CN110334805A - A kind of JPEG domain image latent writing method and system based on generation confrontation network - Google Patents
A kind of JPEG domain image latent writing method and system based on generation confrontation network Download PDFInfo
- Publication number
- CN110334805A CN110334805A CN201910368809.2A CN201910368809A CN110334805A CN 110334805 A CN110334805 A CN 110334805A CN 201910368809 A CN201910368809 A CN 201910368809A CN 110334805 A CN110334805 A CN 110334805A
- Authority
- CN
- China
- Prior art keywords
- network
- image
- matrix
- jpeg
- probability
- 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
- 238000000034 method Methods 0.000 title claims abstract description 37
- 239000011159 matrix material Substances 0.000 claims abstract description 96
- 238000006243 chemical reaction Methods 0.000 claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 18
- 230000003044 adaptive effect Effects 0.000 claims abstract description 5
- 230000009977 dual effect Effects 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 70
- 238000003780 insertion Methods 0.000 claims description 41
- 230000037431 insertion Effects 0.000 claims description 41
- 230000004069 differentiation Effects 0.000 claims description 13
- 238000013527 convolutional neural network Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 11
- 230000000694 effects Effects 0.000 claims description 10
- 238000004891 communication Methods 0.000 claims description 6
- 238000003475 lamination Methods 0.000 claims description 6
- 238000013139 quantization Methods 0.000 claims description 6
- 238000012986 modification Methods 0.000 claims description 5
- 230000004048 modification Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 239000000463 material Substances 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 4
- 238000012546 transfer Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 3
- 230000000644 propagated effect Effects 0.000 claims description 3
- 230000008901 benefit Effects 0.000 claims description 2
- 230000002708 enhancing effect Effects 0.000 claims description 2
- 238000012886 linear function Methods 0.000 claims description 2
- 238000013461 design Methods 0.000 abstract description 2
- 238000001514 detection method Methods 0.000 abstract 1
- 238000004422 calculation algorithm Methods 0.000 description 13
- 238000013528 artificial neural network Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 230000000007 visual effect Effects 0.000 description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 230000006835 compression Effects 0.000 description 2
- 238000007906 compression Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 230000012010 growth Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000000969 carrier Substances 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002401 inhibitory effect Effects 0.000 description 1
- 230000006855 networking Effects 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T9/00—Image coding
- G06T9/002—Image coding using neural networks
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/60—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding
- H04N19/625—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using transform coding using discrete cosine transform [DCT]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Discrete Mathematics (AREA)
- Signal Processing (AREA)
- Editing Of Facsimile Originals (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
- Compression Of Band Width Or Redundancy In Fax (AREA)
Abstract
The invention discloses a kind of based on the JPEG domain image latent writing method for generating confrontation network, generate that carrier image DCT coefficient matrix is corresponding distorts probability matrix by generating network, module, which is embedded in, using analog encoding and can transmit gradient JPEG conversion module generates the corresponding close image of load according to probability matrix is distorted, by differentiating that network distinguishes carrier image with close image is carried, using error in classification as loss function to generating network and differentiating that network carries out dual training, the generation network model that can generate adaptive steganography cost value is finally obtained.Through the model in conjunction with conventional information coding module, secret information is embedded into carrier image and obtains carrying close image.Have design simple compared to traditional JPEG domain image latent writing method, it is easy to accomplish, the features such as anti-detection property is strong.The invention also discloses a kind of JPEG domain image latent writing systems, including based on the obtained generation network module of JPEG domain image latent writing method for generating confrontation network, information coding module and JPEG conversion module.
Description
Technical field
The present invention relates to image latent writing fields, more particularly, to a kind of based on the JPEG domain image for generating confrontation network
Steganography method.
Background technique
Information hiding is that the secret information that will need to encrypt is embedded into generation load ciphertext in bearer documents by certain algorithms
Part, and transmitted by open channel, recipient carries out the secret information carried in ciphertext part by corresponding extraction means
It extracts, and attacker can not obtain secret information from carrying in ciphertext part.The Information hiding of image is mainly by the part number of image
It is replaced or modifies according to such as pixel value etc., secret information is embedded under the premise of not influencing image visual effect and carries close figure
As in, the visual redundancy of human eye is mainly utilized in the Information hiding of image.Since digital picture has information capacity big, distort easily
In realization, and the features such as quantity and many kinds of format, therefore digital picture is a big main carriers of Information hiding.Steganography is calculated
Method is the reversible hiding main implementation of image information, with the continuous development of network and multimedia technology, image conduct
Common information media on network, the steganography and steganalysis of image, which also gradually receive, to be widely applied.
The steganography method of image is broadly divided into the steganography of airspace and domain of variation, root according to the classification of modification image data at present
It is divided into the steganography of picture material adaptivity and non adaptive again according to the mode that secret information is embedded in.It is corresponding with image latent writing
It is the steganalysis of image, for judging whether image is embedded with secret information, steganalysis can be used as steganographic algorithm safety
A kind of evaluation method of property.In general the steganographic algorithm with adaptivity all has relatively high safety.With depth
Application of the degree study in image latent writing analysis, steganalysis performance is greatly improved, while the peace of image latent writing method
Full property is challenged.
Image latent writing algorithm on JPEG domain is first by being transformed into JPEG domain for the pixel value on airspace, by image
DCT coefficient is distorted to be embedded in secret information, in order to enable insertion behavior reduces the change of statistical nature as far as possible, steganography is calculated
Method uses the embedding grammar with adaptivity.In order to reduce visual change, generally embed of information into after modification to figure
As visual effect influences in lesser exchange summation about non-zero DCT coefficients.Conventional method realizes information by setting additivity distortion function
The adaptivity of insertion, but this method height relies on the experience and priori knowledge of designer, and can not be according to steganalysis net
Change make adjustment on strategy in time, therefore its safety still has a certain upgrade space.It does not appear at present
The variation according to steganalysis network of JPEG domain adjusts the steganography cost function self-learning algorithm of itself strategy in time.
Summary of the invention
The present invention provides a kind of JPEG domain image latent writing method and system based on generation confrontation network, using deep learning
In generation confrontation network carry out JPEG domain on steganography.
