The article describes an approach to the operation of a data transmission network protection system against computer attacks based on a hybrid neural network. It is proposed to use a hybrid neural network as a machine learning method. To calculate the output value of neural network signals, the activation function is used. The neural network model consists of recurrent cells - LSTM and GRU. Experiments have demonstrated that the proposed network protection system for detecting computer attacks based on an assessment of the self-similarity of the system functioning parameters using fractal indicators and predicting the impact of cyber attacks by applying the proposed structure of the LSTM neural network has a sufficiently high efficiency in detecting both known and unknown spacecraft. The probability of detecting known spacecraft is 0.96, and the zero-day attack is 0.8.
Keywords: data transmission network, computer attack, neural network, protection system, network traffic, auto-encoder, accuracy, completeness, detection, classifier, self-similarity, recurrent cells with long short-term memory