基于自编码器神经网络的会计异常数据自动审计方案

Automated Auditing Scheme of Abnormal Accounting Data based on Autoencoder Neural Network

大规模会计数据的欺诈检测是财务报表审计的一个难题,为提高会计数据审计的准确度和效率,提出了基于深度自编码器(AE)神经网络的会计数据异常检测方案。从会计信息系统(AIS)中提取日记账分录,并基于正常账目和欺诈账目的属性值特征定义全局异常和局部异常特征,分别对应个别异常属性值和多个异常属性值组合。由此,训练AE网络识别异常数据被给出账目属性的异常评分。在企业年度会计数据集上的实验结果表明,AE网络的检测性能随网络深度的增加而上升,且所提方法在检测精度和误警数占比方面的性能优于其他比较方法,能够有效提高财务审计任务的准确度和效率。

Fraud detection of large-scale accounting data is a difficult problem in financial statement auditing.To improve the accuracy and efficiency of accounting data auditing,an anomaly detection scheme for accounting data based on deep Autoencoder(AE)neural network is proposed.Journal entries are extracted from the Accounting Information System(AIS),and global and local abnormal features are defined based on the characteristics of normal and fraudulent attribute values,corresponding to the occurrence of individual attribute value and co-occurrence of multiple abnormal attribute values,respectively.From this,the AE network is trained to identify anomalous data and provide anomaly scores for journal entries.The experimental results on a company's annual accounting data set show that the detection performance of the AE network increases with the increase of the network depth,and the performance of the proposed method is better than other comparison methods in terms of detection precision and proportion of false alarms,which can effectively improve the accuracy and efficiency of financial audit tasks.