CN111696021B - Image self-adaptive steganalysis system and method based on significance detection - Google Patents

Abstract

The invention belongs to the technical field of image processing, and particularly discloses an image self-adaptive steganalysis system based on saliency detection. Also disclosed is an analytical method, specifically: firstly, inputting an image with errors detected by a discriminator module into a saliency detection module to form a saliency map; then extracting a significance map meeting the requirements by a region screening module, and carrying out image fusion on the significance map and the corresponding original image to form a significance fusion map; and finally, replacing the saliency map which does not meet the requirement with an original image, combining the part of original image and the saliency fusion map into an updated data set, inputting the updated data set into a discriminator module for training, and enabling the discriminator to carry out targeted feature learning on the region with higher coincidence degree with the steganographic region. The method utilizes the significance detection technology to guide the steganalysis model to pay more attention to the characteristics of the image steganalysis area, thereby improving the training effect and the detection accuracy of the model.

Description

An image adaptive steganalysis system and method based on saliency detection

Technical Field

The invention belongs to the technical field of image processing, and relates to an image adaptive steganalysis system and method based on saliency detection.

Background Art

Image steganography is a covert communication technology that embeds secret messages in image carrier files for transmission. Unlike traditional encrypted communication methods that make it difficult for third parties to decipher, image steganography conceals the communication behavior itself, making it difficult for third parties to detect the embedding of secret messages, so it is very deceptive. Image steganalysis is specifically used to detect messages embedded using image steganography. It is a pair of technologies that compete with image steganography and promote each other's development in the process of competition. The difficulty of image steganalysis lies in the fact that the information embedded in the image by the steganographic operation is very small, and the slight difference between the image before and after can be easily covered up by the difference between the image content. In particular, the image adaptive steganography proposed in recent years gives priority to embedding secret information into the complex texture area of the image when performing the steganographic operation. In the case of low embedding rate, it is more difficult to detect, which greatly improves the security of steganography and brings great challenges to image steganalysis.

The main idea of the content adaptive steganography algorithm based on the combination of distortion function and STC includes two parts: quantitative analysis of the change cost based on the distortion function and embedding based on STC. The purpose of the distortion function is to capture the changes of local or global features after embedding through quantitative analysis, such as calculating the possible distortion after each element changes. The purpose of STC is to determine the elements to be changed in the end according to the distortion situation, so as to minimize the overall distortion. The most commonly used adaptive steganography algorithms are HUGO, WOW, S-UNIWARD, UED, and J-UNIWARD. The modification probability of elements in the complex texture area of the image is greater than that of elements in the smooth area. The reason is that the disturbance caused by the steganographic algorithm in these statistically complex areas is less noticeable.

However, most of the existing research work on image steganalysis is to improve the network structure to enhance the detection performance of steganalysis based on the network model. However, when using image adaptive steganography for steganography, the secret message is not embedded in all areas of the image, and the image contains rich information dimensions, so the information contained in the image is not completely beneficial to the training. This situation makes the model have unnecessary redundant interference during training, resulting in reduced detection accuracy. Therefore, it is necessary to propose a steganalysis method that is highly targeted at the steganalysis area, guiding the model to pay more attention to the characteristics of the steganalysis area.

Summary of the invention

In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide an image adaptive steganalysis system and method based on saliency detection, which uses saliency detection technology to guide the steganalysis model to pay more attention to the characteristics of the image steganographic area, thereby improving the training effect of the model and the detection accuracy.

The present invention is achieved through the following technical solutions:

An image adaptive steganalysis method based on saliency detection comprises the following steps:

1) Segment the salient regions of the image in the image with the detected error to form a saliency map;

2) According to the overlap between the salient area and the steganographic area, the saliency map is screened, and the saliency map that meets the requirements is extracted, and the saliency map that meets the requirements is fused with the corresponding original image to form a saliency fusion map; wherein the saliency map that meets the requirements is an image with a high overlap between the salient area and the steganographic area;

3) Replace the saliency maps that do not meet the requirements with the original images, and combine these original images and the saliency fusion maps into an updated dataset;

4) Then use the updated dataset for training.

Furthermore, in step 1), a discriminator module is used to detect erroneous images, and the discriminator module adopts an SRNet model.

