2018
Permanent URI for this communityhttp://crust.ust.edu.ua/handle/123456789/10452
Browse
Item Identifying Threats in Computer Network Based on Multilayer Neural Network(Дніпропетровський національний університет залізничного транспорту імені академіка В. Лазаряна, Дніпро, 2018) Zhukovyts’kyy, Igor V.; Pakhomova, Victoria M.ENG: Purpose. Currently, there appear more often the reports of penetration into computer networks and attacks on the Web-server. Attacks are divided into the following categories: DoS, U2R, R2L, Probe. The purpose of the article is to identify threats in a computer network based on network traffic parameters using neural network technology, which will protect the server. Methodology. The detection of such threats as Back, Buffer_overflow, Quess_password, Ipsweep, Neptune in the computer network is implemented on the basis of analysis and processing of data on the parameters of network connections that use the TCP/IP protocol stack using the 19-1-25-5 neural network configuration in the Fann Explorer program. When simulating the operation of the neural network, a training (430 examples), a testing (200 examples) and a control sample (25 examples) were used, based on an open KDDCUP-99 database of 500000 connection records. Findings. The neural network created on the control sample determined an error of 0.322. It is determined that the configuration network 19-1-25-5 copes well with such attacks as Back, Buffer_overflow and Ipsweep. To detect the attacks of Quess_password and Neptune, the task of 19 network traffic parameters is not enough. Originality. We obtained dependencies of the neural network training time (number of epochs) on the number of neurons in the hidden layer (from 10 to 55) and the number of hidden layers (from 1 to 4). When the number of neurons in the hidden layer increases, the neural network by Batch algorithm is trained almost three times faster than the neural network by Resilient algorithm. When the number of hidden layers increases, the neural network by Resilient algorithm is trained almost twice as fast as that by Incremental algorithm. Practical value. Based on the network traffic parameters, the use of 19-1-25-5 configuration neural network will allow to detect in real time the computer network threats Back, Buffer_overflow, Quess_password, Ipsweep, Neptune and to perform appropriate monitoring.Item Optimal Route Definition in the Network Based on the Multilayer Neural Model(Дніпропетровський національний університет залізничного транспорту імені академіка В. Лазаряна, Дніпро, 2018) Pakhomova, Victoria M.; Tsykalo Igor D.ENG: Purpose. The classic algorithms for finding the shortest path on the graph that underlie existing routing protocols, which are now used in computer networks, in conditions of constant change in network traffic cannot lead to the optimal solution in real time. In this regard, the purpose of the article is to develop a methodology for determining the optimal route in the unified computer network. Methodology. To determine the optimal route in the computer network, the program model "MLP 34-2-410-34" was developed in Python using the TensorFlow framework. It allows toperform the following steps: sample generation (random or balanced); creation of a neural network, the input of which is an array of bandwidth of the computer network channels; training and testing of the neural network in the appropriate samples. Findings. Neural network of 34-2-410-34 configuration with ReLU and Leaky-ReLU activation functions in a hidden layer and the linear activation function in the output layer learns from Adam algorithm. This algorithm is a combination of Adagrad, RMSprop algorithms and stochastic gradient descent with inertia. These functions learn the most quickly in all volumes of the train sample, less than others are subject to reevaluation, and reach the value of the error of 0.0024 on the control sample and in 86% determine the optimal path. Originality. We conducted the study of the neural network parameters based of the calculation of the harmonic mean with different activation functions (Linear, Sigmoid, Tanh, Softplus, ReLU, L-ReLU) on train samples of different volumes (140, 1400, 14000, 49000 examples) and with various neural network training algorithms (BGD, MB SGD, Adam, Adamax, Nadam). Practical value. The use of a neural model, the input of which is an array of channel bandwidth, will allow in real time to determine the optimal route in the computer network.