In order to solve the above technical problems, technical scheme is as follows:
A kind of JPEG domain image latent writing method based on generation confrontation network, which is characterized in that the generation fights net
Network includes generating network, analog encoding insertion module and differentiating network.Its training process the following steps are included:
S1: it by carrier image by the way that gradient JPEG conversion module can be transmitted, obtains in the corresponding DCT domain of the carrier image
Coefficient matrix；
S2: the corresponding DCT coefficient matrix of the carrier image is input in the generation network, is generated by generating network
Corresponding and carrier image is of the same size to distort probability matrix, and the numerical value of described each position for distorting probability matrix represents should
The probability that the corresponding location of pixels DCT coefficient in position is tampered with；
S3: distort probability matrix and the random noise matrix of generation are embedded in module progress analog encoding by analog encoding
It obtaining and carrier image is of the same size distorts matrix, the random noise matrix and the probability matrix of distorting are in the same size,
The value for distorting element in matrix is ± 1 and 0, and random noise matrix is for simulating random secret information；Matrix will be distorted and carried
Matrix in the corresponding DCT domain of body image is added, and obtains the close image of load corresponding with carrier image DCT coefficient matrix in DCT domain
DCT coefficient matrix；
S4: the DCT coefficient matrix of the close image of load in DCT domain is converted to by that can transmit the JPEG conversion module of gradient
The close image of load on airspace；
S5: being input to differentiation network for the S4 carrier image for obtaining carrying close image and S1, by differentiating network to carrier figure
As classifying with the close image of load, the error in classification for differentiating that network generates is fed back to as loss function, and by loss function
Differentiate network and generates network and carry out dual training；
S6: after training, by trained generation network and traditional information coding module and JPEG conversion module
It combines, secret information is embedded into carrier image in an adaptive way according to the insertion cost value that network generates is generated
The close image of load for being used for secret communication is generated in DCT coefficient, the insertion cost value is to distort probability conversion gained.
Preferably, described based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that described to pass
Passing gradient JPEG conversion module includes image real time transfer and matrixing, and wherein image real time transfer includes Tensor data
Batch separation, the piecemeal processing of single batch, the merging of piecemeal and the merging of batch；Gradient JPEG transformation can be transmitted
Matrixing in module include 2-D discrete cosine matrixing, anti-two-dimension discrete cosine transform and corresponding quantization with it is anti-
Quantification treatment, wherein inverse quantization does not use floor operation to ensure that gradient is effectively transmitted.
Preferably, the generation network in the step S2 is a full convolutional neural networks in parallel or a U-shaped structure
Convolutional neural networks, wherein full convolutional neural networks in parallel are divided into four branches, each branch includes three convolutional layers, four
The convolution kernel size of a branch is respectively 2,3,4,5, each convolutional layer includes 50 convolution kernels, is all made of band leakage rectification letter
Number (Leak ReLU) is used as activation primitive.
Preferably, the convolutional neural networks that network is a U-shaped structure of generating in the step S2, including 16 groups of subnets
Network structure, first 8 groups are convolutional layer group, each convolutional layer group includes convolutional layer batch, normalization layer and rectifies function with leakage,
The size of characteristics of image figure is reduced with the increase of the network number of plies；8 groups are warp lamination group afterwards, comprising warp lamination, criticize and return
One changes layer and rectifies function with leakage, and the size of characteristics of image figure becomes larger with the increase of the network number of plies, and first 8 layers and rear 8
Characteristic pattern between the corresponding image network layer of layer is all attached in a manner of series connection (concatenate).
Preferably, analog encoding insertion module is made of optimum code fitting function in step S3, and wherein optimum code is quasi-
Closing the optimum code mode that function is fitted may be expressed as:
In above formula, pijThe each DCT coefficient for making a living into network output corresponding distorts probability, nijFor random noise matrix,
mijTo distort matrix；Optimum code fitting function is characterized in that keeping gradient propagable while can be fitted above-mentioned optimal
Coding mode；The wherein expression formula of optimum code fitting function are as follows:
mij=-k × fs(λ×(pij-2×nij))+k×fs(λ×(pij-2×(1-nij)))+c
Wherein λ is that its value of zoom factor is bigger, and the fitting effect of optimum code fitting function is better, and gradient is propagated
More unfavorable, zoom factor λ value is smaller, is more conducive to the propagation of gradient, but fitting effect can be deteriorated, fsIndicate a Class Activation letter
Number, such as: Tanh, Sigmoid, ReLU, k are weight coefficient, and c is a constant biasing.
Preferably, differentiation network described in step S5 is the steganalysis convolutional neural networks based on DenseNet, step
Differentiation network described in S5 is the steganalysis convolutional neural networks based on DenseNet, including 1 pretreatment layer and 7
Dense convolutional layer group, wherein pretreatment layer includes the strong letter that 16 different types of high-pass filters are used to inhibit picture material
Number, the weak signal of enhancing steganography insertion modification, each dense convolutional layer group includes 4 convolutional layers and 1 pond layer, convolutional layer
All it is connected with its each convolutional layer for being located at front in the same dense convolutional layer group, the reuse of characteristic pattern is realized, using linear
Function is rectified as activation primitive, final class probability is obtained using full articulamentum and softmax function and is exported.
It is reversely passed with the loss function that network is obtained using error in classification is differentiated using gradient preferably for network is generated
The training for carrying out parameter is broadcast, every primary parameter for differentiating network of update then updates the parameter for generating network twice, wherein for more
The loss function of newly-generated network also includes to be embedded in load capacity and targeted loads amount by distorting the fitting that probabilistic information entropy obtains
Between error generate loss function.
Preferably, the iteration update times are greater than 120,000 time.
Preferably, the probabilistic information entropy calculation formula is as follows:
In above formula, H is probabilistic information entropy, and h is the height of image, the width that w is image,Indicate the DCT system for being located at (i, j)
Number carries out+1 probability value distorted,Indicate that the DCT coefficient of position (i, j) carries out -1 probability value distorted,Indicate position
The DCT coefficient of (i, j) is without the probability value distorted.
Preferably, the loss function for differentiating network is the cross entropy of error in classification, specific formula for calculation are as follows:
In above formula, yiIt indicates to differentiate the probability that network divides the image into the i-th class, yi' indicate the actual class label of image.
Preferably, the loss function of network is generated by the loss function of differentiation network and the loss of fitting insertion load capacity
Function collectively constitutes, specific formula are as follows:
lG=α ×-lD+β×lp
Wherein lp=(H- γ × q)2For the loss function of fitting insertion load capacity, γ is non-zero in carrier image DCT coefficient
The quantity of ac coefficient, q are the target insertion load capacity of setting.α and β is the weight coefficient for adjusting two loss functions,
It is characterized in that, the selection of α and β setting value needs to weigh the fitting precision of steganography safety and insertion rate.
Preferably, the corresponding insertion cost value ρ of each DCT coefficient that network generates is generatedI, j, to distort Probability pI, jConversion
Gained, specific conversion formula are as follows:
A kind of JPEG domain image latent writing system, which is characterized in that including hidden based on the JPEG domain image for generating confrontation network
The generation network module that write method obtains, information coding module, JPEG conversion module, the specific steps are as follows:
The airspace carrier image of input is converted to DCT coefficient matrix using JPEG conversion module, then passes through the generation
Network module generates distorting probability matrix and being converted to insertion cost value for corresponding DCT coefficient matrix, according to insertion cost value with
Practical secret information distorts matrix using information coding module generation size identical as image, by the matrix and initial carrier figure
As corresponding DCT matrix phase adduction is input to the close image of load that the generation of JPEG conversion module is used for practical secret communication.Wherein use
The confrontation trained generation group of networks of network training module is generated by passing through in the generation network module of jpeg image steganographic system
At.
