve the problem that data training cannot be centralized due to data privacy.Traditional federated learning uses a centralized parameter server to aggregate,update and distribute models,which is prone to single point of failure and data leakage.By combining a decentralized,immutable blockchain,the trust problem in federated learning can be solved.However,blockchain cannot avoid poisoning attacks caused by malicious actors modifying training samples.Existing research combines aggregation rules or anomaly detection to defend against poisoning attacks,but there are still problems such as poor defense effect,difficulty in accurately identifying malicious nodes,and failure of more than half of the poisoning rate.To solve the above problems,the main contributions of this paper are as follows:Firstly,aiming at the poisoning attack problem in the blockchain-based synchronous federated learning,a blockchain federated learning with synchronous training behavior verification mechanism is proposed.This scheme can effectively identify more than half of malicious nodes in synchronous scenarios,and has better results than the existing schemes Krum,Trimmed-mean and Median,and the average accuracy of poisoning rates above 50% and below50% is improved by 2 % and 10 %.Secondly,aiming at the poisoning attack problem in the blockchain-based asynchronous federated learning,a blockchain federated learning with asynchronous training behavior verification mechanism is proposed.This scheme can effectively identify more than half of malicious nodes in asynchronous scenarios,and has better results than the existing schemes Fed Async and BAFL,and the average accuracy rate of poisoning rate from 0 to 90% is improved by 5%.Finally,a blockchain federated learning system with training behavior verification mechanism is designed and implemented.The core functions are federated learning project evaluation and federated learning project monitoring functions,which can effectively identify malicious local models in synchronous and asynchronous scenarios,and monitor and warn the corresponding nodes.To sum up,this paper mainly aims at the poisoning attack problem encountered in the blockchain-based federated learning in synchronous and asynchronous scenarios,and proposes a synchronous training behavior verification mechanism and an asynchronous training behavior verification mechanism,and uses the two mechanisms to design and implement them.A secure blockchain federated learning system.Characterizing the honesty of honest nodes through training behavior has important theoretical significance for ensuring the security of federated learning and has important application value for improving learning quality.