Research And Application Of Model For Anomaly Alignment Across Multiple Attributed Networks

Anomaly alignment aims to introduce network alignment into anomaly detection to obtain abnormal results that satisfy the alignment relationship among different networks,so as to realize the associated anomaly mining among networks.Anomaly alignment plays an important role in cross-time,cross-type,and cross-domain information mining,such as the mining of homologous attacks among multi-time slice computer networks,the mining of related peak itineraries among multi-type traffic networks,and the mining of related abnormal characters between enterprise network and academic network,and so on.However,the existing anomaly alignment work is mainly based on the spectral method,which is suitable for the mining of associated anomalies between two attributed networks,and it is difficult to meet the needs of complex multi-network scenarios in reality.Therefore,this thesis introduces an alignment method based on representation learning to achieve the goal of anomaly alignment across multiple attributed networks,and conducts research from three aspects: model construction,experimental verification and empirical exploration.The main content and contributions of this thesis are as follows:Firstly,an anomaly alignment model A3MAN(Anomaly Alignment Across Multiple Attributed Networks)is proposed.This model introduces the network alignment theory based on compound embedding to the anomaly subgraph detection,and obtains the set of the largest aligned anomaly subgraphs among multiple attributed networks through the two constraints of anomaly and alignment.In addition,this model can also realize the anomaly mining on the attributeless network by aligning the related abnormal information of attributed networks to the attributeless network.Secondly,the A3 MAN model was verified experimentally.This thesis constructs two real scenarios of multi-time slice computer networks and multi-type traffic networks to verify the A3 MAN model.In the scenario of computer networks,the accuracy rate of abnormal IP detected by A3 MAN is 99%,the number of abnormal anchor links detected by A3 MAN is 8.8 times that of the comparison method,and 268 sets of related abnormal IPs are obtained across multiple periods.In the scenario of traffic networks,based on A3 MAN,the related peak itineraries among different types of traffic,as well as the commonalities of their geographic locations and evolutionary patterns,are obtained.These all confirm the effectiveness,robustness and usability of the model.Finally,an empirical exploration is carried out based on the A3 MAN model.Based on the A3 MAN model and taking business risks as anomalies,this thesis constructs a real enterprise relationship network and academic cooperation network,and conducts associated anomaly entity mining on it.Through the overall analysis and specific case studies of the obtained results(598 companies,1208 executives and 987 related academic staff),the actual risk status of the corresponding entities’ enterprises was explored.In addition,based on the empirical results,this thesis builts a visual interactive platform,which can can more intuitively explore the risk situation of enterprises.In summary,this thesis takes anomaly detection and network alignment as the theoretical cornerstones,and proposes A3 MAN,a model that realizes anomaly alignment across multiple attributed networks.Through experimental verification and empirical application on multiple types of data,the effectiveness of the model is verified.A3 MAN has important application value in associated anomaly mining on multiple attributed networks and anomaly mining on attributeless network.