CN111563263A - Carrier-free information hiding method for migration of any image style - Google Patents
Carrier-free information hiding method for migration of any image style Download PDFInfo
- Publication number
- CN111563263A CN111563263A CN202010301278.8A CN202010301278A CN111563263A CN 111563263 A CN111563263 A CN 111563263A CN 202010301278 A CN202010301278 A CN 202010301278A CN 111563263 A CN111563263 A CN 111563263A
- Authority
- CN
- China
- Prior art keywords
- image
- feature map
- style
- sub
- image feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000013508 migration Methods 0.000 title claims abstract description 65
- 230000005012 migration Effects 0.000 title claims abstract description 65
- 238000000034 method Methods 0.000 title claims abstract description 63
- 239000011159 matrix material Substances 0.000 claims abstract description 24
- 238000013527 convolutional neural network Methods 0.000 claims description 29
- 238000010586 diagram Methods 0.000 claims description 12
- 239000013598 vector Substances 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 8
- 238000012549 training Methods 0.000 claims description 7
- 230000003044 adaptive effect Effects 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000005314 correlation function Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 229910002056 binary alloy Inorganic materials 0.000 claims description 2
- 230000000007 visual effect Effects 0.000 abstract description 6
- 238000013461 design Methods 0.000 abstract description 3
- 230000002194 synthesizing effect Effects 0.000 abstract description 3
- 238000004422 calculation algorithm Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 4
- 238000001514 detection method Methods 0.000 description 3
- 230000007246 mechanism Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 241000178435 Eliokarmos dubius Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 239000004576 sand Substances 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 239000011800 void material Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G06T3/04—
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention provides a carrier-free information hiding method for any image style migration. The method combines the idea of carrier-free information hiding and the operation of non-parameter method image style migration, designs a self-adaptive information hiding matrix according to the self-adaptive secret information coding and adjusting scheme of an input image. And under the guidance of the self-adaptive information hiding matrix, carrying out any image style migration and directly synthesizing an image style migration result driven by the secret information. The method not only keeps the advantages of artistic, more natural and high-fidelity visual effect of the image style migration result, but also has the characteristic that the carrier-free information hiding can resist steganalysis, and is superior to other existing carrier-free information hiding methods in the aspect of embedding capacity.
Description
Technical Field
The invention belongs to the field of carrier-free information hiding, and particularly relates to a carrier-free information hiding method for migration of any image style.
Background
Information hiding is a technique of hiding secret information into information that appears to be normal in an invisible manner. In the conventional image information hiding technique, an image is specified as an appropriate carrier, and then secret information is embedded therein to generate a secret-containing image. With the development of traditional information hiding, although the carrier image is slightly modified by the existing method, whether the secret information is hidden in the image can be detected through a steganalysis algorithm. For this purpose, experts propose bearer-free information hiding algorithms. Carrier-less information hiding can hide secret information and does not make any modification to the carrier image to resist existing steganalysis algorithms.
With the rise of deep learning, Gatys and Ecker and the like originally propose an image style migration method based on a convolutional neural network. However, because each style image needs to be trained by a corresponding network, a large amount of training time is required, and the trained network cannot perform any image style migration.
Disclosure of Invention
After comprehensively considering factors such as visual effect of image style migration, embedding mode and the like, the carrier-free information hiding network for the image style migration is provided. The network combines the idea of carrier-free information hiding and the operation of nonparametric method image style migration, designs a self-adaptive information hiding matrix which can be generated according to the self-adaptive secret information coding and adjusting scheme of an input image. Under the guidance of the self-adaptive information hiding matrix, carrying out any image style migration, and directly synthesizing an image style migration result driven by secret information.
The technical scheme of the invention comprises the following steps:
a carrier-free information hiding method for any image style migration comprises the following steps:
s1, obtaining a style image feature map from a target Relu layer of a convolutional neural network for image style migration, and extracting P with the size of m × m from the style image feature map in an overlapped modeSCarrying out Mean Shift clustering on all the subblocks to obtain K-type style image clustering results;
s2: determining a modified class number K' according to the clustering result of the style images:
wherein: p is the number of m multiplied by m subblocks which can be extracted by the self-adaptive image in a target Relu layer of the convolutional neural network at most in a non-overlapping mode; τ is the minimum number of subblocks included in each cluster;
the method for determining the size h multiplied by w of the self-adaptive image comprises the following steps:
h=min{hC,hS},
w=min{wC,wS}.
