Research On Digital Image Steganalysis Under The Condition Of Mismatch

As a new means of ensuring information security,steganography has been widely valued by researchers and applications in the past two decades.As the opponent of steganography,steganalysis has also been greatly developed.It plays an important role in maintaining commercial information and national security,and has high application value and academic significance.State-of-the-art steganalysis mainly relies on machine-learning based classifier methods.In the academic world steganalysis has achieved good detection results in the laboratory environment,whereas there are some difficult problems when put steganalysis into the real-world complex practical applications,such as CSM(cover source mismatch),steganography algorithm mismatch,and embedding rate mismatch problems,etc.The mismatch between the training data and the test data of the classifier will obviously reduce the performance of steganalysis.This dissertation proposes two methods to alleviate the decline of steganalysis ability caused by mismatch of cover source or steganography.(1)This dissertation proposes an offline approach to solve the CSM problem by matching “similar” training images for the test images.First,a set of frequency features,Ns ND(non-stego noise distribution)features,are designed to describe the image noise distribution,excluding the noise caused by steganography(which we call non-stego noise in this dissertation).Then,based on the designed Ns ND features,a similar image matching method is proposed.In this method,we use the JS(Jensen-Shannon)divergence to measure the similarity between a test image and images from a large image library that includes various types of images,and we match a most “similar” image set as the training set for the test image.Steganalysis experiments are performed for images embedded by adaptive and non-adaptive steganography methods under four types of CSM scenarios.(2)This dissertation proposes a method based on deep transfer learning to solve the problem of mismatch of steganography algorithms.We construct the training set closest to the test image set to reduce the interference caused by the mismatch of the steganography algorithm,and tailor a suitable CNN-based steganalysis system for each batch of images to be detected.First,we compare the steganographic features in the steganographic algorithm library to calculate the probability of each steganographic algorithm used in the test image,and construct a retraining set from the prediction table.Then,using the knowledge of deep transfer learning,the retraining set is used to fine-tune the initial steganalysis system to make its feature extraction ability more closely match the test image,thereby the performance degradation of steganalysis caused by mismatch of steganography is reduced.Large numbers of experiments for the scenarios of CSM and mismatch of steganography algorithm are conducted.The experimental results show that the Ns ND feature proposed in this dissertation can match the homologous images well,thereby improving the detection effect in the case of image source mismatch;and that the convolutional neural network obtained through deep transfer learning and selective fine-tuning can improve the detection accuracy under the mismatch of steganography algorithm.