Furthermore, in step 2), the image fusion is specifically as follows: setting all pixels in the image except those in the salient area to 0, so that the discriminator module only focuses on the image features in the salient area.

Further, in step 1), a salient region of the image is segmented using a saliency detection module;

The saliency detection module adopts the BASNet model, including the prediction module and the multi-scale residual optimization module, to form a saliency map as follows:

The prediction module and multi-scale residual optimization module of the BASNet model are introduced into the network, and a rough saliency map is obtained through the prediction module;

The multi-scale residual optimization module optimizes the rough saliency map of the prediction module by learning the residual between the rough saliency map and the true annotation, and finally obtains the refined saliency map.

Furthermore, in step 2), the saliency map is screened using a region screening module to extract saliency maps that meet the requirements.

Furthermore, in step 2), the calculation method of the overlap degree η between the salient area and the stego area is as follows:

From formula (1), we can get formula (2):

Among them, N represents the total number of pixels in the image, N _coin represents the number of pixels in the overlapping area, N _stego represents the number of pixels in the steganographic area, P _Stego (i, j) and P _SOD (i, j) represent the pixel values of the stegographic point map and the saliency map at position (i, j), respectively.

Furthermore, in step 2), the overlap degree corresponding to the saliency map that meets the requirements is 0.6-1.

The present invention also discloses an image adaptive steganalysis system based on saliency detection, comprising a saliency detection module, a region screening module and a discriminator module;

The saliency detection module is used to generate the salient area of the image to be tested. It adopts the BASNet model, including the prediction module and the multi-scale residual optimization module.

The prediction module is used to obtain a rough saliency map; the multi-scale residual optimization module is used to optimize the rough saliency map of the prediction module by learning the residual between the rough saliency map and the true annotation, and finally obtain the refined saliency map;

The region screening module is used to screen the saliency maps and extract the saliency maps that meet the requirements;

The discriminator module adopts the SRNet model, which is used to provide images with initial detection errors and retrain on the updated dataset.

Furthermore, the multi-scale residual optimization module contains an input layer, an encoder, a bridge layer, a decoder and an output layer.

Furthermore, both the encoder and decoder have four stages, each with only one convolutional layer, and each layer has 64 filters of size 33;

The bridge layer also sets up a convolutional layer with the same parameters as the other convolutional layers.

Compared with the prior art, the present invention has the following beneficial technical effects:

The invention discloses an image adaptive steganalysis method based on saliency detection. First, after processing the image with detection error, a saliency map is formed; then, according to the overlap between the saliency area and the stego area, the saliency map is screened, and the saliency map that meets the requirements is extracted, and image fusion is performed with the corresponding original image to form a saliency fusion map; finally, the saliency map that does not meet the requirements is replaced with the original image, and the updated data set composed of this part of the original image and the saliency fusion map is used for training, and targeted feature learning is performed on the area with a high overlap with the stego area. Compared with the previous convolutional network model, the model can be guided to learn the features of the image stego area, which has better pertinence. The invention performs data statistical analysis on the training set, extracts the image that meets the conditions for processing, and can ensure the effectiveness of the processing. By performing steganalysis on the data set embedded with the adaptive steganography algorithm in the air domain and the JPEG domain, the experiment shows that the invention is universal in the air domain and the JPEG domain, and has good overall performance.

Furthermore, through experimental comparative analysis, it is found that images with a degree of overlap between the salient area and the steganographic area concentrated between 0.6 and 1 are screened out, which has a good training effect on the model.

The invention discloses an image adaptive steganalysis system based on saliency detection, comprising a saliency detection module, a region screening module and a discriminator module, simulating a steganalysis system, wherein the discriminator module detects an erroneous image and inputs it into the saliency detection module to form a saliency map; then the region screening module extracts the saliency map that meets the requirements, and performs image fusion with the corresponding original image to form a saliency fusion map; finally, the saliency map that does not meet the requirements is replaced with the original image, and the updated data set composed of this part of the original image and the saliency fusion map is input into the discriminator module for training, so that the discriminator performs targeted feature learning on the area with a high overlap with the stego area. The image randomly downloaded from the network is tested, and compared with the previous work, it can be used to test any image on the network, improve the generalization of the invention, and intuitively output the probability and result of classification, which is more practical and has certain generalization ability and practicality.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is an overall framework diagram of the method of the present invention;

FIG2 is a diagram showing a comparison experiment of saliency and steganographic regions conducted by the present invention;

FIG3 is a statistical diagram of the overlap between the salient area and the steganographic area according to the present invention;

FIG4 is an experimental comparison diagram of the present invention for different training strategies;

FIG5 is a diagram of a steganalysis system simulated by the present invention and a detection result diagram.

