摘要
算力网络旨在深度融合算力资源与网络资源,实现对泛在离散部署的海量数据和异构资源的高效利用。为了应对算力网络中复杂的多中心计算协同和数据隐私安全的要求,联邦学习在隐私保护和边缘算力资源利用等方面有着先天的优势。然而,基于算力边缘服务器的高度异构和广泛分布,算力网络环境下的联邦学习仍然存在一定的困难。一方面,算力网络中海量边缘服务器之间存在的数据异质性会引起非独立同分布问题,导致联邦学习中局部模型的更新偏离全局最优值。另一方面,由于不同边缘服务器的数据质量存在差异,生成低质量的局部模型会显著影响训练效果。为了解决上述问题,提出了基于区块链的自适应联邦学习(AWFL-BC,adaptiveweightfederatedlearning-blockchain)方案。通过智能合约计算不同边缘服务器的数据分布距离,生成相似度矩阵指导聚合。同时,设计了一种自适应权重聚合算法,以减轻由数据质量差异引起的模型性能和稳定性下降,从而提升模型的准确率,加速模型收敛。最后,结合区块链技术加强了安全保障机制,可有效防止投毒攻击与推理攻击。在3个公有标准数据集上进行的综合实验表明,与最先进的方法相比,AWFL-BC实现了更高的模型准确率,且模型收敛速度更快。
The computational power network was aimed to deeply integrate arithmetic resources and network re‐sources,achieving efficient utilization of massive data and heterogeneous resources in ubiquitous discrete deploy‐ment.To cope with the complicated multicenter computing collaboration and requirements for data privacy and se‐curity in computational power networks,federated learning was recognized for its inherent advantages in privacy protection and edge arithmetic resource utilization.However,federated learning in the computational power net‐work environment faced certain difficulties due to the highly heterogeneous and widely distributed arithmetic edge servers.On one hand,the heterogeneity of data among the massive edge servers in the computational power net‐work caused the Non-IID problem,leading to the deviation of the local model update from the global optimum in federated learning.On the other hand,the generation of low-quality local models could significantly affect the train‐ing effect due to the differences in data quality among different edge servers.To solve the above problems,a blockchain-based adaptive federated learning framework AWFL-BC(adaptive weight federated learning-blockchain)was proposed.Initially,the data distribution distances of different edge servers were calculated through smart contracts to generate a similarity matrix that guided aggregation.Concurrently,an adaptive weight aggrega‐tion algorithm was designed to alleviate the decrease in model performance and stability caused by differences in data quality,thereby improving the accuracy of the model and accelerating model convergence.Finally,the integra‐tion of blockchain technology strengthened the security mechanism,effectively preventing poisoning attacks and in‐ference attacks.Comprehensive experiments on three public standard datasets show that AWFL-BC achieves higher model accuracy and faster model convergence compared to state-of-the-art methods.
作者
刘天瑞
王连海
王启正
徐淑奖
张淑慧
王英晓春
LIU Tianrui;WANG Lianhai;WANG Qizheng;XU Shujiang;ZHANG Shuhui;WANG Yingxiaochun(Key Laboratory of Computing Power Network and Information Security,Ministry of Education,Shandong Computer Science Center(National Supercomputer Center in Jinan),Qilu University of Technology(Shandong Academy of Sciences),Jinan 250014,China;Shandong Provincial Key Laboratory of Computer Networks,Shandong Fundamental Research Center for Computer Science,Jinan 250014,China)
出处
《网络与信息安全学报》
2024年第3期130-142,共13页
Chinese Journal of Network and Information Security
基金
山东省重点研发计划项目(2021CXGC010107)
国家自然科学基金(62102209)
济南市“新高校20条”资助项目(202228017)
泰山学者工程资助(tsqn202312231)。
关键词
联邦学习
区块链
自适应权重
非独立同分布数据
算力网络
federated learning
blockchain
adaptive weight
non-iid data distribution
computational power network
