Research on a financial fraud identification model by fusing a convolutional neural network

Abstract

To address the challenges of low accuracy and poor real-time performance in existing financial fraud identification methods for listed companies, this paper proposes a hybrid identification model (CNN-SVM) that fuses a convolutional neural network with a support vector machine. Utilizing a comprehensive dataset of 7,429 samples from non-financial A-share listed companies penalized for fraud between 2007 and 2022, the study employs random oversampling to rebalance the minority class, resulting in a robust training set of 13,540 samples. The model leverages a CNN architecture to automatically extract high-level features from 87 indicators spanning corporate governance, financial oversight, and operational metrics, which are then classified using a linear-kernel SVM. Experimental results demonstrate that the CNN-SVM model achieves a qualitative leap in performance, yielding an AUC of 0.97, a recall of 0.99, and an F1-score of 0.97, significantly outperforming traditional logistic regression and random forest benchmarks. These findings suggest that the proposed framework effectively balances training efficiency with high precision, providing a superior tool for real-time financial risk control.