Research On Key Techniques Of JPEG Steganalysis Based On Stego Noise Characterization

Steganography embeds secret information into ordinary multimedia files such as images and videos to create covert carriers,and using public channels for transmission to achieve covert communication.This provides convenience for criminals to carry out conspiracies,data theft,malicious attacks,and other criminal activities.The goal of steganalysis is to discover and detect covert carriers,track illegal covert communication behavior,and obtain secret information transmitted through public channels.Nowadays,JPEG images have become a common covert carrier due to their widespread application and information redundancy.Therefore,research on JPEG image steganalysis is of great practical significance for safeguarding national secrets,commercial intelligence,and personal privacy.In the process of JPEG image steganalysis,the accuracy of extracting stego noise plays a decisive role in the detection accuracy of steganalysis algorithms,and can provide theoretical and technical support for tasks such as steganography identification,embedding rate determination,and embedding position analysis.However,existing methods suffer from problems such as difficulty in accurately extracting stego noise in high-quality images and difficulty in accurately identifying stego noise regions in images,resulting in low detection accuracy.To address these issues,this thesis conducts research on JPEG image steganalysis technology based on stego noise characterization,and carries out research work from two perspectives: improving the accuracy of stego noise extraction and enhancing the ability to identify stego noise regions.The main work and innovation points of this thesis are as follows:1.Existing methods do not have optimized goals for extracting stego noise.The stego noise extracted from high-quality images contains a large amount of redundant information,and the accuracy of stego noise extraction is insufficient,which affects the detection accuracy.Therefore,this thesis proposes a JPEG image steganalysis method based on deep extraction of stego noise.This method uses a supervised training strategy to set optimization objectives for the designed deep stego noise extraction network,thereby suppressing the influence of redundant information and improving the accuracy of stego noise extraction.A model evaluation index is proposed to guide the deep stego noise extraction network,obtaining stego noise with both accuracy and discrimination,and improving the detection accuracy.Experimental results on two large-scale public datasets,BOSSBasev1.01 and BOWS2 for two adaptive JPEG image steganographic algorithms,J-UNIWARD and UED-JC,show that compared with the classical methods SRNet and J-Xu Net,the proposed method achieves better detection performance with an accuracy improvement of up to 5.24% and 13.18%,respectively.When the image quality factor is high,the proposed method also outperforms the latest methods,EWNet and CSANet,with an accuracy improvement of up to 2.22% and 3.79%,respectively.2.The distribution of stego noise in space is not continuous,and existing methods have failed to accurately identify the distribution area of stego noise in images when extracting stego noise,resulting in lower detection accuracy.Therefore,this thesis proposes an adaptive stego noise extraction network using parallel dilated convolution.This method increases the network’s receptive field through parallel dilated convolution to improve the network’s adaptability to stego noise areas.It uses inverted bottleneck module containing large convolution kernels to obtain richer and more accurate stego noise features,thereby improving the accuracy of detection.Experimental results on two large public datasets,BOSSBasev1.01 and BOWS2,show that compared with the latest methods,EWNet and CSANet,and classic methods,SRNet and JXu Net,our method achieved significant improvements in the detection of J-UNIWARD and UED-JC,two adaptive JPEG steganographic algorithms.It can increase the detection accuracy by up to 6.25%,7.52%,12.56%,and 18.65%,respectively.Finally,the full dissertation work is summarized,and the next research direction is prospected.