CN114390154A

CN114390154A - Robust steganography method and system for selecting embedded channel based on channel matching network

Info

Publication number: CN114390154A
Application number: CN202111590914.4A
Authority: CN
Inventors: 张祎; 罗向阳; 杨春芳; 刘粉林
Original assignee: Information Engineering University of PLA Strategic Support Force
Current assignee: Information Engineering University of PLA Strategic Support Force
Priority date: 2021-12-23
Filing date: 2021-12-23
Publication date: 2022-04-22
Anticipated expiration: 2041-12-23
Also published as: CN114390154B

Abstract

The invention belongs to the technical field of information hiding, and particularly relates to a robust steganography method and a robust steganography system for selecting an embedded channel based on a channel matching network.A lossy channel transmission test is carried out by utilizing an original and damaged carrier image set, and a repeated transmission network is constructed to depict information such as image processing attack types, parameters and the like in a channel; then, carrying out block matching on the lossy channel according to the robust features of the carrier image, constructing and training a channel matching network based on deep learning, and selecting feature image blocks robust to the channel as candidate embedding positions; and a robust information embedding domain is constructed according to a channel description result, the embedding cost of the carrier element is optimized based on a complex and significant region priority principle, and the secret information embedding is realized by combining an error correcting code and an STC code. The method can effectively utilize the network lossy channel, improve the robustness and invisibility of the embedded information, not only obviously improve the information extraction accuracy after various image processing attacks, but also have stronger detection resistance.

Description

Robust steganography method and system for selecting embedded channel based on channel matching network

Technical Field

The invention belongs to the technical field of information hiding, and particularly relates to a robust steganography method and a robust steganography system for selecting an embedded channel based on a channel matching network.

Background

With the rapid development of smart phones, imaging technologies and social media, mass images acquired, stored and transmitted in a network environment become an important covert communication carrier. However, unlike image steganography in a conventional lossless channel, when a secret-carrying image is transmitted through a social platform, the secret-carrying image is often subjected to image processing operations such as compression, scaling, filtering and the like, so that image degradation and information loss are caused, and therefore, the conventional steganography technology based on least significant bit replacement is difficult to successfully implement covert communication. Therefore, it is necessary to construct a steganographic carrier that can resist various image processing attacks, and design an image steganographic method that has robustness and anti-detection performance and is suitable for a network lossy channel.

The robust watermarking technology is used as a classic robust information hiding technology, and the purpose of inquiring an image source by utilizing the existence detection of watermark information is achieved by embedding the watermark information which can resist various unintentional or existing signal processing attacks into a target image, for example, the restorability of the watermark information to various image processing attacks such as noise adding, compression, scaling, filtering, rotation, stretching and the like is obtained by a classic robust watermarking algorithm designed based on coefficient relation, wavelet transformation, scale invariant characteristics and the like, and important technical support is provided for digital image copyright protection. In contrast, in the image steganography technology, especially the adaptive steganography technology which is widely concerned by researchers at present, by accurately depicting the embedding distortion of carrier elements at different positions and combining with the minimum embedding distortion coding, the texture complex region of an image can be selected to embed secret information in a self-adaptive manner, a high-dimensional statistical model of a carrier image is well kept, and the strong anti-detection performance of an image steganography detection algorithm based on statistical characteristics and deep learning is obtained. However, the concealment and the extraction integrity of the embedded information cannot be guaranteed by the robust watermarking algorithm, and the loss of the transmission channel to the secret-loaded image and the secret message is usually ignored by the steganography algorithm, so that the requirements of the network on the safe and reliable information transmission of the lossy channel covert communication are difficult to meet.

Disclosure of Invention

Therefore, the invention provides a robust steganography method and a robust steganography system for selecting an embedded channel based on a channel matching network, which effectively utilize a network lossy channel, improve the robustness and invisibility of embedded information and ensure the accuracy of extraction of secret-carrying image information.

According to the design scheme provided by the invention, the robust steganography method for selecting the embedded channel based on the channel matching network comprises the following contents:

the method comprises the steps of uploading an original carrier image and a test carrier image as uploading objects of transmission test to a target lossy channel for transmission test and obtaining a corresponding receiving carrier image, comparing and updating the original carrier image and the test carrier image with the receiving carrier image and obtaining a target lossy channel image processing attack type and parameters, wherein the test carrier image is obtained by using a preset repeated transmission network as a lossy channel for transmission test on the original carrier image;

carrying out image processing of simulation attack on an original carrier image by using the acquired target lossy channel image processing attack type and parameters to acquire a damaged carrier image, and acquiring a block region for removing carrier elements in an image smooth region by using block mean square error and mean square residual of the carrier image and a damaged image saliency map; simultaneously, extracting Shi-Tomasi characteristic key points from the carrier image and the damaged image through scaling and blocking processing, selecting image blocks where the characteristic key points are located according to the intensity, selecting the image blocks according to the complexity of the image blocks to construct label blocks for marking whether the label blocks are embedded into the information carrier image blocks, and training a channel matching network by using the constructed label blocks;

constructing an initial robust carrier sequence according to the type of a target lossy channel image processing attack type, acquiring an initial embedded information modification amplitude sequence corresponding to sequence elements by combining a trained channel matching network, and selecting carrier elements of image blocks where feature key points are located and which do not belong to the acquired block areas according to the strength to obtain elements in the robust carrier sequence and embedded information modification amplitude sequences corresponding to the elements;

contrast enhancement is carried out on the original image, and carrier element embedding cost is obtained by combining a robust carrier sequence and an embedded information modification amplitude sequence; scrambling and error correction coding are carried out on the secret message sequence to be embedded, space-time coding is carried out on the secret message sequence after scrambling and error correction coding by using robust carrier sequence elements and embedding cost sequence elements to obtain a secret-carrying element sequence, the secret-carrying element sequence is embedded into a robust steganographic carrier according to an embedding information modification amplitude sequence, and a corresponding secret-carrying image is generated.

The robust steganography method based on the channel matching network selection embedding channel is characterized in that a given carrier image is preprocessed in an original carrier image, the complexity of each image in the given carrier image is calculated, the given carrier image is subjected to image descending arrangement according to the complexity, and the carrier image used for robust steganography is selected as the original carrier image according to a complexity threshold value.

As the robust steganography method for selecting the embedded channel based on the channel matching network, further, each given image is divided into image blocks with fixed size after discrete cosine transform, and the complexity of each given image is obtained through the difference between the maximum value and the minimum value of the coefficient in the DCT blocks with the size of each image block.

The robust steganography method is used for selecting an embedded channel based on a channel matching network, further, an image processing attack set is preset in a repeated transmission network, attack parameters are preset in each attack type in the set, and the parameters at least comprise various attack intensity parameters and attack coefficient parameters; and carrying out image preprocessing of transmission test on the original carrier image by using the attack type and parameters in the preset repeated transmission network to obtain a test carrier image.

As the robust steganography method for selecting the embedded channel based on the channel matching network, the first n with the maximum strength is selected for the Shi-Tomasi characteristic key points of the carrier image and the damaged image_stThe image block where the key point is located is taken as a candidateEmbedding the block, calculating the complexity of the candidate embedded block according to the information embedding complex region priority principle, and selecting the top n with the maximum complexity according to the secret information and the carrier length_tbAnd constructing a label block by marking the image block with the selected complexity and the rest unselected image blocks with different marks to determine whether the image blocks are embedded into the information carrier image blocks, wherein n is_st，n_tbRepresenting a natural number.

As the robust steganography method for selecting the embedded channel based on the channel matching network, the channel matching network further adopts a convolutional neural network AlexNet structure, and the constructed label block is used as a training sample to train the channel matching network so as to obtain the channel matching network with the input as the image block and the output as the label corresponding to the image block.

