Automatic classification of text messages by confidentiality level based on ensemble of artificial neural networks

This study presents an ensemble of convolutional neural networks that was developed to detect sensitive information in text messages and classify them according to their confidential level. The ensemble includes a pre-trained neural network based on the E5 architecture, an autoencoder based on convolutional and fully connected layers that allows you to separate public and private information, as well as several convolutional networks for subsequent classification of messages by confidential level. Convolutional neural networks were trained automatically according to the proposed procedure on organizational data containing confidential documents. However, public documents were not used for training. A feature of the proposed solution is support for 4 levels (or categories) of confidentiality. Low error rates of "missing confidential information" of 0.015% and "false detection of confidential information" of 4.3% were achieved. The proposed solution can be used in corporate messaging systems or in email analysis and is fully compatible with Mandatory Access Control. (Abstract)