Chinese Journal of Computers   Full Text
Title	Detection of PS and Stego Images Based on Noise Models and Features Integration
Authors	LUO Xiang-Yang LIU Fen-Lin YANG Chun-Fang HE Xiong-Fei
Address	(Department of Network Engineering, Information Science and Technology Institute, Zhengzhou 450002)
Year	2010
Issue	No.6(1060—1072)
Abstract &   Background	Abstract In order to reliably classify the stego images and PS images, which are images modified by some normal image processing operation, and improve the detection accuracy of existing steganalysis algorithms, the different noise models of PS images and stego images are analyzed, and a detection algorithm based on noise models and features integration is proposed. First, the wavelet decomposition of images is made, and then a filtering operation is applied to obtain the wavelet subbands of noise images. Second, some high order absolute characteristic function (CF) moments of histogram are extracted from the wavelet coefficient subbands and their noise versions respectively. Third, these features are integrated as feature vectors. Last, a BP neural network is designed to detect images. In addition, two kinds of typical features, namely the probability density function (PDF) moments and CF moments, are analyzed, and the following conclusion is proved: for wavelet subbands of noise image, absolute CF moments are more sensitive to the changes of an image than absolute PDF moments. A series of experiments are made based on the stego images which embedded with methods such as LSB, LTSB, SLSB, PMK, and PS images with typical operations such as image sharpening, contrast enhancing, adding tags and so on. Experimental results show that the proposed method can effectively detect non-natural images from natural images, and can reliably classify images as stego image and PS image. Keywords image detection; PS image; stego image; noise model; features integration Background Steganalysis is a research hotspot in the information security field. How to improve the detection accuracy of steganalysis algorithms, and how to advance the practicality of steganalysis are the main issues and difficulties that the current researches of steganalysis are facing. The wide application of image processing software products and the non-classification of PS images (which are images modified by some normal image processing operations) and stego images by existing steganalysis algorithms, lead to high error detection ratios. This paper focuses on the classification detection for PS images and stego images, and aims at distinguishing PS images from stego images, so that to improve the correct ratio and the practicality of steganalysis. In the former researches, the authors followed and analyzed a great majority of the relevant references, made a survey on the blind detection technique of image steganography; proposed some detection algorithms and improved algorithms for the LSB (Least significant bit) steganography, the LTSB (Least two significant bits) steganography, and the MLSB (Multiple least significant bits) steganography; and presented some universal blind detection algorithms for image steganography. This paper can be regarded as a further extension of the above-mentioned work. Different noise models of PS images and stego images are analyzed, and a detection algorithm based on noise models and feature integration is proposed. This algorithm can effectively distinguish non-natural images from natural images, and can reliably classify images as stego images and PS images. In addition, two kinds of typical features, namely the probability density function (PDF) moments and characteristic function (CF) moments, are analyzed, and the following conclusion is proved: for wavelet subbands of noise images, absolute CF moments are more sensitive to the changes of an image than absolute PDF moments. This conclusion offers a theory basis for the feature selection and extraction of steganalysis algorithms. This work is support by the National Natural Science Foundation of China (Grant Nos.60902102, 60970141, and 60873249), the Fund of Innovation Scientists and Technicians Outstanding Talents of Henan Province (grant No.094200510008), and the Doctoral Dissertation Innovation Fund of Zhengzhou Information Science and Technology Institute (Grant No. BSLWCX200804)