DETAILED DESCRIPTION

The present invention is further described in detail below in conjunction with the accompanying drawings:

The invention discloses an image adaptive steganalysis system based on saliency detection, which is mainly composed of three modules: a saliency detection module, a region screening module and a discriminator module.

As shown in Figure 1, the image detected by the discriminator module with errors is first input into the saliency detection module to form a saliency map; then the region screening module extracts the saliency map that meets the requirements and fuses it with the corresponding original image to form a saliency fusion map; finally, the saliency map that does not meet the requirements is replaced with the original image, and the updated data set composed of this part of the original image and the saliency fusion map is input into the discriminator module for training, so that the discriminator can conduct targeted feature learning on the areas with a high degree of overlap with the steganographic area.

The saliency detection module is used to generate the salient area of the image to be tested, and specifically adopts the BASNet model. The present invention introduces the prediction module and the multi-scale residual optimization module of BASNet into the network. The prediction module is a densely supervised encoder/decoder network similar to U-Net. The network learns to predict the salient area from the input image to obtain a rough saliency map; the multi-scale residual optimization module optimizes the rough saliency map of the prediction module by learning the residual between the rough saliency map and the true annotation, and finally obtains the refined saliency map. The central idea of the model algorithm can be expressed as:

S _refined =S _coarse +S _residual

Among them, S _coarse , S _residual and S _refined represent the residual of the prediction module, the residual of the rough saliency map and the true annotation, and the optimized residual, respectively.

The main architecture of the multi-scale residual optimization module is simpler than that of the prediction module, consisting of an input layer, an encoder, a bridge layer, a decoder, and an output layer. Both the encoder and the decoder have four stages, each with only one convolutional layer, each with 64 filters of size 3×3, followed by batch normalization (BN) and ReLU activation; the bridge layer also sets a convolutional layer with the same parameters as other convolutional layers; the maximum pooling layer (maxpool) is used for downsampling in the encoder, and the bilinear upsampling layer (bilinear upsample) is used for upsampling in the decoder. The output model of the multi-scale residual optimization module is the final saliency map. In order to obtain high-quality regional segmentation and clear boundaries, the model defines a mixed loss l ^(k) during training, expressed as:

This loss combines BCE, SSIM and IoU losses, which helps reduce the false errors caused by cross-propagation learning information on the boundary and makes the boundary more refined.

As shown in Figure 2, the salient area of the image is the white area shown in the second row of pictures in Figure 2, and the steganographic area is the scattered area shown in the third row of pictures in Figure 2. The salient area of the image is compared with the steganographic area. After saliency detection, it can be analyzed from the content of the image that when the salient target in the image is clear (such as 4 on the left), the overlap between the salient area marked in the image and the steganographic area is high; and when the salient target in the image is vague (such as 2 on the right), the overlap between the salient area marked in the image and the steganographic area is low. Therefore, it is necessary to screen the saliency map and extract the images that meet the conditions for processing to ensure the effectiveness of the processing.

The region screening module is used to filter out images that meet the processing requirements to ensure the effectiveness of the processing. In this module, the overlap between the salient area and the stego area is first statistically analyzed. In image adaptive steganography, secret information will be embedded in areas with more complex textures, and these areas are more significant areas for the human eye, so they will be marked as salient areas in saliency detection. The BOSSbase1.01 data set is used for data analysis, which contains 10,000 512×512 digital images, which are steganographically written using the J-UNIWARD image adaptive steganography algorithm with an embedding rate of 0.4bpp. In order to analyze the data distribution as a whole, the overlap between the salient area and the stego area of 10,000 images is statistically analyzed and expressed in the form of a scatter plot, as shown in Figure 3.

