Optimized Convolutional Neural Network for Intelligent Financial Statement Anomaly Detection

Abstract

This study proposes an intelligent detection model based on an improved convolutional neural network to address the complexity and data category imbalance issues in financial statement anomaly detection. This model introduces a deep feature extraction network and attention mechanism based on traditional CNN to enhance the ability to pay attention to key data features. At the same time, by optimizing the loss function and using data enhancement technology, the performance and robustness of the model in anomaly detection tasks are effectively improved. In the experiment, this study selected the real Enron data set for verification and designed comparative experiments to compare the performance with various traditional and deep learning models. Experimental results show that the improved CNN model is significantly better than other comparison models in terms of accuracy, F1 score, recall rate, precision rate and other indicators, especially in scenarios with small data volume and category imbalance. In addition, the adaptability and generalization ability of the improved model are further demonstrated through experimental analysis of different data proportions. The research results not only provide technical support for abnormal detection of financial statements, but also provide theoretical basis and practical value for promoting the intelligent development of financial auditing.