Machine Learning in Image Steganalysis

Chapter 8

Steganalysis in the JPEG Domain

Much of the research in steganography and steganalysis has focused on JPEG images, due to the sheer popularity of the format. It has been in widespread use, known to expert and non-expert users alike, for more than 15 years. Even though other formats may claim similar popularity at one time or another, they have been more short-lived.

Conventional wisdom indicates that steganalysis is most effective when the features are calculated directly from the domain of embedding. This is the ‘firm belief’ of authors like Goljan et al. (2006), and Fridrich et al. (2011a) attribute the idea to Fridrich (2005). The effect of the embedding can also more easily be understood in the embedding domain, and thus it is also easier, in general, to analyse the effect of embedding on features extracted from the same domain.

Since steganographic embedding in the JPEG domain is so popular, a considerable number of JPEG-based feature sets have emerged as well. The most well-known one seems currently to be the 219-dimensional feature vector PEV-219 (also known as PEV-274) of Pevný and Fridrich (2007), and the slightly simpler variation NCPEV-219. These feature vectors are very interesting not only because of their good performance, but also because they combine a range of different techniques. In fact, most of the features we introduce in this chapter are used in NCPEV-219 and PEV-219, and we will highlight those as we go along. Experiments have shown negligible ...