СТАТИСТИЧЕСКИЙ СТЕГАНОАНАЛИЗ ФОТОРЕАЛИСТИЧНЫХ ИЗОБРАЖЕНИЙ С ИСПОЛЬЗОВАНИЕМ ГРАДИЕНТНЫХ ПУТЕЙ

Abstract

Purpose of the article is experimentally test the efficiency of the feature vector based on gradient paths in the spatial domain of the image. Research method is a comparison of steganalytical feature vectors based on the mean square error and the coefficient of determination obtained using SVM-regression in Matlab. The dataset is formed by automating the freeware steganoprograms that implement embedding into the spatial area of the image with sequential and pseudorandom selection of a pixel for embedding. Results of the study: the optimal parameters of the algorithm for seeking gradient paths from the point of view of embedding detection are experimentally obtained. The results of applying machine learning models are obtained and analyzed, the optimal scale of the SVM-regression kernel is determined. The computation durations of feature vectors obtaining, models training, recognizing containers are calculated. It is shown experimentally that the gradient paths feature vector is expedient to use for solving problems where it is necessary to vary the detection accuracy depending on functioning capacity of system, because the proposed feature vector allows to determine the dimension / accuracy ratio. Also, by experiment, a complex 20D vector is selected from several one-dimensional quantitative steganodetectors and the gradient paths feature vector. The effectiveness of result vector is comparable to the 686D feature vector SPAM.