| With the continuous development of computer technology and economy,China’s power system construction is gradually building a smart grid towards digitization and intelligence.The security issues of smart grids are divided into physical security and information security.Information security issues not only threaten the information system of smart grids,but also have an impact on their physical systems and disrupt their normal operation.Image steganography can hide secret information into images,which can not only protect information security,but also be easily used by criminals to threaten the information security of smart grids.Image steganalysis,as the confrontation technology of image steganography,can detect whether secret information is hidden in digital images,and can detect the images transmitted to power enterprises to prevent images containing malicious information from threatening the safe and stable operation of smart grid.Therefore,it is very important to study image steganalysis technology to maintain the information security of smart grid.With the rapid development of deep learning,a large number of deep learning image steganalysis networks are constantly emerging.However,most high-performance deep learning image steganalysis have a large number of parameters and require a large amount of memory space in practical use.In response to this issue,this article proposes a lightweight and effective deep learning image steganalysis network named LWENet.Firstly,the preprocessing section uses a spatial rich model(SRM)high pass filter combined with bottleneck residual blocks(BRB)to improve the signal-to-noise ratio(SNR)of the steganographic signal while maintaining lightweight.Then,multi-view global pooling(MGP)was proposed to generate rich classification features,and in order to maintain lightweight,only one fully connected(FC)layer was selected as the classifier in the classification section.Finally,using depthwise separable convolution(DWSConv)at the end of network feature extraction significantly reduces the number of network parameters and improves performance.The experimental results demonstrate that LWENet has the smallest number of parameters and significant advantages in detection performance compared to other comparative methods,whether in standard datasets or datasets in the field of electrical engineering.In the real-world scenario,the cover images used for training deep learning steganalysis networks is usually different from the source of the detected images,which is known as cover source mismatch(CSM)phenomenon,which can greatly reduce the performance of deep learning image steganalysis networks.In response to this issue,this article designs a deep learning image steganalysis network called CAA-Steg.This network consists of a backbone image steganalysis network and a contrastive domain discrepancy(CDD)based on reliable steganalysis labeling(RSL).Due to its use of sample category information to reduce domain discrepancy between the target domain and the source domain,CDD can be used in image steganalysis to solve CSM problems.However,the clustering algorithm used in CDD cannot provide reliable and effective pseudo labels for weak steganographic signals,resulting in poor detection performance in the target domain.In view of this,this article proposes an RSL based CDD to generate extended target domain samples with pseudo labels to help CDD achieve better detection performance.In addition,CAA-Steg has a certain degree of universality as it can combine various backbone image steganalysis networks.Numerous experiments have shown that compared to J-Net,CAA-Steg has achieved significant performance improvements. |