The robust steganography method for selecting an embedded channel based on a channel matching network is characterized by further constructing a robust carrier sequence and acquiring an embedded information modification amplitude sequence corresponding to a sequence element, if attack types comprise a plurality of image lossy processing operations, constructing a multiple robust embedded domain by using a coefficient difference value of adjacent blocks and acquiring an initial robust carrier sequence of a carrier image, and calculating a corresponding initial embedded information modification amplitude sequence, wherein the plurality of image lossy processing operations at least comprise image scaling, image noise adding and image rotation; and if the attack type only comprises image processing operation, constructing a robust embedding domain based on compression parameters by using a quantization rounding principle in a JPEG (joint photographic experts group) compression process, acquiring an initial robust carrier sequence of the carrier image, and calculating a corresponding initial embedding information modification amplitude sequence, wherein the image processing operation comprises image compression and/or image scaling.

As a robust steganography method for selecting an embedded channel based on a channel matching network, further, carrier elements

Embedding cost

Is calculated byThe formula is shown as:

wherein the content of the first and second substances,

is a carrier element sequence c^rCorresponding embedded information modified amplitude sequence m^r·The ith element, representing a mark as an image block of the embedded information carrier,

are respectively X, Y_iCorresponding spatial domain image, X representing the block of DCT coefficients of the carrier image, Y_iFor corresponding on-carrier elements

The block of DCT coefficients after the information has been embedded,

as a carrier element

The sum of the mean square error residuals of the located significant image blocks, m and n are the image block positions of the significant image blocks,

for image g corresponding to the uv th wavelet coefficient in the first level decomposition of the k subbands, u e {1,2, …, n₁}， v∈{1,2,…,n₂τ is a constant used for stable numerical calculations.

The robust steganography method for selecting an embedded channel based on a channel matching network further comprises the following steps: the extraction of the secret-carrying image robust steganography for obtaining the secret information specifically comprises the following steps: carrying out image processing on the secret-carrying image according to the image processing attack type and parameters to obtain a processed image, selecting a blocking area with a mean square residual of 0 according to a saliency map, a mean square error and a mean square residual of the secret-carrying image and the processed image, carrying out scaling blocking processing on the secret-carrying image and the processed image to extract Shi-Tomasi characteristic key points, selecting image blocks where a plurality of characteristic key points are located according to the intensity, and testing the selected image blocks by using a channel matching network to obtain actual embedded information image blocks; constructing and extracting a robust secret-carrying sequence according to the attack type in the image processing attack type, and extracting a secret-carrying element which is positioned in an actual embedded information image block and does not belong to a block area to obtain a secret-carrying element sequence; and performing space-time code decoding on the secret-carrying element sequence by using the message length to obtain a decoded sequence, and correcting and inverting the decoded sequence to obtain secret information steganographically from the secret-carrying image.

Further, the present invention provides a robust steganography system for selecting an embedded channel based on a channel matching network, comprising: a transmission testing module, a channel matching module, a carrier acquiring module and a message coding module, wherein,

the transmission testing module is used for uploading an original carrier image and a test carrier image as uploading objects of transmission testing to a target lossy channel for transmission testing and acquiring a corresponding receiving carrier image, comparing and updating the original carrier image and the test carrier image with the receiving carrier image and acquiring the target lossy channel image processing attack type and parameters, wherein the test carrier image is acquired by using a preset repeated transmission network as a lossy channel for transmission testing of the original carrier image;

the channel matching module is used for carrying out image processing of simulation attack on the original carrier image by utilizing the acquired target lossy channel image processing attack type and parameters to acquire a damaged carrier image, and acquiring a block area for removing carrier elements in an image smooth area by utilizing the block mean square error and the mean square residual of the carrier image and a damaged image saliency map; simultaneously, extracting Shi-Tomasi characteristic key points from the carrier image and the damaged image through scaling and blocking processing, selecting image blocks where the characteristic key points are located according to the intensity, selecting the image blocks according to the complexity of the image blocks to construct label blocks for marking whether the label blocks are embedded into the information carrier image blocks, and training a channel matching network by using the constructed label blocks;

the carrier acquisition module is used for constructing an initial robust carrier sequence according to the type of a target lossy channel image processing attack type, acquiring an initial embedded information modification amplitude sequence corresponding to sequence elements by combining a trained channel matching network, and selecting carrier elements of image blocks where feature key points are located and which do not belong to the acquired block areas according to the strength to obtain elements in the robust carrier sequence and embedded information modification amplitude sequences corresponding to the elements;

the message coding module is used for enhancing the contrast of the original image and acquiring the embedding cost of the carrier elements by combining the robust carrier sequence and the embedded information modification amplitude sequence; scrambling and error correction coding are carried out on the secret message sequence to be embedded, space-time coding is carried out on the secret message sequence after scrambling and error correction coding by using robust carrier sequence elements and embedding cost sequence elements to obtain a secret-carrying element sequence, the secret-carrying element sequence is embedded into a robust steganographic carrier according to an embedding information modification amplitude sequence, and a corresponding secret-carrying image is generated.

The invention has the beneficial effects that:

the method utilizes an original and damaged carrier image set to carry out a lossy channel transmission test, and constructs a repeated transmission network to depict information such as image processing attack types, parameters and the like in a channel; carrying out block matching on the lossy channel according to the robust features of the carrier image, constructing and training a channel matching network based on deep learning, and selecting a feature image block robust to the channel as a candidate embedding position; and a robust information embedding domain is constructed according to a channel description result, embedding cost of carrier elements is optimized based on a complex and significant region priority principle, and secret information embedding is realized by combining an error correcting code and an STC code, so that a network lossy channel can be effectively utilized, robustness and invisibility of embedded information are improved, and practical scene application is facilitated. Further, experiments show that the scheme not only obviously improves the accuracy of information extraction after various image processing attacks, but also has stronger detection resistance and better application prospect.

Description of the drawings:

FIG. 1 is a flowchart illustrating a robust steganography method for selecting an embedded channel based on a channel matching network in an embodiment;

FIG. 2 is a schematic representation of a lossy channel characterization based on a duplicate transport network in an embodiment;

FIG. 3 is an illustration of carrier transport channel matching based on deep learning in an embodiment;

FIG. 4 is a schematic diagram of the embedding and extracting processes of the CNRAS algorithm in the scheme of the embodiment;

FIG. 5 is a schematic diagram of a deep learning-based network transmission channel matching result in the embodiment;

FIG. 6 is a schematic representation of a secret-carrying image for channel transmission test in an embodiment;

FIG. 7 is an exemplary spatial domain detection error rate;

FIG. 8 is a schematic diagram of the frequency domain detection error rate in the embodiment.

The specific implementation mode is as follows:

in order to make the objects, technical solutions and advantages of the present invention clearer and more obvious, the present invention is further described in detail below with reference to the accompanying drawings and technical solutions.