Compared with prior art, the beneficial effect of technical solution of the present invention is:
1) present invention does not need a large amount of steganography using based on the JPEG domain image latent writing method fighting network and obtaining is generated
The experience and priori knowledge in field.Final steganography model is shown that design is relatively easy by neural network after training, is easy to
It realizes.
2) realize can disease gradient JPEG transformation, to image carry out frequency domain conversion while, do not influence to lose
Backpropagation of the function in neural network.Have many advantages, such as precision it is high, it is high-efficient, do not need pre-training.
3) it is realized under the premise of guaranteeing not influencing the transmitting of neural network gradient to most by optimum code fitting function
The fitting of excellent coding has been obviously improved the safety of steganographic algorithm.
4) proposed by the present invention based on the JPEG domain image latent writing method for generating confrontation network, core concept is made a living networking
Game training between network and differentiation network, by the confrontation between two networks, effectively improves the adaptivity of steganographic algorithm
With safety.
Detailed description of the invention
Fig. 1 is flowage structure figure of the invention；
Fig. 2 is the flow chart based on dual training in the JPEG domain image latent writing method for generating confrontation network；
Fig. 3 is that can transmit gradient JPEG conversion module structure chart in the present invention；
Fig. 4 is the full convolutional network model structure of four branch circuit parallel connections；
Fig. 5 is U-shaped structure Artificial Neural Network Structures figure；
Fig. 6 is double tangent optimum code fitting functions and actual optimum coding function effect contrast figure；
Fig. 7 is the steganalysis neural network structure figure based on DenseNet；
Specific embodiment
The attached figures are only used for illustrative purposes and cannot be understood as limitating the patent；
In order to better illustrate this embodiment, the certain components of attached drawing have omission, zoom in or out, and do not represent actual product
Size；
To those skilled in the art, it is to be understood that certain known features and its explanation, which may be omitted, in attached drawing
's.
The following further describes the technical solution of the present invention with reference to the accompanying drawings and examples.
Embodiment 1
The present embodiment provides a kind of based on the JPEG domain image latent writing method for generating confrontation network, wherein generating confrontation network
Trained detailed process is as shown in Figure 2, comprising the following steps:
S1: initial carrier image, which is input to, can transmit gradient JPEG conversion module, obtain individual load using slice function
Body image carries out every carrier image the piecemeal of 8*8, by carrying out square to each matrix in block form and dct transform matrix
Battle array variation operation, obtains the corresponding DCT matrix of spatial domain picture, then by the amount of DCT coefficient matrix and corresponding compression quality factor Q F
Change the matrix DCT coefficient matrix after being quantified of being divided by finally obtain image by the splicing of piecemeal and correspond on JPEG domain
DCT coefficient matrix.The specific implementation proposed by the present invention for transmitting gradient JPEG conversion module is as shown in Fig. 3.
S2: the DCT coefficient matrix of image being input to and is generated in network, by obtaining and carrier figure after generating network processes
Probability matrix P is distorted as size is consistent, probability matrix is distorted for substituting in traditional steganographic algorithm, is got over wherein distorting probability
Big point will be distorted preferentially.Generation network is a full convolutional neural networks in parallel, as shown in figure 4, it is divided into four branches,
Each branch includes three convolutional layers, and the convolution kernel size of four branches is respectively 2,3,4,5, each convolutional layer includes 50
Convolution kernel is all made of band leakage rectification function as activation primitive.
S3: the random noise matrix (simulation secret information) for distorting probability matrix and identical size that step S2 is obtained
It is input to analog encoding insertion module and generates DCT coefficient and distort matrix, carrier image matrix and step S1 will be distorted obtained
DCT coefficient matrix be added to obtain the DCT coefficient matrix for carrying close image.
Wherein (double tangents are quasi- for the optimum code fitting function one of which specific implementation formula in analog encoding insertion module
Close function) are as follows:
Wherein activation primitive fsAdopt tangent (tanh) function.In above formula, pijThe pixel for making a living into network output corresponding is usurped
Change probability, nijFor random noise matrix, λ is the zoom factor of function, mijTo distort matrix；Zoom factor λ value is bigger,
The effect for being fitted optimal embedding function is better, and it is more unfavorable that gradient is propagated, and zoom factor λ value is smaller, is more conducive to gradient
Propagation, but fitting effect can be deteriorated.
When the zooming parameter λ value for being experimentally confirmed double tangent fitting functions is 100, before not influencing gradient transmitting
It puts, can achieve preferable fitting effect, therefore testing λ value is 100.Attached drawing 6 illustrates double tangents proposed by the present invention
The encoding efficiency of fitting function coding module (λ=100) and actual step function compares.
S4: the DCT coefficient matrix for carrying close image is converted to spatial domain picture by the JPEG conversion module by that can transmit gradient,
Finally obtain the close image of load on airspace.
S5: the carrier image on airspace is input to differentiation network (i.e. steganalysis network, specific structure with close image is carried
Corresponding tag along sort is obtained as shown in Fig. 7), by calculating error in classification:
Corresponding loss function is obtained, l is minimized by back-propagation algorithmsTo train differentiation network, maximization lsTraining
Network is generated, while passing through the probabilistic information entropy of fitting insertion load capacity:
Obtain the loss function l of fitting insertion load capacityp=(H- γ × q)2, utilize loss function:
lG=α ×-lD+β×lp
Come train generate network make the insertion load capacity of steganographic algorithm meet setting value, wherein H be probabilistic information entropy, h
Height, w for image are the width of image,Indicate that the DCT coefficient positioned at (i, j) carries out+1 probability value distorted,Indicate position
The DCT coefficient for setting (i, j) carries out -1 probability value distorted,Indicate the DCT coefficient of position (i, j) without the probability distorted
Value, γ are the quantity of non-zero ac coefficient in carrier image DCT coefficient, and q is the target insertion load capacity of setting.Weight coefficient α
=1, β=10-7。
Wherein differentiate that network uses the steganalysis network based on DenseNet of current better performances, specific structure is such as
Shown in attached drawing 7, including 1 pretreatment layer and 7 convolutional layer groups, connected in a manner of intensive (dense) between characteristic pattern
It connects, wherein pretreatment layer includes that 16 different types of high-pass filters are used for feature extraction, and each dense convolutional layer group includes
4 convolutional layers and 1 pond layer, convolutional layer are all located at each convolutional layer phase of front in the same dense convolutional layer group with it
Even, the reuse of characteristic pattern is realized, using the linear unit function of amendment as activation primitive, using full articulamentum and softmax letter
Number obtains final class probability output.
