首页|基于图卷积网络与社群发现的异常检测方法

基于图卷积网络与社群发现的异常检测方法

扫码查看
深度学习领域对于图结构数据的关注程度日益加深,社交、医疗、电力系统等多个领域已将实体抽象成属性网络的形式,知识图谱等组织方式成功实现了信息的高效组织与管理.在这些信息丰富的网络中,安全问题尤为重要,因为异常实体的存在可能会对整体利益造成威胁.传统方法在图结构数据异常检测方面存在一定困难,尤其是在捕捉网络高维度特征方面表现不佳.深度学习方法尽管强大,但由于网络深度的限制,往往难以获取网络的全局信息.为此,提出一种基于图卷积网络的两阶段异常检测方法,通过图卷积神经网络逐步获取节点的社群信息,克服了传统方法在高维度特征捕捉上的不足;同时考虑节点自身属性,以更好地适应各类复杂网络结构,提升了异常检测性能.实验结果表明,该方法在部分数据集上的AUC分数超过0.9,在各数据集上与基线方法相比可以取得最优或次优性能表现.
Anomaly Detection Method Based on Graph Convolutional Networks and Community Detection
The field of deep learning is paying increasing attention to graph structured data,and multiple fields have abstracted entities into attribute networks.Knowledge graphs and other organizational methods have successfully achieved efficient organization and management of in-formation.In these information rich networks,security issues are particularly important as the presence of anomalous entities may pose a threat to overall interests.Traditional methods face certain difficulties in anomaly detection of graph structured data,especially in capturing high-di-mensional network features.Although deep learning methods are powerful,due to the limitations of network depth,it is often difficult to obtain global information from the network.Therefore,a two-stage anomaly detection method based on graph convolutional neural network is pro-posed,which gradually obtains the community information of nodes through graph convolutional neural network,overcoming the shortcomings of traditional methods in capturing high-dimensional features;Simultaneously considering the node's own attributes to better adapt to various complex network structures and improve anomaly detection performance.The experimental results show that the AUC score of this method ex-ceeds 0.9 on some datasets,and it can achieve optimal or suboptimal performance compared to the baseline method on each dataset.

anomaly detectiongraph convolutional networkscommunity detectionattributed network

夏飞、赵新建、王恺祺、陈石

展开 >

国网江苏省电力有限公司信息通信分公司,江苏 南京 210000

南京大学 计算机学院,江苏 南京 201008

异常检测 图卷积网络 社群发现 属性网络

2024

软件导刊
湖北省信息学会

软件导刊

影响因子:0.524
ISSN:1672-7800
年,卷(期):2024.23(12)