The wide application of new networks such as social networks and the like brings challenges to the traditional steganography technology, and under the lossy network environment, secret information embedded in the traditional steganography technology is easy to lose, so that hidden communication failure is caused. In the robust steganography algorithm appearing in recent years, information is mostly embedded into a robust carrier which keeps stable compression or scaling operation, multiple image processing operations are difficult to resist at the same time, and the anti-statistical detection performance needs to be improved. To this end, an embodiment of the present invention provides a robust steganography method for selecting an embedded channel based on a channel matching network, which is shown in fig. 1 and includes the following contents:

s101, using an original carrier image and a test carrier image as uploading objects of transmission test, uploading the uploading objects to a target lossy channel for transmission test and obtaining a corresponding receiving carrier image, comparing and updating the original carrier image and the test carrier image with the receiving carrier image and obtaining a target lossy channel image processing attack type and parameters, wherein the test carrier image is obtained by using a preset repeated transmission network as a lossy channel for transmission test on the original carrier image;

s102, carrying out image processing of simulation attack on an original carrier image by using the acquired target lossy channel image processing attack type and parameters to acquire a damaged carrier image, and acquiring a block area for removing carrier elements in an image smooth area by using block mean square error and mean square residual of the carrier image and a damaged image saliency map; simultaneously, extracting Shi-Tomasi characteristic key points from the carrier image and the damaged image through scaling and blocking processing, selecting image blocks where the characteristic key points are located according to the intensity, selecting the image blocks according to the complexity of the image blocks to construct label blocks for marking whether the label blocks are embedded into the information carrier image blocks, and training a channel matching network by using the constructed label blocks;

s103, constructing an initial robust carrier sequence according to the type of the target lossy channel image processing attack type, acquiring an initial embedded information modification amplitude sequence corresponding to sequence elements by combining a trained channel matching network, and selecting carrier elements of image blocks where feature key points are located and which do not belong to the acquired block areas according to the strength to obtain elements in the robust carrier sequence and embedded information modification amplitude sequences corresponding to the elements;

s104, contrast enhancement is carried out on the original image, and carrier element embedding cost is obtained by combining a robust carrier sequence and an embedded information modification amplitude sequence; scrambling and error correction coding are carried out on the secret message sequence to be embedded, space-time coding is carried out on the secret message sequence after scrambling and error correction coding by using robust carrier sequence elements and embedding cost sequence elements to obtain a secret-carrying element sequence, the secret-carrying element sequence is embedded into a robust steganographic carrier according to an embedding information modification amplitude sequence, and a corresponding secret-carrying image is generated.

Carrying out a lossy channel transmission test by using the original and damaged carrier image sets, and constructing a repeated transmission network to depict information such as image processing attack types, parameters and the like in a channel; and then carrying out block matching on the damaged channel according to the robust features of the carrier image, constructing and training a channel matching network based on deep learning, and selecting a feature image block which is robust to the channel and has complexity and significance as a candidate embedding position, thereby improving the robustness and invisibility of the embedded information at the same time.

As the robust steganography method for selecting the embedded channel based on the channel matching network in the embodiment of the invention, further, in an original carrier image, the given carrier image is preprocessed, the complexity of each image in the given carrier image is calculated, the given carrier image is subjected to image descending arrangement according to the complexity, and the carrier image used for robust steganography is selected as the original carrier image according to the complexity threshold value. Furthermore, each given image is divided into image blocks with fixed sizes after discrete cosine transform, and the complexity of each given image is obtained through the difference between the maximum value and the minimum value of the coefficient in the DCT blocks with the sizes of the image blocks. In a preset repeated transmission network, setting an image processing attack set, wherein each attack type in the set is preset with an attack parameter which at least comprises a plurality of attack intensity parameters and attack coefficient parameters; and carrying out image preprocessing of transmission test on the original carrier image by using the attack type and parameters in the preset repeated transmission network to obtain a test carrier image. Further, the channel matching network adopts a convolutional neural network AlexNet structure, and the constructed label block is used as a training sample to train the channel matching network so as to obtain the channel matching network with the input as the image block and the output as the label corresponding to the image block.

As the robust steganography method for selecting the embedded channel based on the channel matching network in the embodiment of the invention, further, the top n with the maximum intensity shared by the carrier image and the damaged image is selected according to the Shi-Tomasi characteristic key point of the carrier image and the damaged image_stThe image block where the key point is located is used as a candidate embedded block, the complexity of the candidate embedded block is calculated according to the information embedded complex region priority principle, and the first n with the maximum complexity is selected according to the secret information and the carrier length_tbAnd constructing a label block by marking the image block with the selected complexity and the rest unselected image blocks with different marks to determine whether the image blocks are embedded into the information carrier image blocks, wherein n is_st，n_tbRepresenting a natural number. Further, a robust carrier sequence is constructed and sequence elements are obtainedIn the corresponding embedded information modification amplitude sequence, if the attack type comprises a plurality of image lossy processing operations, constructing a multiple robust embedded domain by using the difference value of the coefficients of adjacent blocks, acquiring an initial robust carrier sequence of a carrier image, and calculating the corresponding initial embedded information modification amplitude sequence, wherein the plurality of image lossy processing operations at least comprise image scaling, image noise adding and image rotation; and if the attack type only comprises image processing operation, constructing a robust embedding domain based on compression parameters by using a quantization rounding principle in a JPEG (joint photographic experts group) compression process, acquiring an initial robust carrier sequence of the carrier image, and calculating a corresponding initial embedding information modification amplitude sequence, wherein the image processing operation comprises image compression and/or image scaling.

Further, the method also comprises the following steps: the extraction of the secret-carrying image robust steganography for obtaining the secret information specifically comprises the following steps: carrying out image processing on the secret-carrying image according to the image processing attack type and parameters to obtain a processed image, selecting a blocking area with a mean square residual of 0 according to a saliency map, a mean square error and a mean square residual of the secret-carrying image and the processed image, carrying out scaling blocking processing on the secret-carrying image and the processed image to extract Shi-Tomasi characteristic key points, selecting image blocks where a plurality of characteristic key points are located according to the intensity, and testing the selected image blocks by using a channel matching network to obtain actual embedded information image blocks; constructing and extracting a robust secret-carrying sequence according to the attack type in the image processing attack type, and extracting a secret-carrying element which is positioned in an actual embedded information image block and does not belong to a block area to obtain a secret-carrying element sequence; and performing space-time code decoding on the secret-carrying element sequence by using the message length to obtain a decoded sequence, and correcting and inverting the decoded sequence to obtain secret information steganographically from the secret-carrying image.

Referring to fig. 2, by performing a lossy channel transmission test on a candidate carrier image set and a test image set subjected to attack, a carrier image repeat transmission network can be constructed, and information such as an image processing attack type and parameters in a channel is characterized, so as to establish a basis for further matching a message embedding channel of covert communication by using side information of a lossy channel. The method specifically comprises the following steps:

(1) for a given network lossy channel, u for a given height_IWidth v of_IOf the carrier image set S_IFirst, a set of possible image processing attacks is defined

And presetting parameters such as various attack strengths, coefficients and the like for each attack

(2) According to the attack types and parameters, carrying image set S_IPreprocessing is carried out, and the attack parameters are

Image processing operation O of_iThe corresponding test image set obtained is recorded as

Further obtaining the image set

(3) Carrying out image transmission test by using a target network lossy channel, and collecting a carrier image set S_IAnd testing the image set

Respectively transmitted through the channel to obtain corresponding receiving image sets

And

(4) preliminarily determining the image processing attack types contained in the network lossy channel by comparing the characteristics of the image size, coefficient and the like of the received image set with the characteristics of the carrier and the test image set to obtain an image processing operation set

And to O¹Selecting attack parameters

(5) Referring to step (2), the carrier image set S is subjected to the above parameters_IProcessing to obtain corresponding test image set

Repeating the step (3) and carrying out image transmission test by utilizing a target network lossy channel to obtain a receiving image set

And

repeating the step (4) and updating the image processing attack type and the parameter set contained in the network lossy channel according to the comparison result

And

(6) repeating the step (5) until the test image is similar to the received image, and storing the image processing operation type and the parameter set

And

and uses it as a network lossy channel parameter.

