Scientific Journal of KubSAU

This article is devoted to the problem of network attacks recognition, which is essential for providing network security. A research of neural network efficiency has been held. Such metaeuristic algorithms as genetic algorithm, gray wolf algorithm and firefly algorithm have been applied for the neural network learning. The algorithms’ fundamentals have been described. Multilayer perseptrone with sigmoid activation function has been selected for the task of network attack presence check. Various configurations of the neural network have been tested in order to find the optimal number of layers and neurons per layer, which ensure the least error. Learning has been performed by minimization of the average squared error between the network’s output and its target value with the help of the listed algorithms. Genetic algorithm requires accurate parameter picking in case of any network’s architecture alteration. Moreover, it is not as fast as firefly and gray wolf algorithms. Gray wolf algorithm appears to be the most effective one. However, it loses its efficiency if the number of layers is increased. Firefly algorithm proves to be the most universal one. Although it is less effective than gray wolf algorithm, it provides the most exact output even if the network’s structure is changed