Analysis of the Multilayer Perceptron Parameters Impact on the Quality of Network Attacks Identification
О. В. Ерохина, Б. Б. Борисенко, I. D. Martishin, А. С. Фадеев · 2021
The paper investigates the architecture of a multilayer perceptron. Using the CSE-CIC-IDS2018 dataset we analyze the quality of computer attack (CA) identification under different parameters of the multilayer perceptron. An artificial neural network (ANN) scheme is obtained empirically, using the analysis of existing works, the optimal number of the hidden layers is identified. The obtained ANN contains three modifiable parameters: activation functions on the first and the second hidden layers, a way to optimize the gradient descent algorithm. The activation functions and optimization methods for the gradient descent algorithm are discussed in this study. A comparison of the ROC-AUC classification characteristics (area under the error curve) of all the constructed ANNs showed which activation function and gradient descent algorithm optimization method is the best of the considered parameter combinations when using a multilayer perceptron scheme to solve CA identification problems.