The opposite generation network of convergence rate due to differentiating network is very fast, and every iteration updates the primary ginseng for differentiating network
Number, it is corresponding to update the parameter for generating network twice.Update is iterated to network parameter using Adam optimizer, batch having a size of
12, the number of iterations is greater than 120,000 time.
After training, by the good generation network of 120,000 time or more repetitive exercise and grid coding will be used
The information coding module and JPEG conversion module of (Syndrometrelliscode, STC) combine, by secret information with certainly
The mode of adaptation is embedded into the close image of load for generating in the DCT coefficient of carrier image and being used for secret communication.
Embodiment 2
The present embodiment provides a kind of based on the JPEG domain image latent writing method for generating confrontation network, wherein generating confrontation network
Trained detailed process is as shown in Figure 2, comprising the following steps:
S1: initial carrier image, which is input to, can transmit gradient JPEG conversion module, obtain individual load using slice function
Body image carries out the division of 8*8 block for every carrier image, by carrying out square to each matrix in block form and dct transform matrix
Battle array variation operation, obtains the corresponding DCT coefficient matrix of spatial domain picture, then by DCT coefficient matrix and corresponding compression quality factor Q F
The quantization matrix DCT coefficient matrix after being quantified of being divided by finally obtain image by the splicing of piecemeal and correspond to JPEG
The DCT coefficient matrix with size identical as carrier on domain.The tool proposed by the present invention for transmitting gradient JPEG conversion module
Body is realized as shown in Fig. 3.
S2: the DCT coefficient matrix of image being input to and is generated in network, by obtaining and carrier figure after generating network processes
Probability matrix P is distorted as size is consistent, probability matrix is distorted for substituting in traditional steganographic algorithm, is got over wherein distorting probability
Big point will be distorted preferentially.It generates network and uses U-shaped structure neural network shown in fig. 5, including 16 groups of sub-network structures, it is preceding
8 groups are convolutional layer group, each convolutional layer group includes convolutional layer batch, normalization layer and rectifies function, characteristics of image with leakage
The size of figure is reduced with the increase of the network number of plies；Afterwards 8 groups be warp lamination group, comprising warp lamination, batch normalization layer with
And function is rectified with leakage, the size of characteristics of image figure becomes larger with the increase of the network number of plies, and first 8 layers are corresponding with latter 8 layers
Characteristic pattern between image network layer is all attached in a manner of series connection (concatenate).
S3: the random noise matrix (simulation secret information) for distorting probability matrix and identical size that step S2 is obtained
It is input to analog encoding insertion module, generates and distorts matrix on Image DCT coefficient, the load that matrix will be distorted and step S1 is obtained
The DCT coefficient matrix of body image is added to obtain the DCT coefficient matrix for carrying close image.
A kind of wherein specific implementation formula (double S type growths of the optimum code fitting function in analog encoding insertion module
Iunction for curve) are as follows:
mij=-sigmoid (λ × (pij-2×nij))+sigmoid(λ×(pij-2×(1-nij))) wherein activation primitive fs
Using S sigmoid growth curve (Sigmoid) function.In above formula, pijThe pixel for making a living into network output corresponding distorts probability, nij
For random noise matrix, λ is the zoom factor of function, mijTo distort matrix；Zoom factor λ value is bigger, is fitted optimal embedding
The effect for entering function is better, but more unfavorable for gradient propagation, and zoom factor λ value is smaller, is more conducive to the propagation of gradient,
But fitting effect can be deteriorated.Wherein λ value is 100.
S4: the DCT coefficient matrix of the close image of carrier/load is converted to airspace by the JPEG conversion module by that can transmit gradient
Image finally obtains the close image of carrier/load on airspace.
S5: the carrier image on airspace is input to differentiation network (i.e. steganalysis network, specific structure with close image is carried
Corresponding tag along sort is obtained as shown in Fig. 7), by calculating error in classification:
Corresponding loss function is obtained, l is minimized by back-propagation algorithmsTo train differentiation network, maximization lsTraining
Network is generated, while passing through the probabilistic information entropy of fitting insertion load capacity:
Obtain the loss function l of fitting insertion load capacityp=(H- γ × q)2, utilize loss function:
lG=α ×-lD+β×lp
It generates network to train the insertion load capacity of steganographic algorithm is made to meet setting value, in above formula, H is probabilistic information
Entropy, h are the height of image, the width that w is image,Indicate that the DCT coefficient positioned at (i, j) carries out+1 probability value distorted,Table
Show that the DCT coefficient of position (i, j) carries out -1 probability value distorted,Indicate the DCT coefficient of position (i, j) without distorting
Probability value, γ are the quantity of non-zero ac coefficient in carrier image DCT coefficient, and q is the target insertion load capacity of setting.Weight system
Number α=1, β=10-7。
Wherein differentiate that network uses the steganalysis network based on DenseNet of current better performances, specific structure is such as
Shown in attached drawing 7, including 1 pretreatment layer and 7 dense convolutional layer groups, wherein pretreatment layer includes 16 different types of
High-pass filter expands insertion weak signal, each dense convolutional layer group includes 4 convolutional layers and 1 for inhibiting picture material
Pond layer, convolutional layer are all connected with its each convolutional layer for being located at front in the same dense convolutional layer group, realize characteristic pattern
It reuses, using linear unit function is corrected as activation primitive, final classification is obtained using full articulamentum and softmax function
Probability output.
The opposite generation network of convergence rate due to differentiating network is very fast, and every iteration updates the primary ginseng for differentiating network
Number, it is corresponding to update the parameter for generating network twice.Update is iterated to network parameter using exchange Adam optimizer, batch size
It is 12, the number of iterations is greater than 120,000 time.
S6 training after, by trained generations network with use grid coding (Syndrometrelliscode,
STC information coding module and JPEG conversion module) combines, and secret information is embedded into carrier figure in an adaptive way
The close image of load for being used for secret communication is generated in the DCT coefficient of picture.
Embodiment 3
The present embodiment provides a kind of to fight net based on the jpeg image steganographic system for generating confrontation network, including based on generating
The generation network module that the JPEG domain image latent writing method of network obtains, information coding module, JPEG conversion module are schemed as shown in figure 1
As shown in steganographic system, its step are as follows:
S1: the carrier image of input is converted into DCT coefficient matrix by JPEG conversion module.
S2: the DCT coefficient matrix of carrier image is input to and generates network module and obtains distorting probability matrix and be converted to
Corresponding insertion cost value.Wherein generating and generating the parameter of network in network module is by generating confrontation network training module instruction
Parameter after white silk.
