摘要
联邦学习作为一种新兴的分布式机器学习框架,在训练过程中模型更新会占用大量网络带宽,这成为联邦学习获得高精度机器学习模型的瓶颈之一。为了解决以上问题,基于Hyperledger Fabric区块链,设计了一个简单而有效的自适应阈值更新算法,并提出了一种自适应权重趋势感知的联邦学习解决方案。通过客户端训练的本地模型与全局模型方向向量矩阵的相关性,来筛除与全局模型偏差较大的客户端模型,同时在训练过程中自适应地调整筛选阈值。实验结果表明,相比于传统的联邦学习方案,提出方案减少了神经网络模型训练过程中超过20%的网络通信开销以及节约超过五倍的训练资源,提高了近4%的模型精确度,并且训练过程可追溯和去中心化,极大地提高了隐私安全保障。
As an emerging distributed machine learning framework,Federated Learning consumes a large amount of network bandwidth for model updates during training,which becomes one of the bottlenecks for Federated Learning to obtain high accuracy machine learning models.To address the above problems,a simple and effective adaptive threshold update algorithm is designed based on the Hyperledger Fabric blockchain and a Federated Learning solution with adaptive weight trend perception is proposed.The correlation between the local model trained by the client and the direction vector matrix of the global model is used to screen out client models that deviate significantly from the global model,while adaptively adjusting the screening threshold during the training process.Experimental results show that over 20%of the network communication overhead and more than 5 times the training resources are saved during the training of neural network models compared to traditional federation learning solutions,nearly 4%of the model accuracy is improved and the training process is traceable and decentralized,which significantly improves privacy and security.
作者
刘振
吴宇
LIU Zhen;WU Yu(School of Cyberspace Security,Dongguan University of Technology,Dongguan 523808,China)
出处
《电子设计工程》
2023年第24期75-80,共6页
Electronic Design Engineering
关键词
联邦学习
自适应
区块链
去中心化
Federated Learning
adaptive
blockchain
decentralized
