Research On Anomaly Detection Algorithm Of Attributes Network Based On Multichannel Autoencoder

The research of anomaly detection based on attribute network aims at finding the nodes whose patterns deviate significantly from the majority of reference nodes.It has been widely used in various security tasks such as network intrusion detection and social spammer detection.However,the potential complex cross-modal interaction between the network structures and node attributes is ignored by most existing methods of anomaly detection,which reduces the quality of node feature learning and therefore degrades the performance of the anomaly detection algorithm.This paper proposes a deep joint representation learning method,and then proposes an anomaly detection model.In order to capture the cross-modality interactions between network structure and node attribute in attribute networks,this paper designs a multichannel autoencoder joint learning framework(MAF).MAF consists of a node encoder,a structure encoder and an attribute encoder to learn both node embedding and attribute embedding jointly in latent space.The embeddings are used as the inputs of attribute decoder and the cross-modality interactions between network structure and node attribute are learned during the reconstruction of node attribute.Finally,the reconstruction errors of network structure and node attributes are reduced by multiple iterations,and so that the reconstructed network resemble the original network to the maximum extent possible.In the anomaly detection task,the abnormal nodes are randomly distributed.In order to reduce the interference of abnormal neighbor information on normal nodes in the process of feature extraction,this paper introduces attention mechanism for node embedding learning.Finally,an anomaly detection algorithm of attribute network based on attention mechanism and multichannel autoencoder(ADAN-AMA)is proposed.In the updating process of node embedding,the feature similarity between the central node and all neighboring nodes is calculated.The information of normal nodes is retained in the feature aggregation process and weakens the influence of abnormal nodes.Experiments on multiple public datasets show that our proposed method consistently outperforms state of the art algorithms,with up to 23.6%,30% and20.3% in terms of AUC score,precision,and recall respectively,which demonstrates the effectiveness of the proposed method.