wherein: h isC×wCFor the content image size of the input convolutional neural network, hS×wSThe size of the style image input into the convolutional neural network;
if K 'is not equal to K, merging the clustered K-type style image clustering results by using a bottom-up AGNES clustering method until the number of classes is equal to K'; if K ═ K, keeping the clustering result of the K-type images unchanged;
s3: clustering results of K' type imagesClass is used as a buffer class, the remaining K' -1 classes are used to hide secret information, and then packets of secret information are computedThe length r-1 is calculated by the formula:
wherein: pCThe maximum number of m × m sub-blocks which can be extracted from a content image in a target Relu layer of a convolutional neural network in a non-overlapping mode;
dividing secret information to be hidden into K' -1 group B1,B2,...,BK′-1In which B isi=(b1b2... br-1), b j0 or 1, i 1,2, K' -1, j 1,2, r 1;
for each BiAdding a flag bit to form B'i:
In the formula: b'j=1-bj;
S4: e to be hiddennBit secret information is converted into B'1,B′2,...,B′K′-1Then, the code is converted from binary system to decimal system and is marked as D1,D2,...,DK′-1D is1,D2,...,DK′-1Respectively as the available data times constraint of 1,2, …, K' -1 types except the buffer type; and (3) assembling the times of using the data of each class into an adaptive information hiding matrix H according to the serial number of the class:
wherein: pbThe number of blocks necessary for said buffer class, when the number of blocks required for the network to perform the image style migration is P, P isbThe calculation formula of (2) is as follows:
s5: and (3) using the self-adaptive information hiding matrix H as the using times of the sub-block clustering in the target Relu layer of the convolutional neural network, and constraining the non-parametric style migration to obtain a secrecy-containing style migration result.
On the basis of the scheme, the steps can be further realized in the following specific mode.
Preferably, the specific method of Mean Shift clustering in S1 is as follows:
s11: randomly selecting one sub-block from all the sub-blocks of the style image feature map which are not classified as a central point, finding out all the sub-blocks with the distance from the central point within a set range, and recording the sub-blocks as a set M;
s12: calculating the offset vectors of the central point and the set M, and moving the central point along the offset vectors;
s13: repeating the step S12 until the size of the offset vector meets the set threshold value, and recording the value of the central point at the moment;
s14: continuously repeating S11-S13 until all the style image feature map sub-blocks are classified; and finally, sorting the coordinates of all the cluster centers according to the dimension order, wherein the coordinates are respectively numbered as 1,2, … and K.
Preferably, the convolutional neural network for image style migration is a VGG-19 network with a hole convolution.
Further, the target Relu layer is a Relu 3-1 layer in the VGG-19 network.
Preferably, the sub-block sizes extracted from the content image feature map and the grid image feature map are both 3 × 3.
Preferably, the specific step of S5 is as follows:
s51, inputting the content image C and the lattice image S into the convolutional neural network, obtaining a content image feature map and a lattice image feature map in a target Relu layer, and extracting all m × m sub-blocks f in a non-overlapping mode on the content image feature mapi(C),1≤i≤PCAll m × m sub-blocks f are extracted in an overlappable manner on the stylized image feature mapj(S),1≤j≤PS,PCAnd PSRespectively represent contentsThe number of small blocks which can be extracted from the image feature map and the style image feature map; and all the style image feature map sub-blocks are divided into K' style image clustering results according to the methods from S1 to S2;
s52: for each content image feature map sub-block fi(C) Selecting a cluster center and f from the K' type image clustering resultsi(C) Clustering the sub-blocks of the latest style image feature map;
s53: for each content image feature map sub-block fi(C) And screening all P 'with the distance from the center of the cluster not more than half of the radius of the cluster in the selected sub-block clusters of the style image feature map'SSub-block f of style image feature mapj(S),j=1,2,...,P′SThen, a sub-block f of the feature map of the content image is determined by using a cross-correlation functioni(C) Best matching intra-class optimal block fi st(C,S):
S54: for each content image feature map sub-block fi(C) Using intra-class optimal block fi st(C, S) replacement fi(C) And the optimal block f in the classi st(C, S) subtracting 1 from the use degree value of the cluster to which the cluster belongs in the information hiding matrix H;
if S54 is executed, the intra-class optimal block fi stThe using number value of the cluster to which (C, S) belongs in the information hiding matrix H is 0, and f isi st(C, S) performing S53 and S54 in the adjacent cluster of the cluster to which the S belongs;
s55: after the sub-blocks of the feature map of all the content images are replaced, a complete content image feature map F is obtained through reconstructionST(C,S)。
Preferably, all sub-blocks extracted from the content image feature map and the grid image feature map contain all channels of the feature map.
