Research On Financial Fraud Detection Method Based On Graph Neural Networks

Financial fraud refers to the act of obtaining financial benefits by dishonest means.Such behavior not only disrupts the financial market order,but also harms social and economic development and breeds other illegal and criminal behaviors.With the popularity of the Internet and online payment methods,many fraudulent activities and money laundering in life have also shifted from offline to online,which poses a great challenge to regulators.How to efficiently detect these financial frauds has become an urgent problem nowadays.Graph neural networks,a deep learning model that exploits interaction relations in graph structures,have been widely used in the field of fraud detection,but there are still some problems.First,fraudulent activities account for only a very small fraction of transaction transfers,which leads to the inevitable label imbalance problem in using graph neural networks to detect fraud.At the same time,fraudsters tend to camouflage their behavior,which exacerbates the negative impact of label imbalance and camouflaged behavior on the final prediction results.In addition,existing studies have ignored the importance of balancing neighbor features and central node features.For example,when the central node has too many neighbors,the features of the central node itself are often ignored.Finally,fraudulent activities and patterns are constantly changing over time,so it is also important to consider the dynamic evolution scenarios of fraud graph edge relationships.Therefore,in this paper,the following studies are conducted to address the above issues.(1)A fraud detection method based on neighbor sampling and neighbor aggregation under graph edge relations.First,a relevant model is defined for the multivariate relationship graph and fraud detection problem.Then,node neighbors are sampled using a sampler based on similarity measure to effectively avoid the influence of relationship camouflage and irrelevant noisy data of fraudulent nodes on the results.Then,the optimal number of neighbors and central node weights are found by reinforcement learning,so as to avoid the disappearance of the central node’s own features in the aggregation process and to better utilize the neighbor node information.Finally,the final node embeddings are obtained by aggregating the embeddings under different relationships.Experiments show that our proposed method outperforms existing algorithms.(2)For the problem of fraud detection in dynamic graphs with changing graph structure in transaction scenarios,a fraud detection method for financial transactions based on spatiotemporal graph convolution is proposed.Firstly,the problem is defined for dynamic graphs.Then,the structural information on the graph is extracted using a graph convolution network,and the temporal features of the dynamic graph are extracted by combining a long and short-term memory network.Then,the fraudulent nodes in the time step of the time window are oversampled so as to balance the label imbalance problem existing in the time step.Finally,in the process of graph convolutional network aggregation,a similarity calculation method based on multilayer perceptron is used to improve the reasonableness of the structural information on the graph extracted by the graph convolutional network.Experiments show that the spatiotemporal graph convolutional fraud detection method proposed in this paper outperforms the existing dynamic graph-based fraud detection algorithms.(3)This paper designs and implements a network transaction fraud detection system.The financial transaction fraud detection system based on graph neural network is a subsystem of financial risk control system.The fraud detection system combines the transaction fraud detection model based on spatio-temporal convolution and the fraud detection model based on neighbor sampling and neighbor aggregation under graph edge relationship.The system implements MySQL master-slave replication to ensure data security and high availability.