基于联邦学习的多源异构网络无数据融合方法

A Multi-Source Heterogeneous Network Compression without Data

在联合作战体系中,数据作为基础性战略资源发挥着重要的底层支撑作用,数据妥善管理和高效利用是推动作战能力整体跃迁和作战样式深度变革的重要动力。为实现不同作战系统间信息的互联互通,提出一种基于联邦学习的多源异构网络无数据融合方法。从多源数据融合面临的安全性和异构性问题出发,利用条件生成对抗网络提取本地知识和全局分布,集成数据信息;结合局部教师模型-全局模型架构,以无数据知识蒸馏的方式对局部模型知识进行迁移,融合异构网络,细化全局模型,实现不同系统间安全、高质量的信息交互,为智能化指挥信息系统建设提供技术支撑。实验结果表明:该方法在结构化数据和图像数据上具有可行性,整体准确率可达到80%以上。

Data serving as a basic strategic resource is playing an important underpinning role in joint combat system.The proper management and efficient use of data are an important driving force in promoting the overall transformation of combat capability and the deep transformation of combat style.In order to realize the information interconnection between different combat systems,a multi-source heterogeneous network data fusion method is proposed based on the federated learning.In view of the security and heterogeneity of multi-source data,the conditional generation adversarial network is utilized for extracting local knowledge and global distribution,and integrating data information.In combination with the local teacher model-global model architecture,the local model knowledge is transferred by distillation of knowledge without data,the heterogeneous network is fused,and the global model is refined to realize safe and high-quality information interaction between different systems,providing technical support for the construction of intelligent command information system.The experimental results show that the proposed method is feasible on structural data sets and image data sets,and the overall accuracy can be more than 80%.