Preferably, a carrier-free information hiding network is constructed based on the carrier-free information hiding method, and the carrier-free information hiding network is trained and adoptedSquare error loss function, loss function LstyleThe (C, S) form is:
wherein | · | purple sweetFIs F norm, wc、hcAnd dcRespectively the width, height and channel number of the content image; f (I) in the training process, calculating a characteristic diagram of a result I in the current middle, wherein rho is a control parameter; l isTV(. cndot.) is the total variation regularization term, which is formulated as:
wherein, Ii,j,kThe k channel pixel value of the ith row and the jth column of the current intermediate calculation result I.
Compared with the prior art, the invention has the following beneficial effects:
the invention combines the advantages of the nonparametric method image style migration network method and realizes the carrier-free information hiding in the image style migration process. Compared with other image style migration networks, the method can realize the migration of any image style under the guidance of the secret information, so that the method can self-adaptively adjust the embedded digit of the secret information while ensuring the visual effect of the image style migration result, and balance the relation between the hidden capacity and the robustness of the carrierless information hiding algorithm. The invention can not only generate the migration result of any image style quickly, but also meet the requirements of safety and anti-detection in carrier-free information hiding. Compared with the existing machine learning-based carrier-free information hiding method, the method not only has larger information hiding capacity, but also can obtain artistic image style migration results.
Drawings
Fig. 1 is a schematic diagram of a carrier-free information hiding network structure for arbitrary image style migration (K ═ 9 is taken as an example).
FIG. 2 is a schematic diagram illustrating the influence of clustering class number of the stylized images on image stylistic migration.
FIG. 3 is a schematic diagram of a hole convolution with a receptive field of 9X 9.
Fig. 4 is a schematic diagram of a plurality of style migration results output by the carrierless information hiding network for arbitrary image style migration.
Detailed Description
The invention will be further elucidated and described with reference to the drawings and the detailed description.
In the invention, a self-adaptive information hiding matrix is generated according to a self-adaptive secret information coding and adjusting scheme of an input image by combining the idea of carrier-free information hiding and the operation of non-parametric method image style migration. And under the guidance of the self-adaptive information hiding matrix, carrying out any image style migration and directly synthesizing an image style migration result driven by the secret information. In addition, the network has good performances in the aspects of steganography capacity, anti-steganography analysis and security. The specific implementation steps of the unsupported information hiding method for arbitrary image style migration are described in detail below.
As shown in fig. 1, a schematic diagram is depicted for describing steps of the unsupported information hiding method for arbitrary image style migration, and the implementation process is as follows:
s1: mean Shift clustering is used for activated patches of a certain layer of the convolutional neural network.
Firstly, a convolutional neural network for image style migration is constructed, a style image feature map is obtained from a target Relu layer of the convolutional neural network, and all P with the size of m × m is extracted from the style image feature map in an overlapping mode (namely partial overlapping is allowed between two sub-blocks)SAnd (4) carrying out Mean Shift clustering on the subblocks to obtain K-type style image clustering results. In the embodiment, the convolutional neural network for image style migration is a VGG-19 network with hole convolution, and the target Relu layer is a Relu 3-1 layer in the VGG-19 network.
The specific method of Mean Shift clustering is as follows:
s11: randomly selecting one sub-block from all the sub-blocks of the style image feature map which are not classified as a central point, finding out all the sub-blocks with the distance from the central point within a set range, and recording the sub-blocks as a set M;
s12: calculating the offset vectors of the central point and the set M, and moving the central point along the offset vectors;
s13: repeating the step S12 until the size of the offset vector meets the set threshold value, and recording the value of the central point at the moment;
s14: continuously repeating S11-S13 until all the style image feature map sub-blocks are classified; and finally, sorting the coordinates of all the cluster centers according to the dimension order, wherein the coordinates are respectively numbered as 1,2, … and K.
S2: and (5) using AGNES clustering to the clustering result of the style images to determine the number K' of the modified classes.
