Стегоанализ цифровых изображений с использованием наивного байесовского классификатора
Steganalysis of digital images includes creation of feature space, formation of learning selection and qualifier training. Some known steganalysis systems offered for practical use work in space with dimension in tens of thousands of features. Therefore researches directed to identification of the most informative features including indicating use when embedding concrete steganographic primitives are relevant. In the real research the set consisting of 43 features is created and assessment of informational content of separate groups of features in relation to the steganalysis of classical steganographic algorithms is carried out. For preparation of the learning and test selection of images three steganographic algorithms are used: F5, JSteg and PM1.Steganalysis is carried out by means of the naive Bayes classifier. It is shown that use of full features set leads to the smaller accuracy of classification in comparison with use of separate features groups entering full set. The research of influence of distortions of non-steganographic character on the steganalysis accuracy is in addition conducted. For this purpose test selection has included a set of the images which aren’t containing attachments, but processed by means of the Prisma application which allows to imitate style of the famous artists’s pictures. It is established that such processing allows to compromise the accuracy of detection of “empty” images, average quantity of the mistakes made at classification of such images is about 40 %. This observation allows to offer new approach to embedding of information into digital images which is in that when embedding to imitate non-steganographic distortions.