Ensemble SW image steganalysis: A low dimension method for LSBR detection

摘要

Blind steganalysis examines digital media for the likely existence of hidden messages, without prior knowledge of the steganographic algorithm that may have been used to hide such messages. This paper puts forward a novel learning-based blind image steganalysis method for tackling spatial domain least significant bit (LSB) flipping. Its key steganalytic feature is the correlation between message length and the regression of the quantity of intensity-identical pixels and channels. A specially designed support vector machine (SVM) kernel is trained to analyze each pixel as an individual analysis unit, with the combined results of the analysis determining the ultimate steganalysis decision. This method makes a number of innovative contributions to the field of blind steganalysis. First, it offers a novel steganalytic feature for measuring similarity between the weight of pixels and channels. Second, it involves pixels in the steganalytic process according to the degree of their detected membership, thus avoiding neutral pixels influencing the process. Third, it extracts reference statistical behavior from cover and stego pixels, thereby enhancing the sensitivity of the steganalyzer. Fourth, its SVM kernel enhances sensitivity using statistical functions combined with trapezoidal fuzzy membership. Finally, with all these innovations it is capable of achieving a sensitivity of 99.626% for 0.25 bpp stego images through only two analysis dimensions.