S3: by the information coding module using grid coding (Syndrometrelliscode, STC) according to insertion cost
What value and practical secret information generated size identical as image distorts matrix.
S4: matrix will be distorted and be added to obtain the DCT coefficient matrix for carrying close image with the DCT coefficient matrix of carrier image.
S5: the close image of the load on airspace is obtained using JPEG conversion module.
The same or similar label correspond to the same or similar components；
The terms describing the positional relationship in the drawings are only for illustration, should not be understood as the limitation to this patent；
Obviously, the above embodiment of the present invention be only to clearly illustrate example of the present invention, and not be pair
The restriction of embodiments of the present invention.For those of ordinary skill in the art, may be used also on the basis of the above description
To make other variations or changes in different ways.There is no necessity and possibility to exhaust all the enbodiments.It is all this
Made any modifications, equivalent replacements, and improvements etc., should be included in the claims in the present invention within the spirit and principle of invention
Protection scope within.
Claims (10)
1. a kind of based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that the generation fights network
Including generating network, analog encoding insertion module and differentiating network, comprising the following steps:
S1: by carrier image by the way that gradient JPEG conversion module can be transmitted, the coefficient in the corresponding DCT domain of the carrier image is obtained
Matrix；
S2: the corresponding DCT coefficient matrix of the carrier image is input in the generation network, generates correspondence by generating network
With carrier image is of the same size distorts probability matrix, the numerical value of described each position for distorting probability matrix represents the position
The probability that corresponding location of pixels DCT coefficient is tampered with；
S3: the probability matrix of distorting of generation is obtained with random noise matrix by analog encoding insertion module progress analog encoding
With carrier image is of the same size distorts matrix, the random noise matrix and the probability matrix of distorting are in the same size, distort
The value of element is ± 1 and 0 in matrix, and random noise matrix is for simulating random secret information；Matrix and carrier figure will be distorted
As the matrix addition in corresponding DCT domain, the DCT that carries close image corresponding with carrier image DCT coefficient matrix in DCT domain is obtained
Coefficient matrix；
S4: the DCT coefficient matrix of the close image of load in DCT domain is converted into airspace by that can transmit the JPEG conversion module of gradient
On the close image of load；
S5: being input to differentiation network for the S4 carrier image for obtaining carrying close image and S1, by differentiate network to carrier image with
It carries close image to classify, the error in classification for differentiating that network generates is fed back into differentiation as loss function, and by loss function
Network and generation network simultaneously carry out dual training；
S6: after training, trained generation network is mutually tied with traditional information coding module and JPEG conversion module
It closes, secret information is embedded into the DCT system of carrier image by the insertion cost value generated according to generation network in an adaptive way
The close image of load for being used for secret communication is generated in number, the insertion cost value is to distort probability conversion gained, specific conversion formula
Are as follows:
2. according to claim 1 based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that described
It includes image real time transfer and matrixing that gradient JPEG conversion module, which can be transmitted, and wherein image real time transfer includes Tensor
The batch of data is separated, the piecemeal processing of single batch, the merging of piecemeal and the merging of batch；Gradient JPEG can be transmitted
Matrixing in conversion module includes 2-D discrete cosine matrixing, anti-two-dimension discrete cosine transform and corresponding quantization
With inverse quantization processing, wherein inverse quantization does not use floor operation to ensure that gradient is effectively transmitted.
3. according to claim 1 based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that described
The convolutional neural networks that network is a full convolutional neural networks or a U-shaped structure in parallel of generating in step S2, wherein
Full convolutional neural networks in parallel are divided into four branches, and each branch includes three convolutional layers, the convolution kernel size of four branches
Respectively 2,3,4,5, each convolutional layer include 50 convolution kernels, are all made of band leakage rectification function (Leak ReLU) conduct
Activation primitive；Totally 16 groups of the convolutional neural networks of U-shaped structure, first 8 groups are convolutional layer group, each convolutional layer group includes convolutional layer
It criticizes, normalize layer and rectify function with leakage, the size of characteristics of image figure is reduced with the increase of the network number of plies；8 groups afterwards
For warp lamination group, function is rectified comprising warp lamination, batch normalization layer and with leakage, the size of characteristics of image figure is with net
The increase of network layers number and become larger, the characteristic pattern between first 8 layers image network layer corresponding with latter 8 layers all with series connection
(concatenate) mode is attached.
4. based on the JPEG domain image latent writing method for generating confrontation network according to claim 1, which is characterized in that step
Analog encoding insertion module is made of optimum code fitting function in S3, the optimal volume that wherein optimum code fitting function is fitted
Code mode may be expressed as:
In above formula, pijThe each DCT coefficient for making a living into network output corresponding distorts probability, nijFor random noise matrix, mijFor
Distort matrix；Optimum code fitting function is characterized in that keeping gradient propagable while can be fitted above-mentioned optimum code
Mode；The wherein expression formula of optimum code fitting function are as follows:
mij=-k × fs(λ×(pij-2×nij))+k×fs(λ×(pij-2×(1-nij)))+c
Wherein λ is that its value of zoom factor is bigger, and the fitting effect of optimum code fitting function is better, and gradient is propagated more not
Benefit, zoom factor λ value is smaller, is more conducive to the propagation of gradient, but fitting effect can be deteriorated, fsIndicate a kind of activation primitive,
Such as: Tanh, Sigmoid, ReLU, k are weight coefficient, and c is a constant biasing.
5. according to claim 1 based on the JPEG domain image latent writing method for generating confrontation network, it is characterised in that: step
Differentiation network described in S5 is the steganalysis convolutional neural networks based on DenseNet, including 1 pretreatment layer and 7
Dense convolutional layer group, wherein pretreatment layer includes the strong letter that 16 different types of high-pass filters are used to inhibit picture material
Number, the weak signal of enhancing steganography insertion modification, each dense convolutional layer group includes 4 convolutional layers and 1 pond layer, convolutional layer
All it is connected with its each convolutional layer for being located at front in the same dense convolutional layer group, the reuse of characteristic pattern is realized, using linear
Function is rectified as activation primitive, final class probability is obtained using full articulamentum and softmax function and is exported.
6. according to claim 1 based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that for
It generates network and differentiates the training that network uses gradient backpropagation to carry out parameter using the loss function that error in classification obtains, often
It updates the primary parameter for differentiating network and then updates the parameter for generating network twice, wherein being used for the loss function of more newly-generated network
Also comprising by distorting the loss letter that error generates between the obtained fitting insertion load capacity of probabilistic information entropy and targeted loads amount
Number.
7. according to claim 6 based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that described
Iteration update times are greater than 120,000 times.
