Electronic invoice(E-invoice) has become the product of the information age, its issue will greatly save the cost of enterprises and achieve the goal of financial process automation. Hence, the generalization of electronic invoice is imperative. However, there exists the risk of malicious attacks in electronic invoice systems, such as sudden invoice of large invoice, invoice at abnormal time, etc. These malicious attacks are difficult to detect through the system itself or manually. To provide a secure service platform for the generalization of electronic invoice, this paper studies the attack detection technology of electronic invoice systems which is mainly based on machine learning to complete two aspects of research. The first is to propose a machine learning-based e-invoice anomaly detection method, which can accurately determine the anomalies occurring in the e-invoice systems. The second is to conduct deep fusion analysis on abnormal behaviors, mining potential threats in the electronic invoice systems, and designing and implementing the electronic invoice depth fusion analysis method based on k-means and Skip-gram. The experimental results indicate that the method we proposed can not only detect the malicious attacks effectively, and also capable of mining the potential threats in the electronic invoice systems.

电子发票已成为信息时代的产物，其发行将大大节省企业成本，达到财务流程自动化的目的。因此，电子发票的泛化势在必行。但是，在电子发票系统中存在恶意攻击的风险，例如大发票的突然发票，异常时间的发票等。这些恶意攻击很难通过系统本身或手动检测出来。为了为电子发票的归纳提供一个安全的服务平台，本文研究了主要基于机器学习的电子发票系统的攻击检测技术，完成了两个方面的研究。首先是提出一种基于机器学习的电子发票异常检测方法，可以准确确定电子发票系统中发生的异常。第二是对异常行为进行深度融合分析，挖掘电子发票系统中的潜在威胁，并设计和实现基于k-means和Skip-gram的电子发票深度融合分析方法。实验结果表明，本文提出的方法不仅可以有效地检测出恶意攻击，还可以挖掘电子发票系统中的潜在威胁。