Prototype of leakage protection system and neural network model for detection of confidential text messages

The study raises the problem of confidential information leaks arising from the actions of internal intruders. Modern software products, including those aimed at protecting confidential data by analyzing user behavior, are considered as methods of countering such attacks. To solve the problem, we propose a prototype of an advanced DLP-system that uses artificial intelligence methods to analyze user activity and prevent such leaks. The developed prototype implements a mechanism for monitoring the context of user behavior and uses machine learning models to detect sensitive information in short text messages. Specifically, a pre-trained neural network based on E5 architecture is used for feature extraction. For confidential information detection, a convolutional neural network of autoencoder type is applied, which is trained exclusively on data containing confidential documents. Experimental results have shown that the proposed model successfully performs binary classification of messages, with the error of the second kind at the level of 5.1% and the absence of errors of the first kind (within the framework of the conducted experiment on its own data set). The proposed complex can serve as a basis for the creation of more complex solutions in the field of information security.