Low-Resource Convolution Neural Network for Keyboard Recognition of the User

The article is devoted to the improvement of neural network tools for recognizing the user's identity by their keyboard handwriting. It is shown that the most promising recognition tools are based on convolutional neural networks, and its drawback is high resource consumption of the network. It is proposed to reduce resource consumption by using modern types of convolutional neural networks that are adapted to the expected conditions of the analysis of keyboard handwriting. As a result of the studies, it was possible to effectively analyze keyboard handwriting using a convolutional neural network of the SqueezeNet type, the main advantages of which are a small amount of memory consumed, high recognition speed, availability of a pre-trained model, the possibility of implementation using proven tools, and also sufficient recognition accuracy. The SqueezeNet model, adapted to the task of recognizing the identity of 10 users by their keyboard handwriting, is implemented using the MATLAB application package. As the analyzed parameters of the keyboard handwriting, the ASCI code of the entered character, the key holding time, and also the time between successive pressing of two keys are used. To encode the input parameters of the neural network model, a special procedure was used to represent the parameters of the keyboard handwriting in the form of a color three-channel square-shaped pattern 227x227 in size. Experimental studies have shown that, along with a fairly low resource consumption, the SqueezeNet model allows for the error in recognizing the user's identity by keyboard handwriting at the level of the best modern systems for this purpose. It is proposed to correlate the paths of further research with the development of a method for adapting the architectural parameters of convolutional neural networks to the recognition of the user's personality and emotional state based on an integrated analysis of several biometric parameters.