The Research On JPEG Image Mismatched Steganalysis Based On Feature Transfer

The rapidly development of Internet,computer and digital communication technology promote the progress of information age.So more and more people pay attention to the information security.Steganography,as an important part of information security,is the art of hiding data into the public digital medium and concealing the existence of the hidden data.However,on the one hand steganography facilitates privacy preserving,on the other hand it can be employed by some criminals for malicious use.Steganalysis,which is as a countermeasure technology to steganography,has got great development.JPEG images have been widely accepted and used in the field of steganography and steganalysis for the characteristics of simple and convenient.The conventional JPEG steganalysis has assumed that both training set and testing set have same feature distribution.Under practical application environment,however,the difference of image cover source,steganographic schemes and image editors can make feature distribution between training set and testing set different,which can result in mismatched phenomenon and the degradation in detection accuracy.Recently years,the research on the JPEG steganalysis has just begun.In this paper,the matched and mismatched JPEG steganalysis are introduced in detail.The mismatched factors and solving methods are summarized.Aiming at mismatched steganalysis under practical application environment,this paper based on feature transfer proposes two JPEG mismatched steganalysis.The mainly research contents of this paper are summarized as follows:(1)This paper studies the basic conception and development background of matched steganalysis under laboratory environment.Then this paper concludes the research framework,including image dataset,steganalysis feature and classification method.Based on these,this paper analyzes mismatched steganalysis under practical application environment,illustrates the basic definition and mismatched factors,and summarizes the development of mismatched steganalysis at present.(2)This paper studies the influence of quantization table and steganographic scheme mismatch for steganalysis.For quantization table and steganographic scheme mismatch under same cover source,this paper proposes a novel JPEG mismatched steganalysis based on multi-order feature alignment,which can obtain new feature representations by low-order feature alignment,higher-order feature alignment and iterative calibration.This method can increase the feature distribution similarity between training set and testing set,and relieve the influence of quantization table mismatch and steganographic scheme mismatch.(3)This paper studies the influence of mixed cover source for steganalysis.For steganographic scheme mismatch under mixed cover source,this paper puts forward JPEG mismatched steganalysis based on contribution feature transfer,which can combine digital forensics and transfer learning.This method can identify the source of each testing image by source analysis stage,and select useful features from training set by transfer stage which can decrease difference of feature distribution between training set and testing set.This method can obtain new feature representations and more discriminated models.The experimental results illustrate that the proposed menthod have improved 5% to 10% of detection accuracy comparing with other existing steganalysis methods.