In recent years, with the significant growth in the field of communications and network data analysis, the use of graphs for representing and analyzing this data has gained tremendous importance. Graphs can clearly represent structural and relational information in network data. However, the volume of network data is rapidly increasing and constantly changing. Furthermore, detecting anomalies in this data is of paramount importance as it can be effective in preventing fraud, the dissemination of fraudulent information, and organizing attacks in networks. Various methods for anomaly detection in graphs have been proposed so far, but most of them have focused on detecting anomalies in edges and paid less attention to detecting anomalies in nodes. This issue can lead to challenges in estimating changes in nodes over time. In this thesis, the primary goal is anomaly detection in nodes within dynamic graphs. Graph Neural Networks (GNNs) are employed to examine the structure and temporal behavior of nodes for anomaly detection in dynamic networks. The results obtained from this approach demonstrate its capability to detect anomalies in real-world scenarios, accurately scrutinizing topological changes in the graphs. This research leverages datasets such as Darpa, EU-CORE, and EU-CORE-L for model eva‎luation. Due to its memory efficiency and parallelization capabilities, this method is applicable to large-scale graphs. The accuracy of anomaly detection for the Darpa dataset reaches approximately 62 percent, representing an improvement compared to previous methods.

زينب مالكي

زينب زالي

ناصر قديري مدرس , عليرضا بصيري