Research On Binary And Multi Classification Steganalysis Based On Deep Learning

Steganography is a very ancient information hiding technology,which was used for secret communication in ancient Greece.Modern steganography technology has developed rapidly with the popularization of Internet media.In the era of big data,digital information such as digital images,videos,audios,texts,and even network traffic has sprung up,and there are inexhaustible carriers for steganography.Digital media exchanged through peer-to-peer communication makes it difficult to discover the information hidden in these media.And images,as the most accessible and easy-to-process digital carrier on the web,are one of the perfect choices for steganography.Any technology has two sides,and modern digital steganography is no exception,but modern digital steganography is used more for the bad side.Due to the low transmission cost and easy availability of digital images,image hiding has developed rapidly in the past decade.At the same time,the proliferation of digital images and development of image information hiding have provided a good medium for criminals to commit crimes.In order to prevent the malicious use of digital image steganography,digital image steganalysis technology has also been developed rapidly.Steganalysis analyzes the principles and characteristics of steganography to determine whether the carrier contains hidden information,which can timely interception,preventing the transmission and misuse of malicious hidden information.Meanwhile,with the development of convolutional neural networks,steganalysis researchers have tried to combine steganalysis and convolutional neural networks to construct more accurate steganalysis models,resulting in many excellent steganalysis algorithms.There are problems such as weak model versatility and practicability,excessive dependence on device computing power,and inability to classify specific steganographic algorithms.This paper proposes solutions based on the problems in the existing models.The specific work is as follows:1.A multi-feature fusion steganalysis method based on SE and Rep Vgg modules is proposed.Firstly,the theoretical implementation of multi-feature extraction and fusion,the roles played by SE module and Rep Vgg module,and the experimental comparison of specific parameters are introduced.By introducing the concept of multi-feature extraction and fusion,the versatility of the model is greatly increased,so that the model has a high detection accuracy rate for multiple steganographic algorithms,and at the same time,it solves the dependence of the network model on the size of the input image,and improves the performance of the model.The practicality of the network model;the experimental results show that under different payloads,the detection accuracy of SFRNet for HUGO,WOW,S-UNIWARD and Mi POD has different degrees of improvement compared to existing algorithms.2.Aiming at the problem that the existing steganalysis model can only classify the secret information and the original image and cannot distinguish the specific algorithm,the multi-classification steganalysis model is discussed and studied.Firstly,the features extracted by the three feature extraction methods are processed;Then the Res Net and Efficient Net V2 networks are respectively transformed to better adapt to the field of steganalysis,and the Gaussian nonlinear activation function is added to make them more sensitive to steganographic signals;Finally three feature extraction methods and the two network structures were combined for experiments,and 8 different results obtained from the experiments were compared and analyzed in an all-round way.Finally,an optimal combination method was selected as the multi-class steganalysis model.The results show that the max SRMd2 feature extraction method combined with the modified Efficientnet V2 model has a high classification accuracy for the four steganography algorithms,the classification accuracy of 63.8% is achieved.