Optimization of neural network inputs by feature selection methods

The main idea of this paper is to compare feature selection methods for dimension reduction of the original dataset to reach optimization of steganalysis process by artificial neural networks (ANN). Feature selection methods are tools based on statistic exploited in pre-processing step of data mining workflow. These methods are very useful in a dimension reduction, removing of insignificant data, increasing comprehensibility and learning accuracy. Dimension reduction leads to reduced computational resource consumptions, which is validated by ANN simulations. Steganalysis is a field of the computer security, which deals with a discovering of hidden information in images which is normally unrecognizable. All dataming processes, which reduce the dimension of ANN input layer, should keep accuracy of steganalysis on the original level. © ECMS.internal grant agency of Tomas Bata University in Zlin [IGA/44/FAI/10/D]; Ministry of Education of the Czech Republic [MSM 7088352101]; Grant Agency of Czech Republic [GACR 102/09/1680]; European Regional Development Fund under the Project CEBIA-Tech [CZ.1.05/2.1.00/03.0089