According to the above lossy channel characterization based on repeated transmission networks, an example is given as follows: aiming at mainstream instant messaging software, namely WeChat and microblog, firstly preliminarily determining that the operations mainly comprise zooming, compression and the like in the image processing process through an image transmission experiment. Then, for the scaling operation, color images with different sizes, which are shot by brand mobile phones such as the headsei, SAMSUNG, IPHONE, OPPO, and VIVO, are scaled to different sizes, and the original image and the test image are subjected to transmission experiments. And then, comparing the original image, the test image after image processing attack and the received image to obtain that the scaling boundary values of the WeChat and the microblog are 1080, namely, the software does not scale the image with the size smaller than 1080. And then, selecting the original images with different sizes to be zoomed, then carrying out repeated transmission experiments by using the communication software, and comparing the sizes of the images of the received image set, the carrier and the test image set to obtain a test conclusion that the software zooms the images with different sizes to 1080 according to the long sides of the images, namely, preliminarily carrying out parameter estimation on the zooming operation in the image transmission process of the WeChat and the microblog. On this basis, for the compression operation therein, it may be first determined through experimental tests that the compression algorithm mainly employed in the above channel is JPEG compression. And then, respectively carrying out JPEG compression attack with quality factor of the zoomed images, and downloading the compressed test images and original images after WeChat and microblog transmission. Then, parameters such as compression ratio, quality factor, quantization table coefficient and the like selected by compression operation in different channels aiming at the uploaded images with different quality factors can be obtained by comparing the quality factors, the quantization tables and the DCT coefficients of the uploaded and downloaded images, namely, parameter estimation of the compression operation in the image transmission process of WeChat and microblog is realized, and further, the corresponding lossy channel parameter description is completed.

In summary, through the above process, a repeated transmission network for a carrier image is constructed by using the side information of the lossy channel, and the image processing attack type and parameters in the channel are obtained by comparing the uploaded image with the downloaded image, thereby realizing network lossy channel characterization.

Based on the lossy channel characterization of the repeated transmission network, a channel matching network based on deep learning is constructed and trained according to robust features of a carrier image by simulating image processing operation in a lossy channel for the carrier image, and the network lossy channel in the transmission process is matched, as shown in fig. 3, so that robust feature image blocks which are less affected by the lossy channel are determined as candidate embedding positions, and the robustness and invisibility of embedded information are further improved, specifically as follows:

(1) for a given height u_IWidth v of_IOf the carrier image set S_IIn section 2.1, on the basis of lossy channel characterization based on a repeating transmission network, a set of image processing operations is performed on the characterized image

And attack parameters

Carrying out channel simulation attack on the target object, and setting the passing parameter as

Image processing operation O of_iThe resulting test image is collected as

Obtaining a set of images

(2) In order to accelerate the training process and improve the algorithm efficiency, the carrier image set S is_IAnd impaired image set

Respectively reduced to a given size u'_I×v′_IAnd 8 multiplied by 8 blocking is carried out, on the basis, Shi-Tomasi key points in the image are extracted, and carrier image sets S are respectively selected_IAnd impaired image set

The first n with the highest common intensity_st8 x 8 image block where key point is located

As candidate embedded partitions.

(3) According to the complex region priority principle of information embedding, carrying image set S_IAnd impaired image set

Selected 8 x 8 image block T_bCalculating the complexity according to formula (1), and selecting the first n with larger complexity according to the secret message and the carrier length_tbAn image block

The mark is '1', the rest image blocks are marked as '0', and a label image block set is obtained

As image blocks for information embedding.

Wherein x is_d＝(i-1)×8+2k-1,y_d＝(j-1)×8+2l-1，

Is a position (x)_d,y_d) The coefficient value of (c).

(4) AlexNet [28 ] using the classical convolutional neural network architecture]Defining a channel matching network N_cAnd using a set of labeled image blocks

It is trained as a training sample such that for an input 8 x 8 patch, the network N_cOutputs its corresponding '0', '1' tag to indicate whether the partition is a carrier image block for embedding information.

(5) In practical application of network lossy channel covert communication, aiming at image set S_IAny one of the images in (1) is first designated as u'_I×v′_ICompressing and 8 x 8 blocking, then extracting Shi-Tomasi key points and selecting the top n with the maximum intensity_stThe image block where the key point is located forms a candidate embedded block T_bAnd then, sending the image block into a channel matching network for testing to obtain a carrier image block marked as '1' and actually used for embedding information, namely, completing the matching of the carrier/secret-carrying image to a specific network lossy channel.

In summary, a carrier image is subjected to simulated attack by using a lossy channel depiction result based on a repeated transmission network, and by using Shi-Tomasi feature points with better robustness for common image processing operation, on the basis of image compression, blocking and marking, in a transmission channel matching scheme based on deep learning, by constructing and training a channel matching network, the embedded position matching of a network lossy channel known or partially known for channel information is realized, so that the communication capacity of the existing robust steganography method for a specific channel is further optimized by using channel side information under the premise of comprehensively considering the robustness of embedded information and the detectability resistance of a secret-loaded image.

Further, based on the above method, an embodiment of the present invention further provides a robust steganography system for selecting an embedded channel based on a channel matching network, including: a transmission testing module, a channel matching module, a carrier acquiring module and a message coding module, wherein,

A Robust Steganography algorithm (CNRAS algorithm for short) for selecting an embedded Channel by using a Channel matching Network in the scheme is designed by using the side information of a Network lossy Channel described by a repeated transmission Network based on a carrier image, combining a Robust embedded domain and an embedded cost measurement algorithm and based on the existing 'error correction-STC' coding structure. In the scheme, the CNRAS algorithm optimizes and selects an information embedding channel by using a network transmission channel matching algorithm based on deep learning aiming at the image processing operation type and parameters contained in the channel on the basis of the characterization of a transmission channel, and combines error correction coding and minimum embedding distortion coding to realize minimum cost embedding of secret information, thereby further improving the robustness of the embedded information aiming at specific image processing operation in the channel and the anti-detection performance of a secret-carrying image. The embedding and extracting process of the CNRAS algorithm is shown in FIG. 4. The embedding process mainly comprises six parts, namely carrier image selection, transmission channel characterization, steganographic channel optimization and selection, robust carrier construction, embedding cost measurement, message encoding and embedding; the extraction process mainly comprises three parts of steganographic channel synchronization, robust secret-carrying extraction and message decoding. The main embedding process of the algorithm can be designed as follows:

(1) selecting a carrier image: in order to improve the robustness and the anti-detection performance of the secret-carrying image, the carrier image needs to be pre-selected so as to delete the carrier image which is too simple in content and texture and not suitable for information hiding. Specifically, for each given carrier image in the candidate image set, the complexity c of the image is first calculated as follows_I：

Wherein the image is DCT transformed and then divided into segments of size t_d×t_dThe small pieces of the rubber,

are respectively an image c_IAfter 8 x 8 blocking DCT, the number of rows and columns, e, of blocks of DCT coefficients_(i,j)The complexity of the block of DCT coefficients located at position (i, j) is defined as follows:

wherein x is_d＝(i-1)u_d+(2k-1)t_d,y_d＝(j-1)v_d+(2l-1)t_dAnd is and

is a position (x)_d,y_d) The coefficient value of (a), and thus the complexity of the candidate carrier image, may be determined by each t_d×t_dThe difference between the maximum and minimum of the coefficients in the DCT block is calculated (t is usually empirically taken)_d2). In addition, if the complexity of a certain image is less than the threshold value e_TThe DCT coefficient block of (2), i.e., the block of coefficients in the image that are too smooth to be steganographically appropriate, sets the complexity of the image to 0. The threshold value e is usually taken empirically_T20. And finally, arranging the images in the candidate carrier image set in a descending order according to the complexity, and sequentially selecting the robust steganographic carrier images according to the order.

(2) Transmission channel characterization: carrying out image processing operation on the candidate carrier image set subjected to pre-selection and sorting by utilizing a lossy channel depiction algorithm to obtain a test image set, carrying out a lossy channel transmission test, constructing a carrier image-based repeated transmission network, and obtaining the image processing operation type in the channel by repeatedly comparing the image size, the coefficient and other characteristics of the received image set and the carrier and the test image set

And parameters

(n_aFor the number of types of image processing operations, q_iThe number of parameters in the i-th image processing operation).

