A new feature extraction scheme in wavelet transform for stego image classification

In steganalysis, there exist a good number of feature extraction techniques; the most commonly used feature extraction technique is based on the Discrete Wavelet Transform (DWT). Almost all of the techniques give a higher detection rate even for short messages, but they are limited to statistical moments such as Characteristic Function (CF) and Probability Density Function (PDF) moments to build their feature sets. A new feature extraction scheme in wavelet transform is presented in this paper. It is based on the adaptation of Zipf’s law to wavelet transform to extract new features obtained from the statistical distributions of wavelet pattern subbands that are represented by a graphical representation called a curve of Zipf. The originality of this work can be shown through two perspectives: the adaptation of Zipf’s law to wavelet subbands, which involves the determination of the pattern size used to count the frequency of pattern appearance, and an encoding model, which makes the distribution of pattern frequencies more significant. Moreover, this novel scheme proposes a new feature set extracted from the curve of Zipf of the wavelet subband coefficients, which is the first such attempt in DWT and the field of steganalysis. A Random Forest classifier (RF) is then used to model the resulting features. Experimental results show that the proposed steganalytic method has the best classification performance in terms of accuracy, which achieved 88.9, 78.25, 84.27 and 60.1%, compared with three feature sets based on CF moments and Markov features, histogram features, and co-occurrence matrix features, respectively.