基于对抗多关系图神经网络的机器账号检测  

Adversarial Multi-relation Graph Neural Network for Bot Account Detection

作　　者：杨英光 李阳阳[2] 彭浩 刘弋锋 谢海永 YANG Yingguang;LI Yangyang;PENG Hao;LIU Yifeng;XIE Haiyong(School of Cyber Science and Technology,University of Science and Technology of China,Hefei,Anhui 230026,China;National Engineering Laboratory for Public Safety Risk Perception and Control by Big Data(PSRPC),China Academy of Electronics and Information Technology,Beijing 100041,China;Beijing Advanced Innovation Center for Big Data and Brain Computing,Beihang University,Beijing 100191,China;Key Laboratory of Cyberculture Content Cognition and Detection,Ministry of Culture and Tourism,Hefei,Anhui 230026,China)

机构地区：[1]中国科学技术大学网络空间安全学院,安徽合肥230026 [2]中国电子科技集团公司电子科学研究院社会安全风险感知与防控大数据应用国家工程实验室,北京100041 [3]北京航空航天大学网络空间安全学院,北京100191 [4]网络文化内容认知与检测文化和旅游部重点实验室,安徽合肥230026

出　　处：《中文信息学报》2023年第7期162-172,共11页Journal of Chinese Information Processing

基　　金：国家自然科学基金(U20B2053);海南省重大科技计划(ZDKJ2019008);重点研究与发展计划专项(SQ2021YFC3300088)。

摘　　要：现有的机器账号检测方法或者依赖于对机器账号的先验知识,或者在检测时只关注单一账号的特征,忽略了与该账号有关系的其他账号所能带来的潜在表征,降低了所提检测方法的有效性。针对上述不足,该文提出了一种基于生成对抗网络的多关系图神经网络检测模型。从社交网络数据集中抽取不同关系,建立多关系图,采样节点,训练生成对抗网络,来动态改变关系图结构;将节点特征和图结构信息输入图神经网络,有选择的聚合邻居节点的特征,得到更加精确的图嵌入向量,将向量输入分类器进行检测。实验结果表明,相比于其他算法,该文所述算法在两个数据集中AUC分别最多提升了24%和9%,Recall值分别最多提升了13%和4%。At present,bot account detection methods either rely on the prior knowledge of the accounts,or focus only on the features of a specific account,without considering the potential features among other accounts related to the account.This paper proposes a model based on multi-relation graph neural network with generative adversarial network for bot account detection.We extract different relationships from the social network data sets to build a multi-relationship graphs.We sample the training nodes to dynamically change the structure of multi-relationship graphs.We input the node features and graph structure information into the graph neural network,selectively aggregating the embedding vectors of neighbor nodes.Experimental results on two data sets show that compared with other algorithms,the proposed algorithm improves by 24%and 9%in AUC,by 13%and 4%in recall,respectively.

关 键 词：机器账号检测 图神经网络 生成对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]