Image Steganography Location Research Based on Pixel Probability Prediction.

In order to further enhance the practicability of image steganalysis, this paper expands the research goal of image steganalysis as steganography location of adaptive steganography and non-adaptive steganography LSB matching, an end-to-end steganography localization network PSL_NET is proposed. Input an image at the input end, and locate the position of the steganographic pixel of the image at the output end. In the preprocessing layer, the high-pass filter of the spatial rich model is used to extract the residual noise images. In the depth residual layer, deep residual learning is used to enhance the expression ability of steganographic features. In the pixel prediction layer, using the mask image which marks the actual position of the steganographic pixel to perform supervised learning, as well as treating the pixels in the smooth or texture area without distinction, the probability whether each pixel is steganographic pixels is predicted, and finally predict the steganographic pixels of the input image. The imbalance problem of positive and negative samples are solved from the perspective of objective function to improve the detection accuracy. In the experiment based on BOSSbase v1.01, when the network predicts the steganographic image of the adaptive steganography algorithm S-UNIWARD at the payload of 0.4 BPP, the pixel detection accuracy is 0.981 74, and the experiments also verify the network can detect the steganographic image embedded by non-content adaptive steganography LSB matching. [ABSTRACT FROM AUTHOR]