8. according to claim 6 based on the JPEG domain image latent writing method for generating confrontation network, which is characterized in that described
Probabilistic information entropy calculation formula it is as follows:
In above formula, H is probabilistic information entropy, and h is the height of image, the width that w is image,Indicate positioned at (i, j) DCT coefficient into
The probability value that row+1 is distorted,Indicate that the DCT coefficient of position (i, j) carries out -1 probability value distorted,It indicates position (i, j)
DCT coefficient without the probability value distorted.
9. described in any item based on the JPEG domain image latent writing method for generating confrontation network, feature according to claim 6 to 8
It is, the loss function for differentiating network is the cross entropy of error in classification, specific formula for calculation are as follows:
In above formula, yiIt indicates to differentiate the probability that network divides the image into the i-th class, yi' indicate the actual class label of image；
The loss function for generating network is embedded in common group of loss function of load capacity by the loss function and fitting of differentiation network
At specific formula are as follows:
lG=α ×-lD+β×lp
Wherein lp=(H- γ × q)2For the loss function of fitting insertion load capacity, γ is non-zero exchange in carrier image DCT coefficient
The quantity of coefficient, q are the target insertion load capacity of setting, and α and β are the weight coefficient for adjusting two loss functions, special
Sign is that the selection of α and β setting value needs to weigh the fitting precision of steganography safety and insertion rate.
10. a kind of based on the jpeg image steganographic system for generating confrontation network, which is characterized in that including based on generation confrontation network
JPEG domain image latent writing method obtained generation network module, information coding module and JPEG conversion module, specific steps are such as
Under:
The airspace carrier image of input is converted to DCT coefficient matrix using JPEG conversion module, then passes through the generation network
Module generates distorting probability matrix and being converted to insertion cost value for corresponding DCT coefficient matrix, according to insertion cost value and reality
Secret information distorts matrix using information coding module generation size identical as image, by the matrix and initial carrier image pair
The DCT matrix phase adduction answered is input to JPEG conversion module and generates the close image of load for being used for practical secret communication.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910368809.2A CN110334805B (en) | 2019-05-05 | 2019-05-05 | JPEG domain image steganography method and system based on generation countermeasure network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910368809.2A CN110334805B (en) | 2019-05-05 | 2019-05-05 | JPEG domain image steganography method and system based on generation countermeasure network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110334805A true CN110334805A (en) | 2019-10-15 |
| CN110334805B CN110334805B (en) | 2022-10-25 |
Family
ID=68139373
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910368809.2A Expired - Fee Related CN110334805B (en) | 2019-05-05 | 2019-05-05 | JPEG domain image steganography method and system based on generation countermeasure network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110334805B (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852316A (en) * | 2019-11-07 | 2020-02-28 | 中山大学 | Image tampering detection and positioning method adopting convolution network with dense structure |
| CN110879894A (en) * | 2019-11-11 | 2020-03-13 | 陕西师范大学 | Image encryption and decryption method based on lazy scale transformation and random layered fusion |
| CN111222583A (en) * | 2020-01-15 | 2020-06-02 | 北京中科研究院 | Image steganalysis method based on confrontation training and key path extraction |
| CN111263165A (en) * | 2019-12-27 | 2020-06-09 | 郑州师范学院 | Embedding method and extracting method of HEVC video steganography information |
| CN111327785A (en) * | 2020-01-15 | 2020-06-23 | 郑州埃文计算机科技有限公司 | Information steganography communication method based on automatic image construction of countermeasure generation network |
| CN111681154A (en) * | 2020-06-09 | 2020-09-18 | 湖南大学 | Color image steganography distortion function design method based on generation countermeasure network |
| CN111768326A (en) * | 2020-04-03 | 2020-10-13 | 南京信息工程大学 | High-capacity data protection method based on GAN amplification image foreground object |
| CN111882476A (en) * | 2020-07-17 | 2020-11-03 | 广州大学 | Image steganography method for automatically learning embedded cost based on deep reinforcement learning |
| CN112199693A (en) * | 2020-09-30 | 2021-01-08 | 东南数字经济发展研究院 | Steganography method based on cartoon image |
| CN112468677A (en) * | 2020-11-27 | 2021-03-09 | 中国农业银行股份有限公司 | Data encryption and decryption method and device and electronic equipment |
| CN112634118A (en) * | 2020-12-24 | 2021-04-09 | 中国科学技术大学 | Batch-resistant steganography method |
| CN112767226A (en) * | 2021-01-15 | 2021-05-07 | 南京信息工程大学 | Image steganography method and system based on GAN network structure automatic learning distortion |
| CN113077377A (en) * | 2021-05-13 | 2021-07-06 | 海南大学 | Color image steganography method based on generation countermeasure network |
| CN113076549A (en) * | 2021-04-08 | 2021-07-06 | 上海电力大学 | Confrontation network image steganography method based on novel U-Net structure generator |
| CN114051082A (en) * | 2021-10-19 | 2022-02-15 | 河南师范大学 | Steganography detection feature selection method and device based on distortion degree and information gain ratio |
| CN114119334A (en) * | 2021-12-06 | 2022-03-01 | 广州大学 | Image steganography method for automatically learning embedded cost based on attention mechanism |
| CN114900586A (en) * | 2022-04-28 | 2022-08-12 | 中国人民武装警察部队工程大学 | Information steganography method and device based on DCGAN |
| CN117939027A (en) * | 2024-03-21 | 2024-04-26 | 南京信息工程大学 | JPEG image steganography method and system based on DCT feature extraction |
Citations (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080175429A1 (en) * | 2007-01-19 | 2008-07-24 | New Jersey Institute Of Technology | Method and apparatus for steganalysis for texture images |
| US20080193031A1 (en) * | 2007-02-09 | 2008-08-14 | New Jersey Institute Of Technology | Method and apparatus for a natural image model based approach to image/splicing/tampering detection |
| US20130208941A1 (en) * | 2012-02-01 | 2013-08-15 | Qingzhong Liu | Steganalysis with neighboring joint density |
| CN103914839A (en) * | 2014-03-27 | 2014-07-09 | 中山大学 | Image stitching and tampering detection method and device based on steganalysis |
| CN106097240A (en) * | 2016-06-13 | 2016-11-09 | 天津大学 | A kind of color image-adaptive steganography method |