(3) Optimizing and selecting a steganographic channel: using image processing operation type O and parameter P_oFor the carrier image set S_IProcessing to obtain image set

And then, performing steganographic channel optimization and selection respectively according to the following modes:

to carrier image set S_IAnd damaged carrier image set

Global and local features of the image are designed by utilizing a Gaussian mixture model, and a saliency map S is respectively obtained by an image abstraction and saliency detection method based on global contrast_cAnd

and calculating 8 multiplied by 8 block mean square error to obtain block mean square error V_bAnd

will V_bAnd

respectively subtracting to obtain mean square residual sequence

And selecting block Z with mean square residual of 0_bAnd further removing the carrier elements located in the image smoothing area.

Secondly, a carrier image set S is collected by using a transmission channel matching algorithm based on a deep learning network_IAnd impaired image set

Respectively reduced to u'_I×v′_IThen 8 multiplied by 8 blocks are carried out to extract Shi-Tomasi key points, and the front n with the maximum intensity is selected_stImage block T where one key point is located_bAnd calculating the complexity, and then selecting the front n with larger complexity_tbOne image block as information embedded partition

And marked as '1', the rest of the blocks are marked as '0', and the label block is obtained

By using

Training a predefined channel matching network N_cThe method is used for information embedding block synchronization after the transmission of the lossy channel.

(4) And (3) robust carrier construction: if the image processing operation type set O in the carved lossy channel comprises a plurality of image lossy processing operations such as image compression, scaling, noise addition, rotation and the like, a multiple robust embedding domain is constructed by using the coefficient difference of adjacent blocks, and a carrier image I is subjected to_cConstruction and extraction of robust carrier element sequences

And calculating a corresponding embedded information modified amplitude sequence

If the image processing operation type set O only comprises processing operations such as image compression, scaling and the like, such as WeChat, microblog and other social software frequently used in an open social network, a robust embedded domain is constructed on the basis of compression parameters such as a quantization table, a quantization step size and the like by utilizing a quantization rounding principle in a JPEG (joint photographic experts group) compression process, and a carrier image I is subjected to_cExtracting robust vector sequences

(. represents '0' or '1', n)_cAs number of carrier elements). On the basis, extracting image blocks with larger intensity at key points

And does not belong to the image block Z with mean square residual error of the saliency map being 0_bThe carrier element in (1) to obtain a carrier element sequence

And its corresponding embedded information modified amplitude sequence

(n_rTo choose the number of robust carrier elements).

(5) The embedding cost measurement: by using the difference of visual quality after embedding information in smooth and complex regions of an image and reducing the embedding cost of carrier elements in the complex regions, the complex and significant regions of the image are preferentially selected as candidate regions for embedding the information, and the carrier image I is subjected to_cContrast enhancement is performed to obtain an enhanced image

And using the mean square error residual sequence of the saliency map

And selecting the carrier image block which is sensitive to image processing operation and has large significance change as an area with a complex structure and rough texture, and preferentially embedding information by reducing the embedding cost. Specifically, considering that modification of carrier elements in the information embedding process mostly occurs in DCT coefficients of the carrier image, carrier element embedding cost can be calculated according to the following formula by using an embedding distortion function in the adaptive steganography algorithm, and then a carrier element sequence c is obtained^rEmbedded information modified amplitude sequence m^·Corresponding embedding cost sequence

Wherein X denotes a block of DCT coefficients of the carrier image, Y_iAre elements on the carrier corresponding thereto

The block of DCT coefficients after the information has been embedded,

respectively, are the corresponding spatial domain images,

is the sum of the mean square error residual sequences of the saliency maps, and

as a carrier element

The significant figure is the sum of the mean square error residuals for the 8 x 8 partitions,

for the uv th wavelet coefficient in the k subband first layer decomposition corresponding to image g (k ═ 1,2,3), u ∈ {1,2, …, n₁}，v∈{1,2,…,n₂τ > 0 is a constant used for stable numerical calculations.

(6) Message encoding and embedding: secret message sequence m to be embedded^sScrambling and error correction coding to obtain a sequence m^eThereby increasing the probability that the embedded information is correctly extracted after being attacked. On the basis of this, use is made of the sequence c of the carrier elements^rEmbedded information modified amplitude sequence m^r·And a corresponding embedded cost sequence d^rTo m^eSTC coding with minimum embedding cost is carried out to obtain corresponding secret-carrying element sequence

Then, the dense element sequence s is carried according to different robust carrier constructions and extraction modes^rEmbedding into corresponding robust steganographic carrier to generate corresponding secret-carrying image I_s。

According to the process, by utilizing the lossy channel description result and through network transmission channel matching based on deep learning, the CNRAS algorithm in the scheme further optimizes the embedded channel in the robust embedded domain, and realizes the secret message embedding with stronger robustness on a specific lossy channel by combining the embedded cost measurement and the coding method, and the specific process can be shown as algorithm 1.

Accordingly, the extraction process of the algorithm can be designed to mainly include the following parts:

(1) steganographic channel synchronization: using image processing operation type O and parameter Po to encrypt image

Processing to obtain image set

And then respectively selecting secret carrying channels according to the following modes:

firstly, a saliency map is obtained by utilizing an image abstraction and saliency detection algorithm in a CPRAS algorithm

And

and respectively calculating 8 multiplied by 8 block mean square deviations thereof to obtain the block mean square deviations

And

will be provided with

And

respectively subtracting to obtain mean square residual sequence

And selecting the block with mean square residual error of 0

For removing the dense elements located in the smooth regions of the image.

Secondly, the secret-carrying image is matched by a transmission channel matching algorithm based on a deep learning network

Is reduced to u'_I×v′_IThen 8 multiplied by 8 blocks are carried out to extract Shi-Tomasi key points, and the front n with the maximum intensity is selected_stImage block T 'where key point is located'_bThrough a channel matching network N_cThe network output obtained by the test is '1', namely the actual embedded information image block

(2) Robust secret carrier extraction: if the image processing operation type set O in the lossy channel comprises a plurality of image lossy processing operations such as image compression, scaling, noise addition, rotation and the like, a multiple robust embedding domain construction algorithm based on coefficient difference value and provided in MREAS algorithm is utilized to carry out dense image processing

Construction and extraction of robust carrier density element sequences

If the image processing operation type set O only comprises processing operations such as image compression, scaling and the like, such as WeChat, microblog and other social software frequently used in an open social network, a robust embedded domain construction algorithm based on a quantitative rounding principle provided in the CPRAS algorithm is utilized to carry out image processing on the image

Extracting robust vector sequences

(. represents '0' or '1', n)_cNumber of secret elements). Then, extracting image blocks with larger intensity at key points

And does not belong to the image block with mean square residual error of 0 in the saliency map

The secret element in the information-embedded secret element sequence is obtained

(n_rNumber of secret-carrying elements synchronized for steganographic channel).

(3) And (3) message decoding: using message length l_mFor the extracted secret-carrying sequence

STC decoding is carried out to obtain a sequence

And performs error correction decoding and inverse scrambling to extract corresponding secret information

The specific extraction process can be designed as shown in algorithm 2:

in summary, in consideration of the situation that channel information is known or partially known in the image transmission process, by using lossy channel characterization based on a repetitive transmission network and network transmission channel matching based on deep learning, the robust steganography embedded channel selection method based on a channel matching network and a corresponding robust steganography algorithm (CNRAS) in the scheme further improve the robustness of the embedded information for specific image processing operation in the channel and the anti-detection performance of the secret-carrying image.

To verify the validity of the present application, the following further explanation is made by an algorithm and experimental data:

in order to analyze the effectiveness of the CNRAS algorithm in the scheme, the feasibility of network transmission channel matching based on deep learning is analyzed by using the universality theorem of the neural network, then the effectiveness of the algorithm is analyzed according to the complex region priority principle in the adaptive steganography embedding by analyzing the distribution of the embedding positions selected after the carrier image is subjected to channel matching, and finally the training time of the network under different numbers of training samples is analyzed, so that the efficiency of the robust steganography embedding channel selection algorithm based on the channel matching network provided by the text is analyzed.

