Transaction Network Graph Neural Networks for Automated and Robust Financial Fraud Detection in Corporate Auditing

Abstract

This paper proposes an automated graph neural network-based method for financial fraud detection in corporate audit transaction data. The study first analyzes the characteristics of fraud in complex transaction networks and models enterprises, accounts, and their relationships as a graph to consider both node attributes and relational dependencies. A detection framework combining graph convolution and graph attention is then constructed to jointly learn node features and multi-layer neighbor information, capturing hidden abnormal patterns in the transaction network. To validate effectiveness, experiments are conducted on a public dataset with systematic evaluation using AUC, F1-Score, Precision, and Recall. The results show that the proposed method outperforms comparison models and effectively identifies fraudulent behavior in complex environments. Furthermore, robustness analysis is performed from three perspectives: hyperparameter sensitivity, environmental sensitivity, and data sensitivity, focusing on the effects of learning rate, number of attention heads, label noise rate, and feature missing rate on performance. The findings demonstrate that the method maintains stable performance under different conditions, confirming its reliability under multi-dimensional perturbations. In conclusion, by integrating graph-based modeling with deep learning techniques, this paper provides an efficient and robust automated solution for financial fraud detection in corporate auditing and shows strong potential for real-world applications.