Through data analysis, it can be seen that the overlap between the salient area and the stego area is concentrated between 0.5-1, and some are still concentrated between 0-0.2, as shown in Figure 2, right 2, which means that not all images are suitable for saliency processing. It is necessary to screen out images that meet the requirements to ensure that the processing results are effective. The calculation method of the overlap between the salient area and the stego area η is as follows:

Where N represents the total number of pixels in the image, N _coin represents the number of pixels in the overlapped area, N _stego represents the number of pixels in the steganographic area, P _Stego (i, j) and P _SOD (i, j) represent the pixel values of the stego point map and saliency map at position (i, j) respectively. The stego point map represents the pixel positions that have changed after being steganographically written by the J-UNIWARD adaptive steganographic algorithm.

In the region screening module, through experimental comparative analysis, the images with the overlap between the salient region and the steganographic region concentrated between 0.6-1 are screened out, which has a good training effect on the model. The comparative experiment is shown in Table 1. The threshold K of the screening area is set to 0.7, and the training effect is the best. After the saliency map is screened out, the image fusion technology is used to fuse the saliency map with the original image, so that the pixels in the image except the pixels in the saliency region are set to 0, that is, the model only focuses on the image features of the saliency region.

Table 1 Detection accuracy under different area screening thresholds (%)

The discriminator module is used to provide images with initial detection errors and retraining of updated data sets. The method of the present invention uses the SRNet model as the discriminator in the experiment. The SRNet model consists of four parts. The two front-end parts (1-7 layers) are responsible for extracting partial residuals of the noise, which are outlined by the first two shadows in the figure. The third part (8-11 layers) aims to reduce the dimension of the feature map. The last part is a standard fully connected layer and a Softmax linear classifier. The 12 layers calculate the statistical moment of each feature map and input the feature vector into the classifier part. All convolutional layers use 3×3 convolution kernels, and all nonlinear activation functions are ReLU.

There are two types of detection errors: one is to identify the steganographic image as the original image, and the other is to identify the original image as the steganographic image. The images given as errors at the beginning are the images that the trained SRNet has identified as errors, in which both the steganographic image and the original image exist.

In order to verify the effectiveness of the method proposed in the present invention for adaptive steganography algorithm detection, steganalysis is performed on the data sets embedded with adaptive steganography algorithms in spatial domain and JPEG domain, as shown in Table 2 and Table 3. It can be seen from the experimental results that the present invention is applicable to both spatial domain and JPEG domain, and performs well overall.

Table 2 Detection accuracy of different steganographic algorithms in the spatial domain %

Table 3 Detection accuracy of different steganography algorithms in JPEG domain %

Note: The SRNet row in the table represents the result of the first processing of the discriminator; the Proposed row represents the result of the second processing of the discriminator after the method of the present invention; the quality factor QF is not distinguished in the spatial domain, but only in the JPEG domain.

FIG4 is an experimental comparison diagram of different training strategies using the present invention. When the replacement strategy is to replace all images, the training effect of the model is very poor, the detection accuracy is very low, and it is difficult to converge. The main reason is that the model loses a lot of features of the image itself during learning, resulting in a decrease in detection performance. Therefore, the second training strategy is changed to replace only the images that are incorrectly detected by the discriminator with the saliency map. The detection accuracy of this training strategy is significantly improved compared with the first training strategy, but it is lower than the SRNet baseline without replacement. Finally, the third training strategy is adopted, that is, only the images that meet the requirements after being processed by the regional screening module are replaced with the saliency map. It is found in the experiment that this training strategy has the best effect and converges faster.

The discriminator is like a person. It cannot discriminate without learning anything. It needs to learn for a period of time (running on a computer) to learn to discriminate. This learning process is called "training". Therefore, the final effect of the present invention is the discrimination accuracy of the discriminator. In essence, the method is to improve the effect of "training".

Figure 5 is a steganalysis system simulated by the present invention and the detection result diagram, where cover represents the original image; stego represents the stego image, and the values after cover and stego represent the probability values of belonging to cover and stego. The larger the probability, the more likely the image belongs to. In this system, the trained model is called to detect any image on the network.