The neural network universality theorem can be expressed as follows: if a feedforward neural network has a linear layer and at least one layer of activation functions with "squeeze" properties (e.g., activation functions such as signmoid, ReLU, etc.), it can approximate any Borel measurable function from one finite dimensional space to another finite dimensional space with any precision given a sufficient number of hidden units, i.e., for any function f (x) (including multivariate functions and vector valued functions), the network can output f (x) or a sufficient approximation of it for all possible inputs x to the function.

According to the theorem, by utilizing the lossy channel characterization result based on the repeated transmission network, when the label image blocks used for training the deep learning network are enough, the obtained channel matching network can extract the image blocks with larger key point intensity for information embedding and extraction aiming at the image processing attack contained in the channel, namely, the selection and synchronization of the embedding channel of robust steganography are realized, so that the robustness and invisibility of the embedded information are further improved.

On the basis, according to the previous lossy channel characterization result, image processing operations and parameters for training and testing the channel matching network are selected as shown in table 1 (wherein the processing such as compression, reduction, filtering and the like of common images also includes the marking processing frequently adopted by social platforms such as microblogs and the like, and the marking position is taken as an image bottom 1/15). Random selection of BossbasTaking u 'from 100 images in e1.01 image library'_I＝v′_I＝128,n_tbTraining a channel matching network N according to a deep learning-based network transmission channel matching algorithm in the scheme as 5_cTaking the image with the number of 142 as an example, the original image I_cAnd the image block label block T where the key point is located_bActual embedding of information image blocks after channel synchronization

As shown in fig. 5 (a), (b), and (c), respectively.

Table 1 image processing operations and parameters included in a lossy channel

From the above experimental results, it can be seen that by using the Shi-Tomasi key point with better robustness to image processing operation, and by using the lossy channel characterization result based on the repeated transmission network, the transmission channel can be trained and matched through the network on the basis of image blocking, so as to obtain an image block which is not affected by the image processing operation in the channel and is located in a complex area of carrier image texture, and the image block is used for embedding the secret message, and the embedded channel synchronization is performed at the receiving end, thereby improving the extraction accuracy of the embedded information after being transmitted through the specific network lossy channel, and optimizing the anti-detection performance of the secret-carrying image.

In addition, in order to evaluate the algorithm efficiency, 100, 500, 1000, 2000 carrier images randomly selected by the bossbase1.01 image library are used to respectively train the channel matching networks according to the processing operation types and parameters of the lossy channel images given in table 1, and after different iteration times are obtained, the accuracy and loss of the network classification result are shown in table 2.

TABLE 2 training results of channel matching networks corresponding to different numbers of carrier images

According to the experimental results, when different numbers of carrier images are used for training, the channel matching network in the scheme can realize the classification of key point image blocks with small loss and high accuracy at a high speed, so that the selection and the matching of the embedded channel of the lossy channel containing the image processing attack of specific types and parameters are completed. Correspondingly, the robust steganography algorithm (CNRAS) which selects the embedded channel by using the channel matching network in the scheme can fully utilize the known information in the channel, select the carrier image block which has better robustness and does not cause obvious change for the image processing operation contained in the channel with higher efficiency, realize the information embedding with lower cost, and further complete the information hiding transmission in the network lossy channel.

In order to verify the performance of the algorithm in the scheme, a classical self-adaptive steganography algorithm S/J-UNIWARD (Spatial/JPEG UNIversal WAvelet Relative translation), a compression-resistant robust steganography algorithm JCRISBE and QTRS, a scaling-resistant robust steganography algorithm IISRS, a multiple robust steganography algorithm MREAS, a CPRAS and other representative steganography algorithms are selected for performance test and comparison with two CNRAS-M/C robust steganography algorithms in the scheme of constructing the multiple robust embedded domain by combining a coefficient difference and a quantization rounding principle. The effectiveness of the algorithm is tested in the aspects of channel synchronization and channel transmission; and the method is compared with the prior representative algorithm in the aspects of robustness and detectability resistance. The method comprises the following specific steps:

firstly, by utilizing a BOSSbase1.01 image library and a UCID image library, 2000 images are randomly selected from the BOSSbase1.01 image library and are subjected to JPEG compression with the quality factor of 65 to serve as a frequency domain candidate carrier image set J_cDecompressing it to generate the corresponding spatial domain candidate carrier image set S_{c_65}. Then, by using representative steganography algorithms such as S/J-UNIWARD, JCRISBE, QTRS, IISRS, MREAS, CPRAS and the like and CNRAS-M/C robust steganography algorithm in the scheme, by using RS error correction coding with parameters of (31,19), embedding randomly generated secret information under embedding ratios of 0.001, K0.01, 0.02, K0.1 bpp/bpnzAC and the like, generating corresponding secret-carrying images (JCRISBE and QTRS algorithms assume that the pre-acquired JPEG compression quality factor is 65, and IISRS SRSThe algorithm assumes a bilinear interpolation scaling mode with a scale of 0.75 as a pre-acquired scaling parameter, the image processing attack types and parameters selected by the CNRAS-M/C algorithm in the scheme are shown in table 1, and the corresponding air/frequency domain algorithms are respectively marked as CNRAS-Ms/CNRAS-M_J、CNRAS-Cs/CNRAS-C_J). The specific experimental setup is shown in table 3.

Table 3 experimental parameter settings

And testing the synchronism of the channel matching network at the embedding and receiving ends. First, for a candidate carrier image set S_{c_65}/J_cIt is attacked using the image processing operations in table 1. Then, training a training channel matching network N for the original image set and the image set subjected to the attack through a network transmission channel matching algorithm based on deep learning_cAnd selecting the blocks with high key point strength and high texture complexity in the image as steganographic blocks. Candidate carrier images in different image sets are obtained, and the training results of the corresponding channel matching networks under different iteration times are shown in the following table 4.

TABLE 4 training results of channel matching networks corresponding to different sets of carrier images

According to the experimental results, candidate images in the carrier image set are selected, and carrier images of the label blocks which cannot be extracted with 100% accuracy after 1000 iterations are eliminated. On the basis, the processing operations such as image compression, scaling and the like in the table 1 are carried out on the secret-loaded image generated by the CNRAS-M/C algorithm in the scheme. And then, for the encrypted image subjected to the attack, acquiring an image block carrying secret information as a steganography channel synchronization result by utilizing steganography channel optimization based on the saliency map and lossy channel selection based on a channel matching network. And then comparing the steganographic image block of the embedded end with the synchronous image block of the receiving end, if the steganographic image block of the embedded end and the synchronous image block of the receiving end are completely consistent, considering that the steganographic channel is successfully synchronized, and otherwise, considering that the steganographic channel is failed to be synchronized. Finally, the synchronous success rate of steganographic channels of the secret-carrying images under different embedding rates of different steganographic algorithms is shown in table 5.

TABLE 5 synchronous accuracy of steganographic channel after image processing attack

The above experimental results show that the embedded channel selection algorithm based on the channel matching network in the scheme can fully utilize the captured lossy channel parameters, and the steganographic channel selection and matching which can resist the specific image processing attack can be realized with lower training cost and higher synchronization accuracy. Specifically, for image processing operations such as JPEG compression with quality factors of 65, 75 and 86, image scaling with parameters of 50% and 75%, and the like, after complex and significant block synchronization operation, the algorithm in the scheme can identify information embedding blocks with 100% accuracy by using a pre-designed and trained channel matching network, thereby providing guarantee for reliable extraction of embedded information. For the 3 × 3 median filtering attack, due to the influence of the parts of images with fuzzy background and foreground regions, information embedding at certain positions can cause the algorithm to be complicated in synchronization and generate certain deviation when the blocks are obviously partitioned, and further, the channel synchronization result of the CNRAS algorithm is influenced. For processing operations frequently adopted by a social platform such as a microblog and the like, the CNRAS algorithm in the scheme deletes candidate carrier images which can not correctly classify the label image blocks in the early-stage image training process, so that 100% of accurate channel synchronization is realized. Therefore, when the lossy channel of the target network comprises the image processing operation of filtering, the carrier image should select an image comprising a salient target, and in the process of selecting the embedded channel, channel optimization without a complex and significant area can be considered, and the steganographic channel is selected only based on the channel matching network, so that the extraction accuracy of the embedded information subjected to median filtering is improved.

