Application of methods for studying a convolutional neural network for people with limited emotions for the purpose of behavioral information security

Mandritsa Igor Vladimirovich, Kopytov Vladimir Vyacheslavovich, Chernyshev Alexander Borisovich, Makarov Anatoly Mikhailovich, Reznikov Dmitry Konstantinovich

The behavior of an employee of an organization in the field of information security must comply with the adopted information security policy in it. A stable working «state» changes to a problematic «unstable» state, and with a high probability leads to information threats to the organization, in the form of information leaks, or irreversible information threats. The emotions of a person (employee) always have a digital «footprint» and are present on the face. This "trace" fills in the metadata for the mathematical model of the organization's information security, where the excess of the calculated probability level determines the «access» of this employee to the organization's business process. For recognition emotion, a neural network consisting of 152 layers is used, with an output layer of 7 neurons, one for each emotion. During the training, a dataset of 28,000 full-face images of people's faces was used, expressing 7 emotions: anger. disgust, fear, happiness, sadness, surprise, calmness. The neural network is written in Python using the PyTorch library.

Keywords:neural network training, emotion recognition, behavioral information security