Locating Secret Bits Flipped By Adaptive Steganography

As an important research topic in the field of information security,digital image steganography and steganalysis have attracted more and more attention.Current steganalysis capable of locating steganographic payload has only been investigated in non-adaptive domain while the researches of locating hidden bits modified by adaptive steganographic algorithms have not been studied yet.This thesis focuses on detecting the existence of steganography and locate the steganographic payload.Aiming at the problem of adaptive steganography existence detection,this work proposes to detect it based on a convolutional neural network.The pre-processing layer filters the lowfrequency signal of the image using a high-pass filter to obtain the noise residual,the convolutional layer automatically extracts the steganographic features,and the fully connected layer implements the classification of features.In addition,operations such like multiple activations,batch normalization,and absolute functions has also improved the detection performance of the network.The experimental results conducted on BOSSbase dataset show that the network can achieve a high detection rate of 82.98%.Aiming at the problem of adaptive steganography payload localization,this work proposes to make fuzzy localization of adaptive steganographic payload based on modification map.The algorithm generates random secret information at the same payload and embeds it into the stego image to obtain the modification map between the stego image and the embedded image.Then,the modification map is extended with a given margin to locate the modified pixels.Experiments on BOSSbase illustrate that when the margin value approaches 9,the proposed locating algorithm can achieve a high recall rate of 95.27%.In order to improve the locating accuracy of the payload location algorithm based on the modification map,this work proposes the localization method of adaptive steganographic payload based on pixel weight.The advanced algorithm takes the Euclidean distance between image pixels as the weight factor and defines a weighted cost matrix of the cover image,and then sorts the pixels within the modification map according to their weighted distortion cost.The experimentalresults on BOSSbase have verified that compared with the payload localization algorithm based on modification map,the proposed method improves the locating accuracy by 17.63% but without reducing the recall rate.