| CN106530200A (en) * | 2016-10-23 | 2017-03-22 | 深圳大学 | Deep-learning-model-based steganography image detection method and system |
| CN106791872A (en) * | 2016-11-18 | 2017-05-31 | 南京邮电大学 | Information concealing method based on SVD |
| CN106920206A (en) * | 2017-03-16 | 2017-07-04 | 广州大学 | A kind of steganalysis method based on confrontation neutral net |
| CN107563155A (en) * | 2017-08-08 | 2018-01-09 | 中国科学院信息工程研究所 | A kind of safe steganography method and device based on generation confrontation network |
| CN107908969A (en) * | 2017-11-21 | 2018-04-13 | 中山大学 | A kind of adaptive steganography method of jpeg image based on Spatial characteristic |
| CN107945204A (en) * | 2017-10-27 | 2018-04-20 | 西安电子科技大学 | A kind of Pixel-level portrait based on generation confrontation network scratches drawing method |
| CN108024028A (en) * | 2017-12-12 | 2018-05-11 | 中国人民解放军战略支援部队信息工程大学 | Jpeg image robust steganography method based on serial concatenation of codes and DCT coefficient relativeness |
| CN108346125A (en) * | 2018-03-15 | 2018-07-31 | 中山大学 | A kind of spatial domain picture steganography method and system based on generation confrontation network |
| CN108416725A (en) * | 2018-03-13 | 2018-08-17 | 中山大学 | A kind of asymmetric JPEG steganography methods based on elimination blocking artifact |
| CN108921764A (en) * | 2018-03-15 | 2018-11-30 | 中山大学 | A kind of image latent writing method and system based on generation confrontation network |
| CN108923922A (en) * | 2018-07-26 | 2018-11-30 | 北京工商大学 | A kind of text steganography method based on generation confrontation network |
| CN108961137A (en) * | 2018-07-12 | 2018-12-07 | 中山大学 | A kind of image latent writing analysis method and system based on convolutional neural networks |
| CN109040763A (en) * | 2018-08-07 | 2018-12-18 | 北京飞搜科技有限公司 | A kind of method for compressing image and system based on production confrontation network |
| CN109214973A (en) * | 2018-08-24 | 2019-01-15 | 中国科学技术大学 | For the confrontation safety barrier generation method of steganalysis neural network |
| CN109587372A (en) * | 2018-12-11 | 2019-04-05 | 北京邮电大学 | A kind of invisible image latent writing art based on generation confrontation network |
| CN109658322A (en) * | 2018-12-11 | 2019-04-19 | 宁波大学 | A kind of large capacity image latent writing method and secret information extraction method |
-
2019
- 2019-05-05 CN CN201910368809.2A patent/CN110334805B/en not_active Expired - Fee Related
Patent Citations (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080175429A1 (en) * | 2007-01-19 | 2008-07-24 | New Jersey Institute Of Technology | Method and apparatus for steganalysis for texture images |
| US20080193031A1 (en) * | 2007-02-09 | 2008-08-14 | New Jersey Institute Of Technology | Method and apparatus for a natural image model based approach to image/splicing/tampering detection |
| US20130208941A1 (en) * | 2012-02-01 | 2013-08-15 | Qingzhong Liu | Steganalysis with neighboring joint density |
| CN103914839A (en) * | 2014-03-27 | 2014-07-09 | 中山大学 | Image stitching and tampering detection method and device based on steganalysis |
| CN106097240A (en) * | 2016-06-13 | 2016-11-09 | 天津大学 | A kind of color image-adaptive steganography method |
| CN106530200A (en) * | 2016-10-23 | 2017-03-22 | 深圳大学 | Deep-learning-model-based steganography image detection method and system |
| CN106791872A (en) * | 2016-11-18 | 2017-05-31 | 南京邮电大学 | Information concealing method based on SVD |
| CN106920206A (en) * | 2017-03-16 | 2017-07-04 | 广州大学 | A kind of steganalysis method based on confrontation neutral net |
| CN107563155A (en) * | 2017-08-08 | 2018-01-09 | 中国科学院信息工程研究所 | A kind of safe steganography method and device based on generation confrontation network |
| CN107945204A (en) * | 2017-10-27 | 2018-04-20 | 西安电子科技大学 | A kind of Pixel-level portrait based on generation confrontation network scratches drawing method |
| CN107908969A (en) * | 2017-11-21 | 2018-04-13 | 中山大学 | A kind of adaptive steganography method of jpeg image based on Spatial characteristic |
| CN108024028A (en) * | 2017-12-12 | 2018-05-11 | 中国人民解放军战略支援部队信息工程大学 | Jpeg image robust steganography method based on serial concatenation of codes and DCT coefficient relativeness |
| CN108416725A (en) * | 2018-03-13 | 2018-08-17 | 中山大学 | A kind of asymmetric JPEG steganography methods based on elimination blocking artifact |
| CN108346125A (en) * | 2018-03-15 | 2018-07-31 | 中山大学 | A kind of spatial domain picture steganography method and system based on generation confrontation network |
| CN108921764A (en) * | 2018-03-15 | 2018-11-30 | 中山大学 | A kind of image latent writing method and system based on generation confrontation network |
| CN108961137A (en) * | 2018-07-12 | 2018-12-07 | 中山大学 | A kind of image latent writing analysis method and system based on convolutional neural networks |
| CN108923922A (en) * | 2018-07-26 | 2018-11-30 | 北京工商大学 | A kind of text steganography method based on generation confrontation network |
| CN109040763A (en) * | 2018-08-07 | 2018-12-18 | 北京飞搜科技有限公司 | A kind of method for compressing image and system based on production confrontation network |
| CN109214973A (en) * | 2018-08-24 | 2019-01-15 | 中国科学技术大学 | For the confrontation safety barrier generation method of steganalysis neural network |
| CN109587372A (en) * | 2018-12-11 | 2019-04-05 | 北京邮电大学 | A kind of invisible image latent writing art based on generation confrontation network |
| CN109658322A (en) * | 2018-12-11 | 2019-04-19 | 宁波大学 | A kind of large capacity image latent writing method and secret information extraction method |
Non-Patent Citations (2)
| Title |
|---|
| HANQI ZI ET AL.: "Steganography with Convincing Normal Image from A Joint Generative Adversarial Framework", 《APSIPA ANNUAL SUMMIT AND CONFERENCE 2018》 * |
| WEIXUAN TANG ET AL.: "Automatic Steganographic Distortion Learning Using a Generative Adversarial Network", 《IEEE SIGNAL PROCESSING LETTERS》 * |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110852316A (en) * | 2019-11-07 | 2020-02-28 | 中山大学 | Image tampering detection and positioning method adopting convolution network with dense structure |
| CN110852316B (en) * | 2019-11-07 | 2023-04-18 | 中山大学 | Image tampering detection and positioning method adopting convolution network with dense structure |
| CN110879894A (en) * | 2019-11-11 | 2020-03-13 | 陕西师范大学 | Image encryption and decryption method based on lazy scale transformation and random layered fusion |
| CN111263165A (en) * | 2019-12-27 | 2020-06-09 | 郑州师范学院 | Embedding method and extracting method of HEVC video steganography information |
| CN111263165B (en) * | 2019-12-27 | 2022-07-01 | 郑州师范学院 | Embedding method and extracting method of HEVC video steganography information |
| CN111327785A (en) * | 2020-01-15 | 2020-06-23 | 郑州埃文计算机科技有限公司 | Information steganography communication method based on automatic image construction of countermeasure generation network |