Aiming at two robust steganography algorithms of CNRAS-M/C in the scheme, the visual quality of the generated secret-carrying image and the information extraction accuracy of a receiving end are tested by using a lossy channel provided by a WeChat friend circle, so that the invisibility of the secret-carrying image generated by the algorithm and the robustness of embedded information are reflected. Considering the practical application scenario, in the experiment, firstly, 20 color JPEG images with the quality factor of 65 in the UCID image set are selected as carrier images, and two robust steganography algorithms of CNRAS-M/C with the embedding ratio of 0.01bpnzAC are selected, and the generated carrier images are respectively used as test images, as shown in fig. 6. Then, the secret-carrying images are respectively uploaded to a WeChat friend circle and downloaded, PSNR values between the original secret-carrying images and the received secret-carrying images are calculated, secret information embedded in the received images is extracted, the average accuracy of information extraction is calculated, and specific experimental results are shown in the following table 6.

Table 6 actual network lossy channel transmission test

The above experimental results show that, in the scheme, based on fully utilizing the information of the lossy channel, the CNRAS-M/C algorithm selects the block embedded information with high intensity and high complexity of texture of the feature points in the image, thereby realizing the information hiding transmission in the lossy channel, i.e. the WeChat friend circle, and maintaining the high visual quality of the secret-loaded image, thereby providing guarantee for the invisibility of the embedded information. Similarly, the hidden communication experiment test performed by using the other images in the image set also achieves a better effect, and further shows the feasibility of the robust steganography algorithm for selecting the embedded channel by using the channel matching network in the scheme in practical application.

In order to compare the CNRAS algorithm with the space/frequency domain self-adaptive steganography algorithm S/J-UNIWARD, the compression-resistant robust steganography algorithms JCRISBE and QTRS, the scaling-resistant robust steganography algorithm IISRS, the multiple robust steganography algorithms MREAS, CPRAS and other representative steganography algorithms attacking different imagesRobustness, after removing the above-mentioned image with failed channel synchronization, the secret-carrying images generated by different algorithms under different embedding ratios are attacked by using the image processing operation in table 1. Subsequently, for each steganographic algorithm, the secret message in the corresponding secret-carrying image under each embedding ratio is extracted, and the average extraction error rate R is calculated_eAs shown in table 7 and table 8 (wherein italics indicates the best result, and bold indicates that the algorithm result in the scheme is better than the robust steganography algorithm result selected without embedding channels in the past).

Table 7 average error rate of information extraction after image processing attack (spatial domain image set S)_{c_65})(×10^-3)

TABLE 8 information extraction average error Rate after image processing attacks (frequency domain image set J)_c)(×10^-3)

The experimental results in table 7 show that, by using the transmission channel matching based on the deep learning network, the CNRAS robust steganography algorithm in the scheme has stronger robustness for processing operations such as JPEG compression, image scaling, median filtering, image labeling and the like. The method mainly comprises the steps that image blocks with high characteristic point strength and high complexity are mainly selected as information embedding positions in the algorithm, so that the expansion of information loss after image processing attack caused by the fact that the prior algorithm only considers the priority of image complex areas is avoided, and the robustness and invisibility of embedded information are considered. For the spatial domain carrier image, after the dense carrier image generated by the algorithm in the scheme is attacked by JPEG (joint photographic experts group) compression, the information extraction error rate is kept below 0.35 per thousand; after suffering from zoom attack, the information extraction error rate is respectively improved by 6.9-26.7 per mill and 0.58-2.80 per mill compared with the prior MREAS and CPRAS robust steganography algorithms, and the limitation that the IISRS algorithm is only suitable for the zoom attack with the zoom factor less than 0.5 is overcome; after the medium filtering attack is suffered, the information error rate is obviously improved compared with the prior MREAS and CPRAS algorithms. Particularly, when the secret-carrying image is attacked by the image marking, the algorithm in the scheme better acquires and utilizes the image processing attack parameters in the channel, thereby avoiding the information embedding at the marking position and ensuring the complete and correct extraction of the embedded information at the receiving end.

Similarly, aiming at the frequency domain carrier image, after the carrier image generated by the algorithm in the scheme is subjected to JPEG compression attack of different quality factors, the information extraction error rate is kept below 0.138%, and the defect that the robust steganography algorithms such as JCRISBE, QTRS and the like can only resist the JPEG compression attack of specific and single parameters is overcome; after suffering from scaling attack, the information extraction error rate is obviously improved compared with the prior MREAS and CPRAS algorithms, and the complete correct extraction of secret information is realized under a plurality of embedding rates; after suffering from median filtering attack, the information error rate is respectively improved by more than 30.1 per mill and 19.6 per mill compared with the prior MREAS and CPRAS algorithms; after suffering from image marking attack, 100% complete and correct extraction of embedded information is realized, so that the effectiveness of using channel matching to select an embedded channel in the scheme for improving the robustness of the embedded information is reflected, and the applicability of the CNRAS algorithm in the scheme in a damaged network channel is reflected.

In order to test the anti-detection performance of the algorithm generated secret-carrying image in the scheme aiming at the steganography detection algorithm based on statistical characteristics, spatial domain steganography detection characteristics such as SPAM and SRM and frequency domain steganography detection characteristics such as CCPEV and DCTR are utilized, multiple robust steganography algorithms such as S/J-UNIWARD self-adaptive steganography algorithm, MREAS and CPRAS and the like of an integrated classifier are combined, and CNRAS-M/C calculation provided in the scheme is usedThe secret-carrying images generated by the method under different embedding ratios are tested. Aiming at each group of secret-carrying images generated by different algorithms under different embedding ratios, 1/2 of the secret-carrying images are randomly used for training in the experiment, and the rest 1/2 of the secret-carrying images are used for testing, so that the detection error rate E of the secret-carrying images corresponding to different methods is obtained_OOBAs shown in fig. 7 and 8.

According to the experimental results, compared with the robust steganography algorithms such as MREAS and CPRAS, the secret-carrying image generated by the CNRAS robust steganography algorithm in the scheme can obtain a higher steganography detection error rate by using the robust steganography embedded channel selection based on the channel matching network. For the spatial domain secret image, when the embedding ratio is 0.001bpp, the CNRAS-M algorithm respectively obtains 49.17% and 43.35% detection error rates aiming at the steganography detection characteristics of SPAM and SRM; the CNRAS-C algorithm respectively obtains 49.53% and 48.82% detection error rates aiming at the steganography detection characteristics of SPAM and SRM, and is respectively improved by 4.34%/5.82% and 0.42%/3.47% compared with the previously proposed robust steganography algorithms such as MREAS, CPRAS and the like. With the increase of the embedding ratio, the algorithm of the scheme shows a trend of increasing before decreasing for the improvement of the anti-detection performance of the secret-carrying image. The method is mainly characterized in that the embedded information is further concentrated in the image blocks with high intensity and high complexity of the characteristic points through a channel matching network on the basis of selecting obvious and complex blocks of the image, so that the anti-detection performance of the dense image is obviously improved under the condition of low embedding rate. When the embedding ratio is increased, the algorithm has fewer selected embedded blocks and more concentrated information embedding, thereby causing the reduction of the anti-detection performance of the secret-carrying image.

