Source Localization of Network Information Propagation via Invertible Graph Diffusion

Abstract

With the development of society, security issues in various types of networks have become increasingly prominent, especially network propagation issues. Accurately locating the diffusion source points of network propagation is an important means to control network propagation. The research on the source location of network propagation also faces problems such as diverse network structure and complex dissemination mechanism. Therefore, this paper studies the problem of source location of network propagation based on graph neural networks, and an invertible graph diffusion model based on graph convolutional networks (GCNIGD) is proposed. In the stage of node susceptibility estimation, the graph convolutional neural network is combined to make full use of the structural information of the network considering the connection relationship between network nodes. In the stage of node feature construction, the graph diffusion theory is combined to spatially localize the information propagation in the network, so that the graph-based model can be enhanced by learning from multi-hop information. In the stage of source localization, the graph traceability problem is transformed into the inverse problem of graph diffusion, a reversible graph network is constructed to accurately estimate the source node, and the ill-posed problems in network traceability are solved. Finally, extensive experiments are conducted on six real-world datasets, and the results show that the proposed method outperforms the state-of-the-art methods. This study has important guiding significance for network security issues such as false information traceability, network attack traceability, etc.

Keywords

• source localization; graph neural networks; graph diffusion; inverse problem