Abstract
Information hiding for covert communication is rapidlygaining momentum. With sophisticated techniques beingdeveloped in steganography, steganalysis needs to beuniversal. In this paper we propose Universal Steganalysisusing Histogram, Discrete Fourier Transform and SVM(SHDFT). The stego image has irregular statisticalcharacteristics as compare to cover image. Using Histogramand DFT, the statistical features are generated to train One-Class SVM to discriminate the cover and stego image.SHDFT algorithm is found to be efficient and fast since thenumber of statistical features is less compared to theexisting algorithm.Keywords: Universal Steganalysis, Cover-Medium, Payload,Redundant Bits, Total Cover Capacity, Histogram, DFT,SVM.
I. INTRODUCTION
Since long time ago, surreptitious communicationand exchange of information is well known. There arenumerous illustrations depicting covertness used incommunication. Steganography, water marking,cryptography are some of the techniques used for hidinginformation [1]. There are several mediums for hidinginformation, such as digital images, text, audio, webpages, video, etc. Steganography is a skill and disciplineof embedding a secret payload into a cover medium. Theredundant bits in the cover medium are identified andreplaced with the payload. The intent of steganography isensuring that the presence of hidden message isundetectable. The different techniques used forembedding of data are Least Significant Bit (LSB),Discrete Cosine Transform (DCT), Discrete WaveletTransform (DWT), Spread Spectrum (SS) and Palettetechnique. There are software tools available over theInternet for embedding the payload in digital images andvideos viz., S-Tools, J-Steg, Outguess, F5 and Steghide.The steganography is being used by criminals andterrorists to exchange information regarding their illegalactivities. The government and other standardorganizations are using steganalysis to restrain theseillegal activities. Steganalysis is to discover andrecognize the extent of a hidden message and can beeither blind (universal) or non-blind (embeddingspecific). In case of universal technique, the embeddingscheme is unknown; so attempt is made to detect theexistence of hidden data. Whereas in embedding specific,the embedding technique is known; therefore payloadsize is estimated. There are Steganalysis Tools availablefreely or as commercial software. The Stegdetect Tool,detects Jsteg, Jphide, invisible secrets, Outguess, F5,camouflage Steganographic schemes in JPEG images.The Tools developed using the Chi-Square analysis,performs a statistical attack to detect hidden data in BMPimages.The success of steganography depends on variousaspects such as the algorithm used for embedding,compression algorithm used and alteration of imageproperties. LSB is most commonly used technique forimage steganography, which uses bitwise methods tomanipulate LSB of the cover image. The minute changesof LSB are imperceptible to human eye and the methodof hiding information in LSB can be analogized byadding noise to the image. Reliable algorithms forcompression used in Steganography are Windows Bitmap(BMP), Graphics Interchange Format (GIF) and JointPhotographic Experts Group (JPEG), to ensure thehidden information is not lost after transformation.Lossless compression algorithms such as BMP and GIFformats are chosen for LSB techniques in whichmodifications are usually made in spatial domain. JPEGis losy compression algorithm where data hiding isusually done in frequency domain.General Steganalysis method has not been developedsince every Steganographic method utilizes differentmethods of embedding the payload. In classicalSteganographic schemes, the security lies in theconcealment of the encoding technique whereas themodern schemes adopt Kerchoff’s Principle ofCryptography, and hence the security depends on thesecret key that is used to encode the payload.There are several Classifiers being used for patternrecognition: Bayesian Multi-Variate, Fischer LinearDiscriminant (FLD), Neural Network (NN), and SupportVector Machines (SVM). SVM is based on statisticallearning theory which is highly dependent upon thekernel functions viz., Gaussian Radial Basis Function,Polynomial, Exponential Radial Basis Function, Multi-Layer Perceptron, Splines, BSplines, Additive Kernelsand Tensor Product are used for mapping features. SVMcan be classified as One-Class and Multi-Class. In OneACEEEInternational Journal on Signal and Image Processing Vol 1, No. 1, Jan 201018ISSN 2152-5056 © 2010 ACEEEDOI: 01.ijsip.01.01.04Class SVM, only the features of cover images arerequired for classification of cover and stego imagewhereas in Multi-Class SVM, the features of both thecover and stego-images are required for discriminatingthe cover and stego image. Owing to simple geometry,SVMs are used more predominantly than NeuralNetworks. Biggest limitation of SVM is choosing of thekernel function, speed and size, both in training andtesting.Contribution: In this paper, we have proposedSHDFT Universal Steganalysis Method. Histogram andDiscrete Fourier Transform of color images are computedto obtain the PDF moments. In addition, the energy of theimage is considered. The non-linear Classifier SupportVector Machine is used for classification of cover andstego image.


Download full report
http://academypublisherijrte/vol01/no01/ijrte0101367371.pdf