First, randomly download an image from the Internet. Since most of the images on the Internet are color images, the three-channel RGB image is first converted into a single-channel grayscale image, named cover.jpg; secondly, the J-UNIWARD adaptive steganography algorithm is used for steganography, and the embedding rate is 0.1-0.4bpp. The embedded images are named stego-0.1.jpg, stego-0.2.jpg, stego-0.3.jpg, and stego-0.4.jpg respectively. The four steganographic images cannot be distinguished by the naked eye. Finally, the four steganographic images are input into the steganalysis detection system for detection. The function of the system is to determine whether the input image is the original image or the steganographic image, and to display the probability value of each class and the detection results. It can be seen from the detection results that as the embedding rate increases, the probability value of the steganographic image increases, which further illustrates the accuracy of the detection results. The present invention as a whole has good generalization ability and practical value.

Claims (10)

1. An image adaptive steganalysis method based on saliency detection, characterized in that it comprises the following steps: 1) Segment the salient regions of the image in the image with the detected error to form a saliency map; 2) According to the overlap between the salient area and the steganographic area, the saliency map is screened, and the saliency map that meets the requirements is extracted, and the saliency map that meets the requirements is fused with the corresponding original image to form a saliency fusion map; wherein the saliency map that meets the requirements is an image with a high overlap between the salient area and the steganographic area; 3) Replace the saliency maps that do not meet the requirements with the original images, and combine these original images and the saliency fusion maps into an updated dataset; 4) Then use the updated dataset for training. 2. According to the image adaptive steganalysis method based on saliency detection according to claim 1, it is characterized in that in step 1), a discriminator module is used to detect erroneous images, and the discriminator module adopts an SRNet model. 3. According to the image adaptive steganalysis method based on saliency detection according to claim 2, it is characterized in that in step 2), the image fusion is specifically: setting the remaining pixels in the image except the pixels in the salient area to 0, so that the discriminator module only focuses on the image features of the salient area. 4. The image adaptive steganalysis method based on saliency detection according to claim 1, characterized in that, in step 1), a saliency detection module is used to segment the salient area of the image; The saliency detection module adopts the BASNet model, including the prediction module and the multi-scale residual optimization module, to form a saliency map as follows: The prediction module and multi-scale residual optimization module of the BASNet model are introduced into the network, and a rough saliency map is obtained through the prediction module; The multi-scale residual optimization module optimizes the rough saliency map of the prediction module by learning the residual between the rough saliency map and the true annotation, and finally obtains the refined saliency map. 5. The image adaptive steganalysis method based on saliency detection according to claim 1 is characterized in that, in step 2), a region screening module is used to screen the saliency map to extract the saliency map that meets the requirements. 6. The image adaptive steganalysis method based on saliency detection according to claim 1 is characterized in that, in step 2), the calculation method of the overlap degree η between the saliency area and the stego area is as follows:

From formula (1), we can get formula (2):

Among them, N represents the total number of pixels in the image, N _coin represents the number of pixels in the overlapping area, N _stego represents the number of pixels in the steganographic area, P _Stego (i, j) and P _SOD (i, j) represent the pixel values of the stegographic point map and the saliency map at position (i, j), respectively. 7. The image adaptive steganalysis method based on saliency detection according to claim 1 is characterized in that, in step 2), the overlap degree corresponding to the saliency map that meets the requirements is 0.6 to 1. 8. An image adaptive steganalysis system for implementing the image adaptive steganalysis method based on saliency detection according to any one of claims 1 to 7, characterized in that it comprises a saliency detection module, a region screening module and a discriminator module; The saliency detection module is used to generate the salient area of the image to be tested. It adopts the BASNet model, including the prediction module and the multi-scale residual optimization module. The prediction module is used to obtain a rough saliency map; the multi-scale residual optimization module is used to optimize the rough saliency map of the prediction module by learning the residual between the rough saliency map and the true annotation, and finally obtain the refined saliency map; The region screening module is used to screen the saliency maps and extract the saliency maps that meet the requirements; The discriminator module adopts the SRNet model, which is used to provide images with initial detection errors and retrain on the updated dataset. 9. The image adaptive steganalysis system based on saliency detection according to claim 8, characterized in that the multi-scale residual optimization module comprises an input layer, an encoder, a bridge layer, a decoder and an output layer. 10. The image adaptive steganalysis system based on saliency detection according to claim 9, characterized in that both the encoder and the decoder have four stages, each stage has only one convolutional layer, and each layer has 64 filters of size 33; The bridge layer also sets up a convolutional layer with the same parameters as the other convolutional layers.

Images