面向天基信息网络的智能模型分布式训练技术  

Distributed Training Techniques for Intelligent Model in Space-Based Information Networks

作　　者：栗渊钧 杨德伟[1] 李佳宁 冯笑 LI Yuanjun;YANG Dewei;LI Jianing;FENG Xiao(School of Information and Electronics,Beijing Institute of Technology,Beijing 100081,China;The 15th Institute of China Electronics Technology Group Corporation,Beijing 100083,China)

机构地区：[1]北京理工大学信息与电子学院,北京100081 [2]中国电子科技集团公司第十五研究所,北京100083

出　　处：《天地一体化信息网络》2025年第1期24-34,共11页Space-Integrated-Ground Information Networks

基　　金：国家重点研发计划资助项目(No.2022YFB2902703)。

摘　　要：针对天基信息网络中智能模型的分布式训练存在数据分布异构、模型陈旧以及隐私安全等问题,提出基于区块链的智能模型联邦学习架构和安全高效训练方法,引入差分隐私噪声机制和参数评估方法,有效应对隐私泄露、中毒攻击和单点故障威胁;采用基于时延最小的模型聚合方法,通过轨道内外的模型广播及区块广播过程,加速模型训练。仿真结果表明,所提方法能使不同结构的智能模型快速收敛,缩短训练时间,并有效应对安全隐私威胁。In addressing the issues of data distribution heterogeneity,outdated models,and data privacy and security in distributed training of intelligent models,a federated learning architecture of intelligent models was designed based on blockchain technology and applied to space-based information networks.A secure and efficient training method for intelligent models was proposed based on this architecture,where a differential privacy noise mechanism,the blockchain technology and a parameter evaluation method were introduced to effectively deal with privacy leakage,poisoning attacks and single-point failure threats.Meanwhile,using a model aggregation method based on the minimized delay,the model training was accelerated via the processes of intra-orbit and inter-orbit model broadcasting and block broadcasting.The simulation results indicated that the proposed method enables intelligent models of different structures to converge rapidly,shorten the model training time,and effectively deal with security and privacy threats.

关 键 词：天基信息网络 联邦学习 智能模型 区块链技术 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]