For the frequency domain secret-carrying image, when the embedding ratio is 0.001bpnzAC, the CNRAS-M algorithm respectively obtains 49.38% and 49.27% of detection error rates aiming at CCPEV and DCTR steganography detection characteristics; the CNRAS-C algorithm respectively obtains 49.50% and 48.07% detection error rates aiming at CCPEV and DCTR steganography detection characteristics, and the detection error rates are respectively improved by 2.26%/3.52% and 0.47%/1.79% compared with the steganography algorithms before improvement, such as MREAS, CPRAS and the like. Along with the improvement of the embedding ratio, the detection resistance performance of the embedded image is similar to the detection resistance performance change condition of the spatial secret-carrying image and is influenced by gradual concentration of the embedded information, the improvement of the detection resistance performance of the secret-carrying image generated by the algorithm also shows the trend of ascending before descending, but still keeps the detection error rate equivalent to that of the CPRAS algorithm, thereby reflecting the stronger detection resistance performance of the CNRAS algorithm in the scheme.

Based on the experimental results, the scheme can fully mine and utilize channel information, and efficiently select robust and complex robust carrier embedded information, so that the communication capability of a robust steganography algorithm on a specific network lossy channel is optimized. On the basis, the effectiveness of the algorithm is tested from the aspects of channel synchronization and channel transmission, and a comparison result with the existing robust steganography algorithm is given from the aspects of robustness and detectability resistance. Experiments show that after secret information embedded in the CNRAS algorithm in the scheme is attacked by multiple image processing such as compression, scaling, filtering and marking, the information extraction accuracy of the space/frequency domain secret-carrying images is respectively improved to 99.96%/99.86%, 90.43%/86.3%, 89.8%/91.6% and 100%/100%, and the defects that the existing IISRS, JCRISBE and QTRS robust steganography algorithm is narrow in application range and only can resist single and specific parameter image processing attacks are overcome; the information extraction accuracy of the existing MREAS, CPRAS and other multiple robust steganography algorithms is improved, and the method has stronger anti-detection performance and better application prospect.

Unless specifically stated otherwise, the relative steps, numerical expressions, and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.

Based on the foregoing method and/or system, an embodiment of the present invention further provides a server, including: one or more processors; a storage device for storing one or more programs which, when executed by the one or more processors, cause the one or more processors to implement the method described above.

Based on the above method and/or system, the embodiment of the invention further provides a computer readable medium, on which a computer program is stored, wherein the program, when executed by a processor, implements the above method.

In all examples shown and described herein, any particular value should be construed as merely exemplary, and not as a limitation, and thus other examples of example embodiments may have different values.

It should be noted that: like reference numbers and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.

Finally, it should be noted that: the above-mentioned embodiments are only specific embodiments of the present invention, which are used for illustrating the technical solutions of the present invention and not for limiting the same, and the protection scope of the present invention is not limited thereto, although the present invention is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: any person skilled in the art can modify or easily conceive the technical solutions described in the foregoing embodiments or equivalent substitutes for some technical features within the technical scope of the present disclosure; such modifications, changes or substitutions do not depart from the spirit and scope of the embodiments of the present invention, and they should be construed as being included therein. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A robust steganography method for selecting an embedded channel based on a channel matching network is characterized by comprising the following contents:

2. The robust steganography method based on channel matching network selection embedding channel as claimed in claim 1, wherein in an original carrier image, a given carrier image is preprocessed, the complexity of each image in the given carrier image is calculated, the given carrier image is subjected to image descending order according to the complexity, and the carrier image used for robust steganography is selected as the original carrier image according to a complexity threshold.

3. The robust steganography method for selecting an embedded channel based on a channel matching network as claimed in claim 2, wherein each given image is divided into image blocks of fixed size after discrete cosine transform, and the complexity of each given image is obtained by the difference between the maximum value and the minimum value of the coefficients in the DCT blocks of each image block size.

4. The robust steganography method for selecting an embedded channel based on a channel matching network as claimed in claim 1, wherein an image processing attack set is set in a preset repetitive transmission network, and attack parameters are preset in each attack type in the set, wherein the parameters at least comprise a plurality of attack intensity parameters and attack coefficient parameters; and carrying out image preprocessing of transmission test on the original carrier image by using the attack type and parameters in the preset repeated transmission network to obtain a test carrier image.

5. The robust steganography method for selecting an embedded channel based on a channel matching network as claimed in claim 1, wherein the top n with the maximum intensity is selected for the key points of the Shi-Tomasi features of the carrier image and the damaged image_stThe image block where the key point is located is used as a candidate embedded block, the complexity of the candidate embedded block is calculated according to the information embedded complex region priority principle, and the first n with the maximum complexity is selected according to the secret information and the carrier length_tbAnd constructing a label block by marking the image block with the selected complexity and the rest unselected image blocks with different marks to determine whether the image blocks are embedded into the information carrier image blocks, wherein n is_st，n_tbRepresenting a natural number.

6. The robust steganography method based on channel matching network selection embedding channel as claimed in claim 1 or 5, wherein the channel matching network adopts a convolutional neural network AlexNet structure, and the constructed label block is used as a training sample to train the channel matching network so as to obtain the channel matching network with input as the image block and output as the label corresponding to the image block.

7. The robust steganography method for selecting an embedded channel based on a channel matching network as claimed in claim 1, wherein in constructing a robust carrier sequence and obtaining an embedded information modification amplitude sequence corresponding to a sequence element, if the attack type includes a plurality of image lossy processing operations, constructing a multiple robust embedded domain by using the difference of coefficients of adjacent blocks and obtaining an initial robust carrier sequence of a carrier image, and calculating a corresponding initial embedded information modification amplitude sequence, wherein the plurality of image lossy processing operations at least include image scaling, image denoising, and image rotation; and if the attack type only comprises image processing operation, constructing a robust embedding domain based on compression parameters by using a quantization rounding principle in a JPEG (joint photographic experts group) compression process, acquiring an initial robust carrier sequence of the carrier image, and calculating a corresponding initial embedding information modification amplitude sequence, wherein the image processing operation comprises image compression and/or image scaling.

8. The robust steganography method for selecting embedded channel based on channel matching network as claimed in claim 1, wherein the carrier element

Embedding cost

Is expressed as:

wherein the content of the first and second substances,

are respectively X,Y_iCorresponding spatial domain image, X representing the block of DCT coefficients of the carrier image, Y_iFor corresponding on-carrier elements

The block of DCT coefficients after the information has been embedded,

as a carrier element

for an image g corresponding to the uv-th wavelet coefficient in the first-level decomposition of the k subbands, u ∈ {1,2₁}，v∈{1,2,...,n₂τ is a constant used for stable numerical calculations.

9. The robust steganography method for selecting an embedded channel based on a channel matching network as claimed in claim 1, further comprising: the extraction of the secret-carrying image robust steganography for obtaining the secret information specifically comprises the following steps: carrying out image processing on the secret-carrying image according to the image processing attack type and parameters to obtain a processed image, selecting a blocking area with a mean square residual of 0 according to a saliency map, a mean square error and a mean square residual of the secret-carrying image and the processed image, carrying out scaling blocking processing on the secret-carrying image and the processed image to extract Shi-Tomasi characteristic key points, selecting image blocks where a plurality of characteristic key points are located according to the intensity, and testing the selected image blocks by using a channel matching network to obtain actual embedded information image blocks; constructing and extracting a robust secret-carrying sequence according to the attack type in the image processing attack type, and extracting a secret-carrying element which is positioned in an actual embedded information image block and does not belong to a block area to obtain a secret-carrying element sequence; and performing space-time code decoding on the secret-carrying element sequence by using the message length to obtain a decoded sequence, and correcting and inverting the decoded sequence to obtain secret information steganographically from the secret-carrying image.

10. A robust steganography system that selects an embedded channel based on a channel matching network, comprising: a transmission testing module, a channel matching module, a carrier acquiring module and a message coding module, wherein,