First, the size of the content image input to the convolutional neural network is determined to be hC×wCInput the size h of the stylized image of the convolutional neural networkS×wSTaking the minimum value of the two as the size h × w of the adaptive image, namely:
h=min{hC,hS},
w=min{wC,wS}.
according to the adaptive image size h × w, the maximum number of m × m (m ═ 3) sub-blocks that can be extracted in a non-overlapping manner (i.e., no overlapping portion is allowed between any two sub-blocks) in the target Relu layer of the convolutional neural network is P, the minimum number of sub-blocks included in each cluster is τ, and then the modified number of clusters K' is:
k' is the cluster number which is suitable for the self characteristics of the image and keeps better robustness.
If K 'is not equal to K, merging the clustered K-type style image clustering results by using a bottom-up AGNES clustering method until the number of classes is equal to K'; and if K ═ K, keeping the clustering result of the K-type images unchanged. A schematic diagram of the impact of the number of clusters of stylized images on image style migration is shown in FIG. 2. In this embodiment, K' is 9, and thus remains unchanged.
Thus, in step S2, a K' type image clustering result can be obtained.
S3: the packet length r-1 is calculated from the number of modified classes K'.
Clustering results of K' type imagesClass as buffer class: (Representing rounding up), the remaining K' -1 class is used for hiding secret information, the number of m × m small blocks obtained by non-overlapping division of the content image in the target Relu layer of the convolutional neural network is maximum, and the number is PCThe average number of blocks per class that can be used for information hiding is(in the formulaRepresenting a rounding down), the packet length r-1 is:
dividing the secret information to be hidden into K' -1 group, and marking as B1,B2,...,BK′-1. Wherein B isiHaving the r-1 position, i.e. Bi=(b1b2... br-1), b i0 or 1, i 1,2, K' -1, j 1,2, r 1; then according to the following rule pair BiAdding a flag bit to form B'i:
In the formula: b'j=1-bj,j=1,2,...,r-1。
Taking secret information EnBit, converting the subsequence into B 'according to a coding mechanism'1,B′2,...,B′K′-1Converting the updated code from binary to decimal and recording as D1,D2,...,DK′-1. Will D1,D2,...,DK′-1As a constraint on the number of times data is available for use by the remaining classes other than the buffer class.
S4: and encoding the secret information into an adaptive information hiding matrix H by adopting a new encoding mechanism.
Taking secret information EnBit, converting the subsequence into B 'according to a coding mechanism'1,B′2,...,B′K′-1Then, the updated code is converted from binary to decimal and is marked as D1,D2,...,DK′-1D is1,D2,...,DK′-1Respectively as the number of data times constraints available for 1,2, …, K' -1 classes other than the buffer class. The total number of data constraints P available for use of the K' -1 class for encoding secret informationnComprises the following steps:
the number of blocks needed by the network for image style migration is P, and the number of blocks needed by the buffer class is PbComprises the following steps:
Pb=P-Pn.
summarizing the times of using the data of each class into an information hiding matrix H according to the serial number of the class, namely
S5: and (3) using the self-adaptive information hiding matrix H as the using times of the sub-block clustering in the target Relu layer of the convolutional neural network, and constraining the non-parametric style migration to obtain a secrecy-containing style migration result. The specific implementation process of the step is as follows:
s51: inputting the content image C and the lattice image S into the convolutional neural network (VGG-19 network with void convolution), and obtaining the content image characteristic in the target Relu layerExtracting all m × m sub-blocks f on the content image feature map in a non-overlapping modei(C),1≤i≤PCAll m × m sub-blocks f are extracted in an overlappable manner on the stylized image feature mapj(S),1≤j≤PS,PCAnd PSThe sub-blocks extracted from the content image feature map and the style image feature map are 3 × 3, i.e. m is 3, and all the sub-blocks of the style image feature map are divided into K' style image clustering results according to the methods from S1 to S2.
