Signal Processing Research Volume 2 Issue 1, March 2013 www.seipub.org/spr 17 Texture Based Steganalysis of Grayscale Images Using Neural Network Arooj Nissar *1 , A. H. Mir 2 1 Department of Information Technology, 2 Department of Electronics and Communication, National Institute of Technology, Srinagar, India *1 chotz786@yahoo.com; 2 ajazhmir@yahoo.com Abstract A method of steganalysis of grayscale images based on texture analysis using Spatial Gray Level Dependence Method is presented. A neural network trained with the texture related statistics, extracted from cover images and stego images that are created using one embedding algorithm, is adopted as a classifier. The classifier reliably detects clean images and stego images despite the small data embedding rate of 0.0077. Steganographic tools tested include both spatial and transform domain image hiding techniques. It is worth mentioning that the clean images are effectively distinguished from the stego images on the basis of image texture alone, regardless of the embedding algorithm used. Keywords Steganography; Steganalysis; Texture Analysis; SGLDM; Spatial Gray Level Dependence Method; Neural Network Introduction Steganalysis is the art and science of detecting the presence of hidden information in digital media embedded using steganography (N. F. Johnson, 1998; J. Fridrich et al., 2001; J. Fridrich, 2002). A number of steganalysis techniques have been put forward in the literature. Classification of these techniques is given in (Arooj, 2010). Some of the techniques given in (N. F. Johnson, 1998; M. Niimi et al., 2001; J. Fridrich et al., 2001; Tariq Al Hawi et al., 2004) exploit the signatures introduced by certain steganographic algorithms while as the techniques given in (I. Avcibas et al., 2001; I. Avcibas et al., 2002; J.J. Harmsen, 2003; R. Chandramouli, 2003; J. Fridrich et al., 2003; S. Dumitrescu et al., 2003; K. Sullivan et al., 2004; K. Sullivan et al., 2005; Xiang-dong Chen et al., 2006) identify characteristic side effects in statistics caused due to embedding algorithm and exploit the same for steganalysis. The above techniques prove useful only if the steganographic algorithm used is known. The other set of techniques given in (S. Lyu, 2002; S. Lyu, 2004; Wen-Nung Lie, 2005; G. Xuan et al., 2005; Yun Q. Shi et al., 2005; Xiaochuan Chen et al., 2006; Patricia Lafferty, 2004) find out some appropriate statistics that are changed due to any steganographic algorithm and hence do not depend on the behavior of embedding algorithm. Such techniques can be fruitful in a practical scenario of steganalysis wherein there is neither any scope of comparison with the original images nor the embedding algorithm is known apriori. Because of nature of field, the steganographic algorithms are constantly enhanced rendering the proposed steganalytic techniques useless. Hence with the ever increasing number of steganographic techniques, it will be reasonable to devise a steganalytic technique that can detect the presence of secret information embedded using any steganographic algorithm. Work has been carried out for steganalysis of images based on texture analysis and is reported in (Patricia Lafferty, 2004; Shaohui Liu et al., 2004). The technique for steganalysis of images given in (Shaohui Liu et al., 2004) analyses texture in wavelet domain and the technique proposed in (Patricia Lafferty, 2004) analyses texture in spatial domain by deriving first order statistics (statistical texture analysis). Literature survey (Arooj, 2010) reveals that second order texture statistics has not been analysed for the purpose of steganalysis. In this paper, we propose a new steganalysis technique based on statistical texture analysis in spatial domain deriving second order statistics using Spatial Gray Level Dependence Method (SGLDM) (R. M. Harlick et al., 1973). The rest of the paper is organized as follows: section 2 deals with texture analysis of images. In section 3, neural network is trained with the features