基于Transformer风格卷积网络的上市公司财务造假侦查技术研究  

Research on Financial Fraud Investigation Technology in Listed Companies Based on Transformer-style Convolutional Network

作　　者：刘珊 邓志伟 LIU Shan;DENG Zhiwei(Department of Business Administration,Anhui Vocational College of Grain Engineering,Hefei 231635,China;School of Instrumentation Science and Optoelectronic Engineering,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]安徽粮食工程职业学院工商管理系,安徽合肥230011 [2]合肥工业大学仪器学院与光电工程学院,安徽合肥230009

出　　处：《中国人民公安大学学报（自然科学版）》2023年第3期102-108,共7页Journal of People’s Public Security University of China(Science and Technology)

基　　金：安徽省质量工程教学研究项目(2022jyxm420);安徽省科学研究项目—重点项目(2023AH053268)。

摘　　要：在当今数字时代下,上市公司财务造假具有专业性、隐蔽性、智能化特点,涉案数据体量大、数据类型多、数据关系复杂等是上市公司财务造假侦查面临的主要挑战。如何利用前沿人工智能(AI)技术对财务造假进行智能化数据挖掘和侦查是经济侦查亟待解决的问题。由于上市公司财务数据的不同特征字段之间存在潜在的依赖关系,为此提出了一种Transformer风格卷积神经网络模型(TSCNN)对上市公司的财务造假进行监测识别,该技术针对深度学习中流行的卷积神经网络普遍应用小卷积核,无法对全局特征依赖性进行建模等问题,利用Transformer架构中自注意力机制捕获全局的特征依赖,同时嵌入的大卷积核卷积层编码特征权重,增强感知特征间依赖关系的能力。此外,TSCNN通过线性复杂度的Hadamard积简化自注意力机制的运算。研究表明,所提出的侦查技术在上市公司财务数据上的鉴别性能优于现有流行的ResNet和MLP等人工智能算法。该方法是人工智能技术在经济侦查应用中的探索,在创新升级传统经济侦查手段基础上进一步提升办案效率。Financial fraud of listed companies in the digital age has the characteristics of professionalism,concealment and intelligence.The main challenges for the financial fraud investigation in listed companies are large data volumes,multiple data sources,and complex data relationships.How to apply advanced artificial intelligence(AI)technology for intelligent data mining and investigation of financial fraud is an urgent problem to be solved in economic investigation.Due to the potential dependence between different features of the financial data,a transformer-style convolution neural network(TSCNN)is proposed to monitor and identify financial fraud in listed companies.This technology aims at the problems such as the incapability to model global feature dependencies due to the widespread use of small convolutional kernels of popular convolutional neural networks in deep learning.The self-attention mechanism in transformer architecture is able to capture global feature dependencies,and the embedded convolutional layer of large convolutional kernels encodes features weights to enhance the ability to perceive the feature dependency relationship.Moreover,TSCNN simplifies the operation of self-attention mechanism through the Hadamard product of linear complexity.The results show that the proposed investigation technology outperforms the existing popular ResNet and MLP algorithms for discriminating financial data of listed companies.This technology is an exploration of the application of artificial intelligence technology in economic investigation,and further improves the efficiency of case handling based on innovative upgrading of traditional economic investigation methods.

关 键 词：上市企业 财务造假 人工智能 经济侦查 

分 类 号：D918.2[政治法律—法学]