S52: for each content image feature map sub-block fi(C) Respectively calculating the distance between the cluster center and the cluster center of the K 'type image cluster, and further selecting the cluster center and the f from the K' type image cluster resulti(C) Clustering the sub-blocks of the feature map of the latest style image;
s53: for each content image feature map sub-block fi(C) And screening all P 'with the distance from the center of the cluster not more than half of the radius of the cluster in the selected sub-block clusters of the style image feature map'SSub-block f of style image feature mapj(S),j=1,2,...,P′SThen, a sub-block f of the feature map of the content image is determined by using a cross-correlation functioni(C) Best matching intra-class optimal block fi st(C,S):
S54: for each content image feature map sub-block fi(C) Using intra-class optimal block fi st(C, S) replacement fi(C) And the optimal block f in the classi st(C, S) subtracting 1 from the use degree value of the cluster to which the cluster belongs in the information hiding matrix H;
if S54 is executed, the intra-class optimal block fi stUse degree value of cluster to which (C, S) belongs in information hiding matrix HIs 0, then fi st(C, S) performing S53 and S54 in the adjacent cluster of the cluster to which the S belongs;
s55: after the sub-blocks of the feature map of all the content images are replaced, a complete content image feature map F is obtained through reconstructionST(C,S)。
Constructing a carrier-free information hiding network (CSST-Net) based on the carrier-free information hiding method, wherein the carrier-free information hiding network is trained by adopting a mean square error loss function and a loss function LstyleThe (C, S) form is:
wherein | · | purple sweetFIs F norm, wc、hcAnd dcRespectively the width, height and channel number of the content image; f (I) in the training process, calculating a characteristic diagram of a result I in the current middle, wherein rho is a control parameter; l isTV(. cndot.) is the total variation regularization term, which is formulated as:
wherein, Ii,j,kThe k channel pixel value of the ith row and the jth column of the current intermediate calculation result I.
Therefore, the carrier-free information hiding method can be used for embedding the secret information into the style migration image, and carrier-free information hiding is realized in the image style migration process.
For easier understanding, the invention further provides a specific extraction step of the secret information at the receiving end, which is detailed as follows:
step 1: the VGG-19 network with the hole convolution of CSST-Net is used for extracting the feature map of the style migration result, and the Relu 3-1 layer in the VGG-19 network is also used for obtaining the feature map at the same time as the encryption end. The resulting signature is divided into m × m sub-blocks in a non-overlapping manner.
Step 2: clustering the obtained feature map small blocks by using the Mean Shift algorithm to obtain C1,C2,...,CK′。
And step 3: counting the number of the characteristic diagram subblocks in each type after sorting, C1The number of the characteristic diagram subblocks in the class is n1, C2The number of small blocks of the class-like feature graph is n2,…,CK′The number of the characteristic diagram subblocks in the class is nK′. Assembling the sub-block quantity of each class of feature map into a matrix according to the numbering sequence of the classesNamely:
and 4, step 4: using the inverse procedure of the information hiding matrix H, the subblock number matrix is formedThe inverse solution results in an embedded secret information bit stream.
Therefore, the invention designs a self-adaptive secret information coding and adjusting scheme according to the input image by combining the idea of carrier-free information hiding and the operation of non-parametric method image style migration, and generates the self-adaptive information hiding matrix. And under the guidance of the adaptive information hiding matrix, carrying out arbitrary image style migration and directly combining the arbitrary image style migration and the arbitrary image style migration into an image style migration result driven by secret information. The method not only inherits the visual effect of the image style migration result and has the advantages of artistry, naturalness and high fidelity, but also has the characteristic that the carrier-free information hiding can resist the steganalysis, and is superior to other existing carrier-free information hiding methods in the aspect of embedding capacity. In order to show the effects achieved by the present invention, the method is applied to a specific embodiment, the specific steps are not described again, and the specific parameters and technical effects are mainly shown below.
Examples
In this embodiment, an image conversion task of image style transition is performed using CSST-Net. The implementation details and visual effect evaluation thereof will be analyzed in detail later. In addition, another important task of CSST-Net is to generate a codebook without carrier information hiding. Therefore, the present embodiment also performs experimental analysis of steganographic capacity, detectability resistance and security for the carrier-free information hiding operation.
1. Image style migration results
The CSST-Net feedforward network uses pre-trained VGG-19 with hole convolution, the inverse network ref [3] is trained using the Microsoft COCO (MSCOCO) data set, 15000 pictures are randomly drawn from the validation set as the training set, and three cycles are trained. Because Mean shift clustering and AGES clustering are adopted in the method, additional time is consumed in image style migration. All the following generated results are obtained in 18-25 seconds on a GPU of NVIDIA GeFoce GTX 10606 GB. The output of CSST-Net with both artistic and fidelity is shown in FIG. 4.
2. Network speed
Since the convolutional neural network is used for both the information hiding and the information extraction processes, the image style training speed and the style category of CSST-Net are considered aspects. The comparative results are shown in table 1:
TABLE 1 training comparison of Style migration networks
The specific practice of the above comparison method is as follows:
[1]Gatys L A,Ecker A S,Bethge M.A neural algorithm of artistic style[J]. arXiv preprint arXiv:1508.06576,2015.
