Financial frauds, often executed through asset transfers and profit inflation, aim to reduce taxes and secure credits. To enhance the accuracy and efficiency of accounting data auditing, this study proposes an anomaly detection scheme based on a deep autoencoder neural network. Financial statement entries are extracted from the accounting information system, and global and local anomaly features are defined based on the attribute values of normal and fraudulent accounts, corresponding to individual and combined anomaly attribute values. The AE network is trained to identify anomalies using account attribute scores. Results demonstrate classification accuracies of 91.7%, 90.3%, and 90.9% for sample ratios of 8:2, 7:3, and 6:4, respectively. The precision, recall, and F1 score reach 90.85%, 90.77%, and 90.81%, respectively. Training takes 95.81ms, with recognition classification requiring only 0.02ms. The proposed deep neural network achieves high recognition accuracy and speed, significantly improving the detection of financial statement anomalies and fraud.