Crucial business data is an essential asset for each business establishment. Data security is vital when sensitive data are transmitted over the Internet in a business environment. Steganography is the art of obscuring data inside a regular file of similar or different forms. For digital forensics, hiding data has always been necessary. The current information hiding method based on deep learning models can not directly use the original data as carriers, which means the method can not use the prevailing data in big data to hide information. Hence, this paper proposes a Deep neural network-based invisible text steganalysis (DNNITS) for business data hiding. This paper uses a word embedding layer to extract the syntax and semantic word features. A rough set of relative information entropy has been employed based on information features, and the optimized feature matrices are determined. The information in the optimized feature matrices are weighted, and the hiding information weighted feature is acquired. The findings reveal that our model can safely hide secret messages conveniently, quickly, and with no restriction on the business environment's data amount. The experimental results show that the suggested DNNITS model enhances the extraction rate of 95.4%, significance rate of 97.5%, the performance ratio of 89.6%, an efficiency ratio of 98.7%, recall ratio of 90.4%, and the lower error rate of 10.2% compared to other existing models.

关键业务数据是每个业务机构的重要资产。在商业环境中通过 Internet 传输敏感数据时，数据安全至关重要。隐写术是在类似或不同形式的常规文件中隐藏数据的艺术。对于数字取证，隐藏数据一直是必要的。目前基于深度学习模型的信息隐藏方法不能直接以原始数据为载体，也就无法利用大数据中的主流数据来隐藏信息。因此，本文提出了一种基于深度神经网络的隐形文本隐写分析（DNNITS）用于业务数据隐藏。本文使用词嵌入层来提取句法和语义词特征。基于信息特征采用了一组粗略的相对信息熵，并确定优化的特征矩阵。对优化后的特征矩阵中的信息进行加权，得到隐藏信息加权特征。研究结果表明，我们的模型可以方便、快速地安全地隐藏秘密消息，并且对商业环境的数据量没有限制。实验结果表明，所提出的DNNITS模型提升了95.4%的提取率、97.5%的显着性、89.6%的性能比、98.7%的效率、90.4%的召回率和10.2的较低错误率% 与其他现有模型相比。快速，不受业务环境数据量限制。实验结果表明，所提出的DNNITS模型提升了95.4%的提取率、97.5%的显着性、89.6%的性能比、98.7%的效率、90.4%的召回率和10.2的较低错误率% 与其他现有模型相比。快速，不受业务环境数据量限制。实验结果表明，所提出的DNNITS模型提升了95.4%的提取率、97.5%的显着性、89.6%的性能比、98.7%的效率、90.4%的召回率和10.2的较低错误率% 与其他现有模型相比。