[2]Li Y,Fang C,Yang J,et al.Diversified texture synthesis with feed-forward networks[C].IEEE Conference on Computer Vision and PatternRecognition.2017: 3920-3928.
[3]Chen T Q,Schmidt M.Fast patch-based style transfer of arbitrarystyle[J]. arXiv preprint arXiv:1612.04337,2016.
3. information hiding capacity
Since the image carrier-free information hiding method is not perfect enough, the capacity is not comparable to that of the traditional information hiding method, so that the comparison can be carried out only among a plurality of image carrier-free information hiding directions. The size of the information hiding capacity of the invention is self-adaptively determined according to the size characteristics of the content image and the style image. In the experiment, the size of 640 × 360 images had a hidden capacity of 48 bits when the number of clusters was equal to 9. The results of the specific comparisons are shown in Table 2.
TABLE 2 comparison of Carrier-free information hiding Capacity
The specific practice of the above comparison method is as follows:
[4] peripheral, YI, Sun Star, based on the Bag-of-Words model, application science bulletin, 2016,34(5):527 + 536.
[5]Zhou Z,Sun H,Harit R,et al.Coverless image steganography withoutembedding[C].International Conference on Cloud Computing andSecurity.Springer, Cham,2015:123-132.
[6]Yi C,Zhou Z,Yang C N,et al.Coverless information hiding based onFaster R-CNN[C].International Conference on Security with IntelligentComputing and Big-data Services.Springer,Cham,2018:795-807.
4. Resistance to detection and safety
CSST-Net directly hides the secret information in the image style migration result in an encoding mode, and does not modify the secret-contained image in the process of transferring the secret information. The results show that the image style migration result obtained by CSST-Net has good visual effect, can be better disguised as daily transmission images and is more difficult to cause the suspicion of attackers. Detection of statistical-based information hiding analysis algorithms can be fundamentally resisted.
The above-described embodiments are merely preferred embodiments of the present invention, which should not be construed as limiting the invention. Various changes and modifications may be made by one of ordinary skill in the pertinent art without departing from the spirit and scope of the present invention. Therefore, the technical scheme obtained by adopting the mode of equivalent replacement or equivalent transformation is within the protection scope of the invention.
Claims (8)
1. A carrier-free information hiding method for any image style migration is characterized by comprising the following steps:
s1, obtaining a style image feature map from a target Relu layer of a convolutional neural network for image style migration, and extracting P with the size of m × m from the style image feature map in an overlapping modeSCarrying out MeanShift clustering on the subblocks to obtain K-type style image clustering results;
s2: determining a modified class number K' according to the clustering result of the style images:
wherein: p is the number of m multiplied by m subblocks which can be extracted by the self-adaptive image in a target Relu layer of the convolutional neural network at most in a non-overlapping mode; τ is the minimum number of subblocks included in each cluster;
the method for determining the size h multiplied by w of the self-adaptive image comprises the following steps:
h=min{hC,hS},
w=min{wC,wS}.
wherein: h isC×wCFor the content image size of the input convolutional neural network, hS×wSThe size of the style image input into the convolutional neural network;
if K 'is not equal to K, merging the clustered K-type style image clustering results by using a bottom-up AGNES clustering method until the number of classes is equal to K'; if K ═ K, keeping the clustering result of the K-type images unchanged;
s3: clustering results of K' type imagesThe class is used as a buffer class, the remaining K' -1 class is used for hiding the secret information, and then the packet length r-1 of the secret information is calculated by the following formula:
wherein: pCThe maximum number of m × m sub-blocks which can be extracted from a content image in a target Relu layer of a convolutional neural network in a non-overlapping mode;
dividing secret information to be hidden into K' -1 group B1,B2,...,BK′-1In which B isi=(b1b2... br-1),bj0 or 1, i 1,2, K' -1, j 1,2, r 1;
for each BiAdding a flag bit to form B'i:
In the formula: b'j=1-bj;
S4: e to be hiddennBit secret information is converted into B'1,B′2,...,B′K′-1Then, the code is converted from binary system to decimal system and is marked as D1,D2,...,DK′-1D is1,D2,...,DK′-1Respectively as the available data times constraint of 1,2, …, K' -1 types except the buffer type; and (3) assembling the times of using the data of each class into an adaptive information hiding matrix H according to the serial number of the class:
wherein: pbThe number of blocks necessary for said buffer class, when the number of blocks required for the network to perform the image style migration is P, P isbThe calculation formula of (2) is as follows:
s5: and (3) using the self-adaptive information hiding matrix H as the using times of the sub-block clusters in the target Relu layer of the convolutional neural network, and constraining the non-parametric style migration to obtain a secrecy-containing style migration result.
