The description of the developed neural network attack detection algorithm is given, the peculiarity of which is the possibility of launching two parallel processes: searching for the optimal model of an artificial neural network and normalizing the training sample data. It is shown that the choice of the artificial neural network architecture is carried out taking into account the loss function for a limited set of attack classes. The use of libraries (frameworks) TensorFlow and Keras Tuner for the software implementation of the attack detection algorithm is shown. The description of the experiment on choosing the architecture of the neural network and its training is given. The accuracy obtained in experiments reaches 94-98 % for different classes of attacks.