| CN111222583B (en) * | 2020-01-15 | 2023-11-17 | 北京中科研究院 | Image steganalysis method based on countermeasure training and critical path extraction |
| CN111327785B (en) * | 2020-01-15 | 2022-04-15 | 郑州埃文计算机科技有限公司 | Information steganography communication method based on automatic image construction of countermeasure generation network |
| CN111222583A (en) * | 2020-01-15 | 2020-06-02 | 北京中科研究院 | Image steganalysis method based on confrontation training and key path extraction |
| CN111768326A (en) * | 2020-04-03 | 2020-10-13 | 南京信息工程大学 | High-capacity data protection method based on GAN amplification image foreground object |
| CN111768326B (en) * | 2020-04-03 | 2023-08-25 | 南京信息工程大学 | High-capacity data protection method based on GAN (gas-insulated gate bipolar transistor) amplified image foreground object |
| CN111681154A (en) * | 2020-06-09 | 2020-09-18 | 湖南大学 | Color image steganography distortion function design method based on generation countermeasure network |
| CN111681154B (en) * | 2020-06-09 | 2022-06-21 | 湖南大学 | Color image steganography distortion function design method based on generation countermeasure network |
| CN111882476B (en) * | 2020-07-17 | 2023-07-07 | 广州大学 | Image steganography method for automatic learning embedding cost based on deep reinforcement learning |
| CN111882476A (en) * | 2020-07-17 | 2020-11-03 | 广州大学 | Image steganography method for automatically learning embedded cost based on deep reinforcement learning |
| CN112199693A (en) * | 2020-09-30 | 2021-01-08 | 东南数字经济发展研究院 | Steganography method based on cartoon image |
| CN112468677B (en) * | 2020-11-27 | 2022-09-27 | 中国农业银行股份有限公司 | Data encryption and decryption method and device and electronic equipment |
| CN112468677A (en) * | 2020-11-27 | 2021-03-09 | 中国农业银行股份有限公司 | Data encryption and decryption method and device and electronic equipment |
| CN112634118A (en) * | 2020-12-24 | 2021-04-09 | 中国科学技术大学 | Batch-resistant steganography method |
| CN112634118B (en) * | 2020-12-24 | 2022-09-30 | 中国科学技术大学 | Anti-batch steganography method |
| CN112767226A (en) * | 2021-01-15 | 2021-05-07 | 南京信息工程大学 | Image steganography method and system based on GAN network structure automatic learning distortion |
| CN112767226B (en) * | 2021-01-15 | 2023-09-12 | 南京信息工程大学 | Image steganography method and system based on automatic learning distortion of GAN network structure |
| CN113076549B (en) * | 2021-04-08 | 2023-05-12 | 上海电力大学 | Novel U-Net structure generator-based countermeasures network image steganography method |
| CN113076549A (en) * | 2021-04-08 | 2021-07-06 | 上海电力大学 | Confrontation network image steganography method based on novel U-Net structure generator |
| CN113077377A (en) * | 2021-05-13 | 2021-07-06 | 海南大学 | Color image steganography method based on generation countermeasure network |
| CN114051082A (en) * | 2021-10-19 | 2022-02-15 | 河南师范大学 | Steganography detection feature selection method and device based on distortion degree and information gain ratio |
| CN114051082B (en) * | 2021-10-19 | 2023-10-27 | 河南师范大学 | Steganography detection feature selection method and device based on distortion degree and information gain ratio |
| CN114119334A (en) * | 2021-12-06 | 2022-03-01 | 广州大学 | Image steganography method for automatically learning embedded cost based on attention mechanism |
| CN114900586A (en) * | 2022-04-28 | 2022-08-12 | 中国人民武装警察部队工程大学 | Information steganography method and device based on DCGAN |
| CN114900586B (en) * | 2022-04-28 | 2024-04-16 | 中国人民武装警察部队工程大学 | Information steganography method and device based on DCGAN |
| CN117939027A (en) * | 2024-03-21 | 2024-04-26 | 南京信息工程大学 | JPEG image steganography method and system based on DCT feature extraction |
| CN117939027B (en) * | 2024-03-21 | 2024-06-07 | 南京信息工程大学 | JPEG image steganography method and system based on DCT feature extraction |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110334805B (en) | 2022-10-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110334805A (en) | A kind of JPEG domain image latent writing method and system based on generation confrontation network | |
| CN112634117B (en) | End-to-end JPEG domain image steganography method based on generation of countermeasure network | |
| CN112288658B (en) | Underwater image enhancement method based on multi-residual joint learning | |
| CN111275647A (en) | Underwater image restoration method based on cyclic generation countermeasure network | |
| CN103927531B (en) | It is a kind of based on local binary and the face identification method of particle group optimizing BP neural network | |
| CN108346125A (en) | A kind of spatial domain picture steganography method and system based on generation confrontation network | |
| CN113222800A (en) | Robust image watermark embedding and extracting method and system based on deep learning | |
| CN107977932A (en) | It is a kind of based on can differentiate attribute constraint generation confrontation network face image super-resolution reconstruction method | |
| CN109300090A (en) | A kind of single image to the fog method generating network based on sub-pix and condition confrontation | |
| CN109658322B (en) | A kind of large capacity image latent writing method and secret information extraction method | |
| CN109598732B (en) | Medical image segmentation method based on three-dimensional space weighting | |
| CN111681154B (en) | Color image steganography distortion function design method based on generation countermeasure network | |
| CN112528830A (en) | Lightweight CNN mask face pose classification method combined with transfer learning | |
| CN112287973A (en) | Digital image countermeasure sample defense method based on truncated singular value and pixel interpolation | |
| CN108710896A (en) | The field learning method of learning network is fought based on production | |
| CN110490265A (en) | A kind of image latent writing analysis method based on two-way convolution sum Fusion Features | |
| CN112132737B (en) | Image robust steganography method without reference generation | |
| CN113076549B (en) | Novel U-Net structure generator-based countermeasures network image steganography method | |
| CN114339258A (en) | Information steganography method and device based on video carrier | |
| CN113938254A (en) | Attention mechanism-based layered source-channel joint coding transmission system and transmission method thereof | |
| CN107578365B (en) | Wavelet digital watermark embedding and extracting method based on quantum weed optimizing mechanism | |
| Kumari et al. | Flower pollination-based K-means algorithm for medical image compression | |
| CN114596229A (en) | Image noise reduction method and system | |
| CN109087114A (en) | A kind of red wine packaging anti-counterfei waterprint intelligent measurement method for tracing | |
| CN110223224A (en) | A kind of Image Super-resolution realization algorithm based on information filtering network |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20221025 |