2. The carrier-free information hiding network for arbitrary image style migration according to claim 1, wherein the specific method of Mean Shift clustering in S1 is as follows:
s11: randomly selecting one sub-block from all the sub-blocks of the style image feature map which are not classified as a central point, finding out all the sub-blocks with the distance from the central point within a set range, and recording the sub-blocks as a set M;
s12: calculating the offset vectors of the central point and the set M, and moving the central point along the offset vectors;
s13: repeating the step S12 until the size of the offset vector meets the set threshold value, and recording the value of the central point at the moment;
s14: continuously repeating S11-S13 until all the style image feature map sub-blocks are classified; and finally, sorting the coordinates of all cluster centers according to the dimensional order, wherein the coordinates are respectively numbered as 1,2, … and K.
3. The method for hiding the information without the carrier for the arbitrary image style migration according to claim 1, wherein the convolutional neural network for the image style migration is a VGG-19 network with a hole convolution.
4. The method according to claim 3, wherein the destination Relu layer is Relu 3-1 layer in VGG-19 network.
5. The method according to claim 1, wherein the sub-block sizes extracted from the content image feature map and the style image feature map are 3 x 3.
6. The method for hiding unsupported information for migrating any image style according to claim 1, wherein the specific steps of S5 are as follows:
s51, inputting the content image C and the lattice image S into the convolutional neural network, obtaining a content image feature map and a lattice image feature map in a target Relu layer, and extracting all m × m sub-blocks f in a non-overlapping mode on the content image feature mapi(C),1≤i≤PCAll m × m sub-blocks f are extracted in an overlappable manner on the stylized image feature mapj(S),1≤j≤PS,PCAnd PSRespectively representing the number of small blocks which can be extracted from the content image feature map and the style image feature map; and all the style image feature map sub-blocks are divided into K' style image clustering results according to the methods from S1 to S2;
s52: for each content image feature map sub-block fi(C) Selecting a cluster center and f from the K' type image clustering resultsi(C) Clustering the sub-blocks of the latest style image feature map;
s53: for each content image feature map sub-block fi(C) And screening all P 'with the distance from the center of the cluster not more than half of the radius of the cluster in the selected sub-block clusters of the style image feature map'SSub-block f of style image feature mapj(S),j=1,2,...,P′SThen, a sub-block f of the feature map of the content image is determined by using a cross-correlation functioni(C) Best matching intra-class optimal block fi st(C,S):
S54: for each content image feature map sub-block fi(C) Using intra-class optimal block fi st(C, S) replacement fi(C) And the optimal block f in the classi stNumber of uses of cluster to which (C, S) belongs in information hiding matrix HSubtracting 1 from the numerical value;
if S54 is executed, the intra-class optimal block fi stThe using number value of the cluster to which (C, S) belongs in the information hiding matrix H is 0, and f isi st(C, S) performing S53 and S54 in the adjacent cluster of the cluster to which the S belongs;
s55: after the sub-blocks of the feature map of all the content images are replaced, a complete content image feature map F is obtained through reconstructionST(C,S)。
7. The method according to claim 1, wherein all sub-blocks extracted from the content image feature map and the style image feature map contain all channels of the feature map.
