ARTIFICIAL NEURAL NETWORK MODEL OPTIMIZATION FOR NETWORK PACKAGE CLASSIFICATION

Abstract

The classification of internet networks is widely used by various parties for saving, allocating, limiting, and administration of internet resources. Various techniques are used, one of which is a machine learning approach, especially an artificial neural network algorithm, hereinafter abbreviated as ANN. ANN works by imitating the way the nerves of the human brain work. ANN in this study is used to classify network packets. The dataset used consists of 990558 rows of data, 4 input columns (X) including protocol, port, time before previous packet, packet length. Output column (Y) consisting of 5 labels (App, Social Media, Game, Browsing, Streaming). Furthermore, this dataset is divided into 3 datasets namely training, test, and validation. The purpose of this study is, first, to determine the ability of the ANN algorithm for classifying packets on a per-packet basis. Second, find the most optimal ANN model for the above problem. The conclusion obtained is that the neural network algorithm can be used in the classification of network packets, but if the dataset used has characteristics such as the number of variables X is small, the data on the variable X is very wide distances such as (port, packet size, time to previous packet, protocol) then high accuracy is very difficult to obtain. Second, if the problem solved has similarities with this study, the most optimal ANN model architecture is: one input layer, one hidden layer, one output layer. The number of input layer neurons is 7 times the number of variables X,  The number of hidden layer neuron is 1 / 2.8 times the number of input neurons, dropout value is 0.33, tanh-tanh-softmax activation method, adamax optimization method. Stable accuracy values are obtained at the 600th iteration (epoch), stable loss values are obtained at the 1000th epoch.