8. The method according to claim 1, wherein the carrierless information hiding network is constructed based on the carrierless information hiding method, and the carrierless information hiding network is trained by using a mean square error loss function, a loss function LstyleThe (C, S) form is:
wherein | · | purple sweetFIs F norm, wc、hcAnd dcRespectively the width, height and channel number of the content image; f (I) in the training process, calculating a characteristic diagram of a result I in the current middle, wherein rho is a control parameter; l isTV(. cndot.) is the total variation regularization term, which is formulated as:
wherein, Ii,j,kThe k channel pixel value of the ith row and the jth column of the current intermediate calculation result I.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010301278.8A CN111563263B (en) | 2020-04-16 | 2020-04-16 | Carrier-free information hiding method for migration of arbitrary image style |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010301278.8A CN111563263B (en) | 2020-04-16 | 2020-04-16 | Carrier-free information hiding method for migration of arbitrary image style |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111563263A true CN111563263A (en) | 2020-08-21 |
CN111563263B CN111563263B (en) | 2023-04-25 |
Family
ID=72073153
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010301278.8A Active CN111563263B (en) | 2020-04-16 | 2020-04-16 | Carrier-free information hiding method for migration of arbitrary image style |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111563263B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112288622A (en) * | 2020-10-29 | 2021-01-29 | 中山大学 | Multi-scale generation countermeasure network-based camouflaged image generation method |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190220746A1 (en) * | 2017-08-29 | 2019-07-18 | Boe Technology Group Co., Ltd. | Image processing method, image processing device, and training method of neural network |
CN110533570A (en) * | 2019-08-27 | 2019-12-03 | 南京工程学院 | A kind of general steganography method based on deep learning |
CN110677552A (en) * | 2019-08-30 | 2020-01-10 | 绍兴聚量数据技术有限公司 | Carrier-free information hiding method based on complete packet bases |
-
2020
- 2020-04-16 CN CN202010301278.8A patent/CN111563263B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190220746A1 (en) * | 2017-08-29 | 2019-07-18 | Boe Technology Group Co., Ltd. | Image processing method, image processing device, and training method of neural network |
CN110533570A (en) * | 2019-08-27 | 2019-12-03 | 南京工程学院 | A kind of general steganography method based on deep learning |
CN110677552A (en) * | 2019-08-30 | 2020-01-10 | 绍兴聚量数据技术有限公司 | Carrier-free information hiding method based on complete packet bases |
Non-Patent Citations (3)
Title |
---|
DORIN COMANICIU 等: "Mean shift: A robust Approach Toward Feature Space Analysis" * |
FISHER YU: "MULTI-SCALE CONTEXT AGGREGATION BY DILATED CONVOLUTIONS" * |
TIAN QI CHEN 等: "Fast Patch-based Style Transfer of Arbitrary Style" * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112288622A (en) * | 2020-10-29 | 2021-01-29 | 中山大学 | Multi-scale generation countermeasure network-based camouflaged image generation method |
Also Published As
Publication number | Publication date |
---|---|
CN111563263B (en) | 2023-04-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107340993B (en) | Arithmetic device and method | |
US9619749B2 (en) | Neural network and method of neural network training | |
US9390373B2 (en) | Neural network and method of neural network training | |
CN108346125A (en) | A kind of spatial domain picture steganography method and system based on generation confrontation network | |
WO2017185412A1 (en) | Neural network operation device and method supporting few-bit fixed-point number | |
CN110570346A (en) | Method for performing style migration on calligraphy based on cyclic generation countermeasure network | |
CN112464003A (en) | Image classification method and related device | |
CN112489164B (en) | Image coloring method based on improved depth separable convolutional neural network | |
CN112132258B (en) | Multitask image processing method based on deformable convolution | |
CN103914527B (en) | Graphic image recognition and matching method based on genetic programming algorithms of novel coding modes | |
CN109300128A (en) | The transfer learning image processing method of structure is implied based on convolutional Neural net | |
CN111563263A (en) | Carrier-free information hiding method for migration of any image style | |
CN104103042A (en) | Nonconvex compressed sensing image reconstruction method based on local similarity and local selection | |
CN109388959B (en) | Generation type information hiding method combining difference clustering and minimum error texture synthesis | |
CN108804933B (en) | Method for carrying out scale conversion on big data in information hiding technology | |
CN111031191B (en) | Image encryption method based on controlled alternate quantum walking and DNA sequence operation | |
CN111783989A (en) | S box optimization method based on improved genetic algorithm | |
CN110086955A (en) | A kind of large capacity image latent writing method | |
Zhang et al. | Adaptive robust blind watermarking scheme improved by entropy-based SVM and optimized quantum genetic algorithm | |
CN113516199B (en) | Image data generation method based on differential privacy | |
CN113610878B (en) | Structural information hiding method based on automatic contour generation | |
Chiu et al. | An XOR-based progressive visual cryptography with meaningful shares | |
Usha et al. | Cognitive prediction of the most appropriate image steganography approach | |
CN111444985A (en) | Image matching method based on histogram matching | |
CN113222160A (en) | Quantum state conversion method and device |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |