CN115086503B - Information hiding method, device, equipment and storage medium - Google Patents
Information hiding method, device, equipment and storage medium Download PDFInfo
- Publication number
- CN115086503B CN115086503B CN202210583727.1A CN202210583727A CN115086503B CN 115086503 B CN115086503 B CN 115086503B CN 202210583727 A CN202210583727 A CN 202210583727A CN 115086503 B CN115086503 B CN 115086503B
- Authority
- CN
- China
- Prior art keywords
- information
- carrier image
- hidden
- model
- image
- 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.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N1/00—Scanning, transmission or reproduction of documents or the like, e.g. facsimile transmission; Details thereof
- H04N1/32—Circuits or arrangements for control or supervision between transmitter and receiver or between image input and image output device, e.g. between a still-image camera and its memory or between a still-image camera and a printer device
- H04N1/32101—Display, printing, storage or transmission of additional information, e.g. ID code, date and time or title
- H04N1/32144—Display, printing, storage or transmission of additional information, e.g. ID code, date and time or title embedded in the image data, i.e. enclosed or integrated in the image, e.g. watermark, super-imposed logo or stamp
- H04N1/32149—Methods relating to embedding, encoding, decoding, detection or retrieval operations
- H04N1/32267—Methods relating to embedding, encoding, decoding, detection or retrieval operations combined with processing of the image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T1/00—General purpose image data processing
- G06T1/0021—Image watermarking
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Editing Of Facsimile Originals (AREA)
- Image Processing (AREA)
Abstract
The invention provides an information hiding method, an information hiding device, information hiding equipment and a storage medium, wherein the information hiding method comprises the following steps: acquiring semantic information of an original carrier image; embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information; the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image. In the method of the embodiment of the invention, the embedding position of the information to be hidden is guided by using the semantic information, so that the imperceptibility is better.
Description
Technical Field
The present invention relates to the field of computer technologies, and in particular, to an information hiding method, an information hiding device, and a storage medium.
Background
The information hiding technology is a technology of embedding some information data with secret value into a common image which does not carry secret information through a technical means, so as to hide the secret data. Different from the true meaning of hidden data in traditional cryptography, the purpose of the information hiding technology is to hide the existence of information, and then recover the hidden information through a corresponding algorithm, thereby realizing the purpose of secret information transmission. Information hiding can be largely classified into character string information, image information, and audio information for different hidden objects. An image that does not carry confidential information is referred to as a carrier image, and information or an image recovered from the carrier image is referred to as secret information or a secret image. Information hiding technology is widely used in various fields, such as copyright protection, information authentication, and the like.
Information hiding algorithms are generally evaluated from three aspects, embedding capacity, robustness, and imperceptibility, respectively. The embedded capacity is the ratio of the length of the embedded secret information to the carrier image, and the larger the ratio is, the larger the embedded capacity is, and the stronger the information carrying capacity of the algorithm is. Robustness refers to the ability to extract secret information, the ultimate goal of information hiding is to pass information in secret, and if there is a large error in the extracted information, even if the information that is desired to be communicated cannot be recovered, the robustness of the algorithm is very poor. Imperceptibility represents the difference between an image after information is embedded (also called a carrier image) and a carrier image without the information is originally embedded, the smaller the difference is, the smaller the influence of information embedding on the original image is, the better the imperceptibility is, the harder the carrier image is distinguished from the carrier image, and the better the information hiding effect is.
With the development of deep learning technology, an information hiding algorithm based on deep learning is continuously proposed in recent years, but the information hiding algorithm based on deep learning is poor in perceptibility, and particularly a flat area of a carrier image. Therefore, it is highly desirable for those skilled in the art to implement an information hiding method with better imperceptibility.
Disclosure of Invention
The invention provides an information hiding method, an information hiding device, information hiding equipment and a storage medium, which are used for solving the defect of poor imperceptibility of information hiding in the prior art and realizing the information hiding method with good imperceptibility.
The invention provides an information hiding method, which comprises the following steps:
acquiring semantic information of an original carrier image;
embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
According to the information hiding method provided by the invention, the embedding model comprises a deconvolution module, an overlapping layer and a convolution module, the information to be hidden is embedded into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information, and the information hiding method comprises the following steps:
performing up-sampling processing on the original carrier image by using a deconvolution module of the embedded model based on the semantic information to obtain first carrier image information; the first carrier image information has the same size as the information to be hidden obtained by preprocessing;
And superposing the information to be hidden, the first carrier image information and the original carrier image which are obtained by preprocessing on the basis of the semantic information, and carrying out downsampling on the superposed information by utilizing a convolution module of the embedded model on the basis of the semantic information to obtain the target carrier image.
According to the information hiding method provided by the invention, before the information to be hidden, the first carrier image information and the original carrier image obtained by the preprocessing are subjected to superposition processing based on the semantic information, the information hiding method further comprises the following steps:
preprocessing the information to be hidden by using an information preprocessing model to obtain the preprocessed information to be hidden;
the information preprocessing model comprises the following steps: a linear connection layer, a reconstruction layer, and a convolution module; the linear connection layer is used for expanding the information to be hidden to obtain information with the length of N, the reconstruction layer is used for reconstructing the information with the length of N to obtain vectors with the same number of dimensions as the original carrier image, the convolution module is used for carrying out convolution processing on the vectors output by the reconstruction layer to obtain the preprocessed information to be hidden, and N is an integer larger than 1.
According to the information hiding method provided by the invention, the extraction model comprises the following steps: deconvolution module, convolution module, and linear connection layer, the method further comprising:
performing up-sampling processing on the target carrier image by using a deconvolution module of the extraction model to obtain second carrier image information;
performing downsampling processing on the second carrier image information by using a convolution module of the extraction model to obtain third carrier image information;
and processing the third carrier image information by using the linear connection layer of the extraction model to obtain extraction information.
According to the information hiding method provided by the invention, the semantic information comprises at least one of the following: edge information, gradient information, texture information.
According to the information hiding method provided by the invention, the semantic information comprises multi-scale semantic information of the original carrier image.
According to the information hiding method provided by the invention, the information to be hidden comprises information of K different tasks, wherein K is an integer greater than 1.
According to the information hiding method provided by the invention, the method further comprises the following steps:
based on the original carrier image and the target carrier image, acquiring difference information of the original carrier image and the target carrier image by utilizing a discrimination model; the difference information is represented using probabilities.
According to the information hiding method provided by the invention, the method further comprises the following steps:
optimizing at least one model based on the error information;
the at least one model includes at least one of: the system comprises an embedded model, an information preprocessing module, a judging model and an information extracting model;
the error information includes at least one of: error between original carrier image and target carrier image, error between information to be hidden and extracted information, cross entropy of discrimination model and error between semantic information and difference image of original carrier image; the difference image is a difference image between the original carrier image and the target carrier image.
According to the information hiding method provided by the invention, the convolution module comprises the following steps: a convolutional layer, a regularization layer, and an activation layer.
The deconvolution module includes: deconvolution, regularization and activation layers.
According to the information hiding method provided by the invention, the embedded model is obtained based on training data, and the training data comprises: information to be hidden, position of information to be hidden, carrier image.
The invention also provides an information hiding device, comprising:
The acquisition module is used for acquiring semantic information of the original carrier image;
the processing module is used for embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
The invention also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the information hiding method as described above when executing the program.
The present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the information hiding method as described in any one of the above.
The invention also provides a computer program product comprising a computer program which, when executed by a processor, implements a method of hiding information as described in any one of the above.
According to the information hiding method, device, equipment and storage medium, the embedding position of the information to be hidden is guided by adding semantic information, the information to be hidden is embedded into the original carrier image by utilizing the trained embedding model to obtain the target carrier image, information embedding of a flat area is reduced, and as the human eyes have strong perception of tiny change of the flat area and have lower perceptibility of change of other areas, the imperceptibility is improved through the scheme.
Drawings
In order to more clearly illustrate the invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the invention, and other drawings can be obtained according to the drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic flow chart of an information hiding method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a prior art information hiding scheme;
FIG. 3 is a schematic diagram of an information hiding method according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a second embodiment of an information hiding method according to the present invention;
FIG. 5 is a test result of the information hiding method provided by the embodiment of the present invention on a div2k dataset;
FIG. 6 is an illustration of ablation experimental results provided by an embodiment of the present invention;
FIG. 7 is a schematic diagram of a detection result of an image of a target carrier according to an embodiment of the present invention;
FIG. 8 is a schematic diagram of an embedding position result of an information hiding method without edge information guidance according to an embodiment of the present invention;
FIG. 9 is a schematic diagram of an embedding position result of an information hiding method under the guidance of added edge information according to an embodiment of the present invention;
FIG. 10 is a comparison result of the information hiding method according to the embodiment of the present invention with other algorithms;
FIG. 11 is a schematic diagram of an information hiding apparatus according to an embodiment of the present invention;
fig. 12 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
According to the method provided by the embodiment of the invention, the information to be hidden is embedded into the original carrier image by adding the guidance of the image semantic information and utilizing the trained embedding model to obtain the target carrier image, so that the information embedding of the flat area is reduced, and the imperceptibility can be improved.
The following describes the technical solution of the embodiment of the present invention in detail with reference to fig. 1 to 12. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.
Fig. 1 is a flow chart of an information hiding method according to an embodiment of the present invention. As shown in fig. 1 to 3, the method provided in this embodiment includes:
step 101, acquiring semantic information of an original carrier image;
specifically, the purpose of semantic information guiding is to enable the model to realize smaller change on a flat place of an image, and more information is embedded into semantic information such as edges, textures, gradients and the like. Since the human eye is very sensitive to small changes in the flat area according to the basic principle of human vision, the embedding model should be guided to embed as little information to be hidden into the flat area of the image as possible in order to increase the imperceptibility of the algorithm.
Step 102, embedding information to be hidden into an original carrier image to obtain a target carrier image by utilizing a trained embedding model based on semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
Specifically, the embedded model may be obtained by training a pre-established machine learning model based on training data, where the training data includes: information to be hidden, the location of the information to be hidden (e.g., the location corresponding to the semantic information of the carrier image), the carrier image.
The machine learning model may be a model built based on a neural network algorithm or the like.
The input data of the embedded model are: according to the method, the embedding model embeds the information to be hidden into the original carrier image under the guidance of the semantic information, namely the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image, so that the information to be hidden is hidden at the position corresponding to the semantic information of the carrier image and the position nearby corresponding to the semantic information as much as possible, the embedding of a flat area is reduced, and the imperceptibility is improved.
According to the method, the embedding position of the information to be hidden is guided by adding semantic information, the information to be hidden is embedded into the original carrier image by utilizing the trained embedding model to obtain the target carrier image, the information embedding of the flat area is reduced, and the human eyes have strong perception of tiny change of the flat area and have lower perceptibility of change of other areas, so that the imperceptibility is improved through the scheme.
Optionally, the semantic information includes at least one of: edge information, gradient information, texture information.
As shown in fig. 3, the purpose of edge information guidance is to enable the model to embed more information into the edge pixels of the image, while achieving less modification where the image is flat. Since the human eye has a low perceptibility of changes at the boundary of the edges and a strong perception of small changes in the flat area according to the basic principle of human vision, the embedding model should be guided to embed as much information to be hidden into the edge pixels of the image as possible in order to improve the imperceptibility of the algorithm.
The gradient information and the texture information are relatively more complex than the change of the flat area, and the human eye has lower perceptibility of the change, so that the information embedding of the flat area is reduced based on the guidance of the gradient information and the texture information, and the imperceptibility is improved.
Specifically, a multi-scale deep learning edge detection algorithm can be used to obtain multi-scale edge information, and the multi-scale edge information can be guided to not only achieve the purpose of embedding more information into the image edge, but also guide the embedding model to embed a small number of pixels into positions near the edge pixels, so that the embedding model is less embedded into a flat area.
The gradient information and texture information are similar and will not be described in detail here.
For example, the carrier image is 3×h×w, and edge information can be extracted based on h×w,2h×2w,3h×3w, and/or 4h×4w, where H and W represent the height and width of the image, respectively.
Before introducing the model, firstly, introducing structures used for multiple times in the two embodiments of the invention, namely a convolution ConvBNRerlu module, comprising a convolution layer, a batch regularization layer and an activation layer (for example, using a ReLU activation function) with different step sizes, wherein the asynchronous length is used for controlling convolution operations with the same size and up-down sampling convolution operations; the operation of the batch regularization layer can facilitate parameter searching, so that the selection of the super parameters by the neural network is more stable, and the model has nonlinear fitting capability due to the action of the activation layer. The second is the deconvolution TransConvBNRerlu structure comprising: deconvolution layer, batch regularization layer, and activation layer. The deconvolution layer realizes the up-sampling function of the carrier image, makes more image cavity pixels, facilitates the later embedding of information, and the operation of the batch regularization layer can make the parameter searching problem easier, so that the neural network is more stable in the selection of super parameters, and the model has nonlinear fitting capability due to the action of the activation layer.
Optionally, the embedding model includes a deconvolution module, an overlap layer, and a convolution module, and step 102 may be implemented by:
the method comprises the steps that an deconvolution module of an embedded model is utilized to carry out up-sampling processing on an original carrier image based on semantic information, so that first carrier image information is obtained; the first carrier image information has the same size as the information to be hidden obtained by preprocessing;
and superposing the information to be hidden, the first carrier image information and the original carrier image which are obtained by preprocessing on the basis of the semantic information, and carrying out downsampling on the superposed information by utilizing a convolution module of the embedded model on the basis of the semantic information to obtain a target carrier image.
Optionally, before the superimposing process, the method may further include:
preprocessing information to be hidden by using an information preprocessing model;
the information preprocessing model comprises the following steps: a linear connection layer, a reconstruction layer, a convolutional convbnrinl module; the linear connection layer is used for expanding information to be hidden to obtain information with the length of N, the reconstruction layer is used for reconstructing the information with the length of N to obtain vectors with the same number of dimensions as the original carrier image, and the convolution module is used for carrying out convolution processing on the vectors output by the reconstruction layer.
Specifically, assuming that the size of the original carrier image is 3×h×w, for example, H is 128, W is 128, K is greater than 0 for K tasks, each task includes information with a length L, and the information is first converted into information output with a length N by a linear connection layer, and then the size is changed by expansion and reconstruction, and the information is repeated until the same size as the image information sampled on the original carrier image becomes 3×4h× 4w (information with the same number of dimensions as the original carrier image), where the information preprocessing may use a tile repetition manner for K tasks, i.e., the information with a length of 3 kx4h× 4w is obtained. A convolutional convbnrinl module is then connected to obtain information with a first dimension M, m×4h×4w, for example, M is 32. For example, the step sizes of the convolution layers in the convolution module are all 1, and the same-size convolution is adopted.
In the related art, single information repetition is adopted, the obtained output is (n+3+32) ×4h×4w, according to practical situations, N is generally an information amount exceeding 100, and the output of the information preprocessing of the embodiment of the present invention is 32×4h×4w information, so that the capacity of embedded information is greatly improved, and the calculation amount of a model can be reduced.
As shown in fig. 4, the semantic information in fig. 4 is illustrated by taking edge information as an example, the linear connection layer is used to expand the information to be hidden to obtain information with a length of N, if the task number is K, the information with a size of k×n is obtained, the reconstruction layer is used to reconstruct the information with a length of N to obtain a vector with a size of 3k×4hx4w, the size of the vector in the channel dimension is K times the number of channels of the original carrier image, the size of the other dimensions is the same as the size of the corresponding dimension of the first carrier image information after the original carrier image is subjected to the up-sampling processing, finally, the vector output by the reconstruction layer is subjected to the convolution processing by the convolution module to obtain the preprocessed information to be hidden, and the size of the vector is 32×4hx4w, which is the same as the size of the carrier image information after the original carrier image is subjected to the up-sampling processing.
At present, the processing of the original carrier image is generally the processing of the same scale or downsampling, the pixel positions of the original carrier image which can be used for embedding are not increased, and therefore the embedded information capacity is low. In the embodiment of the invention, the deconvolution is adopted to up-sample the original carrier image, so that more pixel gaps are manufactured, and more pixel spaces are provided for embedding secret information. In addition, the method of the embodiment of the invention has the information hiding capability of processing multitasking. According to the experimental results, the network of the embodiment of the invention is obviously superior to other information hiding algorithms in capacity and robustness.
In the implementation of step 102, the deconvolution transconvbnreiu module is first used to upsample the carrier image, where the original 3-channel image is changed to an M-channel image, e.g., m=32. And then, carrying out superposition processing on the information to be hidden, the up-sampled carrier image information and the original carrier image obtained by preprocessing in the channel dimension, sending the superposed information to a convolution ConvBNRerlu module for carrying out down-sampling operation, and carrying out down-sampling to the size of the original carrier image similar to the up-sampling operation of the carrier image, wherein the number of channels is reduced to 3, so as to obtain a three-channel image embedded with information, namely a target carrier image, and the size of the target carrier image is the same as that of the original carrier image. For example, the downsampling process is performed by the three convolution modules in fig. 4.
Optionally, the size of the first dimension of the superimposed information is the sum of the size of the first dimension of the information to be hidden, the first carrier image information and the original carrier image obtained by preprocessing; the second dimension and the third dimension of the superimposed information are respectively the same as the second dimension and the third dimension of the first carrier image information.
As shown in fig. 4, the size of the information to be hidden obtained by the preprocessing is 32×4h×4w, the size of the up-sampled carrier image information is 32×4h×4w, the original carrier image is 3×4h×4w after 4 repetitions, the superimposition processing is performed in the channel dimension, and the size of the superimposed information is 67×4h×4w. The embedded model in fig. 4 is processed by the deconvolution module twice and the convolution module three times.
In the above embodiment, the up-sampling process is performed on the carrier image by the deconvolution module, so that the pixel positions where the original carrier image can be used for embedding are increased, and thus the capacity of embedded information is improved.
Optionally, extracting the model includes: deconvolution module, convolution module, and linear connection layer, the method further comprising:
up-sampling the target carrier image by using a deconvolution module of the extraction model to obtain second carrier image information;
Performing downsampling processing on the second carrier image information by using a convolution module of the extraction model to obtain third carrier image information;
and processing the third carrier image information by using the linear connection layer of the extraction model to obtain extraction information.
Specifically, for the target carrier image in which information is embedded, the information extraction process is similar to the embedding process. As shown in fig. 4, the deconvolution TransConvBNRelu module is used to upsample the information embedded target carrier image, the 2 deconvolutions were followed by a change from original bx3×128×128 (bx3×hxw) to bx32×512×512 (bx32×4hx4W), then the processing using the 2 convolutions convbnrinlu module was again reduced in size to bx3 k×h×w, and then extracting information channels by using a linear connection layer, reducing the number of channels to 3K by using the same size as the convolution ConvBNRerlu module, wherein each 3 channels are information of one task, then flattening the information of the 3 channels (R/G/B) to reduce the size of KxN, and finally reducing the information of KxN to the length KxL of the original information through the linear connection layer.
In the embodiment, the extraction model is utilized to process the target carrier image to obtain the extraction information, so that the efficiency is high, and the extraction information can be used for optimizing the model.
Optionally, the method further comprises:
based on the original carrier image and the target carrier image, acquiring difference information of the original carrier image and the target carrier image by utilizing a discrimination model; the difference information is represented by probabilities.
Specifically, the discrimination model is mainly used for supervising the difference between the target carrier image and the original carrier image after the information is embedded, so that the discrimination model cannot distinguish whether the information is embedded or not, namely, the smaller the difference between the target carrier image and the original carrier image is, the better the difference is. For example, the larger the probability, the smaller the difference between the original carrier image and the target carrier image.
Optionally, the discriminant model comprises: the device comprises a convolution module, a pooling layer and a linear connection layer; for example, as shown in fig. 4, the input original carrier image and target carrier image pass through three step-1 convolved convbnrinlu modules, a pooling layer (e.g., an adaptive averaging pooling layer), and a linear connection layer. The input information is bx 3×128×128 (i.e., bx 3×h×w), the output is b×1, and B is the amount of data input per batch.
In the embodiment, the imperceptibility is improved by utilizing the difference between the target carrier image and the original carrier image after the information is embedded by the discrimination model.
Optionally, the method further comprises:
optimizing at least one model based on the error information;
the at least one model includes at least one of: the system comprises an embedded model, an information preprocessing module, a judging model and an information extracting model;
the error information includes at least one of: error between original carrier image and target carrier image, error between information to be hidden and extracted information, cross entropy of discrimination model and error between semantic information and difference image of original carrier image; the difference image is a difference image between the original carrier image and the target carrier image.
Specifically, the internal parameters of the model may be optimized based on the error results of the multiple loss functions, as shown in fig. 4, where the error results of the loss functions include, for example, at least one of the following: mean square error between carrier image and target carrier image obtained by embedding informationMessage to be hiddenMean square error of information and extracted information>Cross entropy of discriminant model>And mean square error between edge image and difference image ∈>The smaller and better the loss values are, the different weight coefficients of the loss function represent the importance degree of the loss.
Since the edge information of the image is more hidden than the flat area, the human eye is less likely to perceive the change of the image information. Therefore, the information hiding strategy used in the embodiments of the present invention is to hide information in edge pixels, i.e. to make more information embedded in edge pixels of the carrier image. It should be noted that, the method of the embodiment of the present invention does not necessarily require that all pixels be embedded in the edge pixels, but the guiding embedding model embeds as many edge pixels as possible, and because the embedding capacity of the edge pixels is limited, for large-capacity and multi-task information to be hidden, it is unlikely that all the information to be hidden is embedded in the edge pixels, so that the imperceptibility of the image may be poor, and therefore, the method of the embodiment of the present invention embeds the information in the edge pixels only in a guiding manner. And the multi-scale edge information is used, and the embedded information of the positions, close to the edges, of the two sides of the edge information is fully utilized.
The method of the embodiment of the invention is described below by combining experimental results, and firstly, an image quality evaluation index and main parameter information in the embodiment of the invention are described; then, carrying out attack experiments of a plurality of different parameters on the method of the embodiment of the invention, and testing the robustness and imperceptibility of the method of the embodiment of the invention on embedding capacity and multitasking; finally, a great number of comparison experiments are used for explaining the advantages of the information hiding method provided by the embodiment of the invention compared with other information hiding algorithms. The following experimental results are analyzed for the semantic information as edge information.
1. Evaluation index
The method for measuring information hiding generally adopts Peak Signal-to-Noise Ratio (PSNR), and compares the degree of difference between an image I' (x, y) embedded with information and a carrier image I (x, y), and the PSNR has the following calculation formula:
where m×n is the size of the image, I (x, y) is the pixel value of the carrier image at the (x, y) position, and I' (x, y) is the pixel value of the image at the (x, y) position after embedding the information. Peak signal-to-noise ratio is typically used to evaluate the imperceptibility of the image after embedding the information, the greater the PSNR value, the less the image distortion and the better the imperceptibility.
Structural similarity (Structural Similarity, SSIM) to construct a fitness function, the structural similarity SSIM function is defined as follows:
in the formula (2): u (u) x ,u y Respectively mean values of x and y; sigma (sigma) xy Is covariance; k (K) 1 ,K 2 Are all a small constant, e.g. take K 1 ＝0.02，K 2 ＝0.01。
In the embodiment of the invention, the accuracy of the extracted information can be evaluated by using the Bit Error Rate (BER), namely the number of the error bits in the extracted information compared with the original information, wherein the lower the bit error rate is, the more accurate the extracted information is represented, and the better the robustness of the algorithm is.
The experimental platform of the embodiment of the invention can be realized based on PyTorch and accelerated by using a graphics card. Alternatively, the cocval 2014 dataset was used, with 4 ten thousand images of 128 x 128 resolution, 80% of which were used for training and the rest for testing. Alternatively, the image data may also be normalized and processed using a random center cropping strategy. Further to illustrate the versatility of the method of the present embodiments, a Mirlickr dataset with 2.5 ten thousand images and a Div2k dataset with 100 images may also be tested. For example, the total number of iterations is 300. The batch size B may be set to 16. Alternatively, an optimizer, such as Adam optimizer, may be used during the experiment, and the initial learning rate may be 0.001.
Table 1 ablation experimental results
Wherein, 1 in three columns of methods in table 1 represents use, and 0 represents no use.
The results of the ablation experiments on the three data sets are shown in table 1. Baseline represents the same size convolution without edge loss, where edge loss represents the addition of edge information at the time of model optimization, edge loss weight coefficients control the amount of information embedded in the image edges, and larger edge loss weight coefficients represent more information embedded in the edges. For information of length 100 bits, baseline achieves very high results in terms of PSNR and SSIM values, but the bit error rate has exceeded 25%. As can be seen from the last 50 times of the residual image in fig. 6, the embedded information alters the image significantly. After adding up-sampling (i.e. deconvolution), the information is perfectly extracted by 100-bit embedding. However, the imperceptibility is reduced and the average PSNR value is 39dB. After adding the edge information, imperceptibility is sacrificed in order to enable 100 bits of information to be fully embedded and extracted. The PSNR value is less than 36 and the SSIM value is less than 0.9. The last two columns show our algorithm with PSNR exceeding 41dB on all three data sets, preferably 41.36dB on Div2K data set. At the same time, the error rate is also all zero. It can be seen from the residual image that edge pixels are embedded with more information under edge guidance.
Figure 5 shows the results of a method of an embodiment of the invention trained on a COCO dataset and tested on a Div2k dataset. The residual information after the edge information guidance is added is mainly concentrated on the edge portion of the image, and the amount of embedded information is small in the flat portion of the image, as can be seen from the enlarged result of the fourth five columns in the first row. The observation from the angle of naked eyes can not distinguish the carrier image and the image embedded with the information, so that the method provided by the embodiment of the invention has good imperceptibility.
The embodiment of the invention uses the strategy of guiding the embedded information based on the image edge information, thereby greatly helping to improve the imperceptibility of the information hiding method.
Furthermore, the method of the embodiment of the invention also has the capability of processing multiple tasks, and experiments are carried out on the task number K from 2 to 5, wherein the length of the task number K is 50 bits. The results in table 2 below show that the method of the present invention maintains a bit error rate of less than 0.1% even with 5 tasks. To alleviate the number of parameters of the model, a different number of convolutional layers are designed adaptively for different numbers of tasks. The parameters of the model are reduced as much as possible under the condition of ensuring the error rate to be zero. The number of processing layers after stacking of information is adaptively changed when designing the multitasking model, and the number of tasks in table 2 also represents the number of convolution layers after embedding of information. The more the number of tasks is, the fewer the number of corresponding convolution layers is, and the scheme can ensure the performance of the algorithm and reduce the parameters of the model.
TABLE 2 Multiplexed experiments
For further safety of the method of the embodiment of the present invention, for example, the detection rate of a plurality of deep learning information hiding algorithms on StegaExpose may be selected to compare with the method of the embodiment of the present invention. The closer the detection rate is to a random guess, the higher the algorithm security is. It should be noted that, as can be seen in fig. 7, the method of the embodiment of the present invention is very close to random guess, which illustrates that the StegaExpose has limited detection capability for the method of the embodiment of the present invention. The reason is that the StegaExpose mainly detects the embedding condition of the least significant bit of the image, and the method of the embodiment of the invention not only embeds information in the least significant bit, but also embeds the information in the middle position, so that the detection rate of the detection algorithm for the method of the embodiment of the invention is close to random guess.
Further, in order to explain the position of the embedded information in the embodiment of the present invention, comparing fig. 8 and fig. 9 with the verification analysis of the valid bit of the image embedded with the information, the method of the embodiment of the present invention not only embeds the information in the least valid bit (the position corresponding to the horizontal axis equal to 7 in the figure), but also embeds the information in the middle position, especially the R and B channel information in the positions of the middle positions 3 and 4 exceeds the information amount of the least significant bit, and the information of the most significant bit falls below 3%, which illustrates that the method of the embodiment of the present invention can well transfer the information of the most significant bit to the middle position for embedding, and well solves the problem of lower imperceptibility of other methods. Because the high order information is more easily perceived, whereas the embedding of the medium and low order information is not easily perceived.
In summary, as shown in fig. 10, the method and the stegasamp algorithm according to the embodiment of the present invention can achieve perfect information extraction, and the bit error rate is zero. In contrast, the stegas algorithm uses the error correction algorithm of BCH to reduce the bit error rate, and one 100 bits of information can only be processed by 56 bits. The method of the embodiment of the invention can process 100 bits of information without an error correction algorithm. The method of the embodiment of the invention is superior to other methods in four evaluation indexes. The evaluation indexes PSNR and SSIM are obviously improved by more than 0.992 and 41dB, which are respectively higher than the Stegasamp algorithm on the HCISNet and Div2k data sets on the COCO data set.
The information hiding device provided by the invention is described below, and the information hiding device described below and the information hiding method described above can be referred to correspondingly.
Fig. 11 is a schematic structural diagram of an information hiding apparatus according to an embodiment of the present invention. As shown in fig. 11, the information hiding apparatus provided in this embodiment includes:
an obtaining module 210, configured to obtain semantic information of an original carrier image;
the processing module 220 is configured to embed information to be hidden into the original carrier image to obtain a target carrier image based on the semantic information by using the trained embedding model;
The semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
Optionally, the embedding model includes a deconvolution module, an overlap layer and a convolution module, and the processing module 220 is specifically configured to:
performing up-sampling processing on the carrier image by using a deconvolution module of the embedded model based on the semantic information to obtain first carrier image information; the first carrier image information has the same size as the information to be hidden obtained by preprocessing;
and superposing the information to be hidden, the first carrier image information and the original carrier image which are obtained by preprocessing on the basis of the semantic information, and carrying out downsampling on the superposed information by utilizing a convolution module of the embedded model on the basis of the semantic information to obtain the target carrier image.
Optionally, the processing module 220 is further configured to:
preprocessing the information to be hidden by using an information preprocessing model to obtain the preprocessed information to be hidden;
the information preprocessing model comprises the following steps: a linear connection layer, a reconstruction layer, and a convolution module; the linear connection layer is used for expanding the information to be hidden to obtain information with the length of N, the reconstruction layer is used for reconstructing the information with the length of N to obtain vectors with the same number of dimensions as the original carrier image, the convolution module is used for carrying out convolution processing on the vectors output by the reconstruction layer to obtain the preprocessed information to be hidden, and N is an integer larger than 1.
Optionally, extracting the model includes: deconvolution module, convolution module and linear connection layer, the processing module 220 is further configured to:
performing up-sampling processing on the target carrier image by using a deconvolution module of the extraction model to obtain second carrier image information;
performing downsampling processing on the second carrier image information by using a convolution module of the extraction model to obtain third carrier image information;
and processing the third carrier image information by using the linear connection layer of the extraction model to obtain extraction information.
Optionally, the semantic information includes at least one of: edge information, gradient information, texture information.
Optionally, the semantic information includes: and the multi-scale semantic information of the original carrier image.
Optionally, the information to be hidden includes information of K different tasks, where K is an integer greater than 1.
Optionally, the processing module 220 is further configured to:
based on the original carrier image and the target carrier image, acquiring difference information of the original carrier image and the target carrier image by utilizing a discrimination model; the difference information is represented using probabilities.
Optionally, the processing module 220 is further configured to:
Optimizing at least one model based on the error information;
the at least one model includes at least one of: the system comprises an embedded model, an information preprocessing module, a judging model and an information extracting model;
the error information includes at least one of: error between original carrier image and target carrier image, error between information to be hidden and extracted information, cross entropy of discrimination model and error between semantic information and difference image of original carrier image; the difference image is a difference image between the original carrier image and the target carrier image.
Optionally, the convolution module includes: a convolutional layer, a regularization layer, and an activation layer.
The deconvolution module includes: deconvolution, regularization and activation layers.
Optionally, the embedded model is trained based on training data, the training data comprising: information to be hidden, position of information to be hidden, carrier image.
The device of the embodiment of the present invention is configured to perform the method of any of the foregoing method embodiments, and its implementation principle and technical effects are similar, and are not described in detail herein.
Fig. 12 illustrates a physical structure diagram of an electronic device, as shown in fig. 12, which may include: processor 810, communication interface (Communications Interface) 820, memory 830, and communication bus 840, wherein processor 810, communication interface 820, memory 830 accomplish communication with each other through communication bus 840. Processor 810 may invoke logic instructions in memory 830 to perform an information hiding method comprising: acquiring semantic information of an original carrier image;
Embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
Further, the logic instructions in the memory 830 described above may be implemented in the form of software functional units and may be stored in a computer-readable storage medium when sold or used as a stand-alone product. Based on this understanding, the technical solution of the present invention may be embodied essentially or in a part contributing to the prior art or in a part of the technical solution, in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, a network device, etc.) to perform all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: a U-disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.
In another aspect, the present invention also provides a computer program product comprising a computer program, the computer program being storable on a non-transitory computer readable storage medium, the computer program, when executed by a processor, being capable of performing the information hiding method provided by the above methods, the method comprising: acquiring semantic information of an original carrier image;
embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
In yet another aspect, the present invention also provides a non-transitory computer readable storage medium having stored thereon a computer program which, when executed by a processor, is implemented to perform the information hiding method provided by the above methods, the method comprising: acquiring semantic information of an original carrier image;
embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and are not limiting; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (12)
1. An information hiding method, comprising:
acquiring semantic information of an original carrier image;
embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image;
the embedding model comprises a deconvolution module, an overlapping layer and a convolution module, the embedding of the information to be hidden into the original carrier image to obtain a target carrier image based on the semantic information by using the trained embedding model comprises the following steps:
Performing up-sampling processing on the original carrier image by using a deconvolution module of the embedded model based on the semantic information to obtain first carrier image information; the first carrier image information has the same size as the information to be hidden obtained by preprocessing;
superposing the information to be hidden, the first carrier image information and the original carrier image obtained by preprocessing on the basis of the semantic information, and carrying out downsampling on the superposed information by utilizing a convolution module of the embedded model on the basis of the semantic information to obtain the target carrier image; the semantic information is used for guiding the embedding position of the information to be hidden.
2. The information hiding method according to claim 1, wherein before the superimposing processing is performed on the information to be hidden, the first carrier image information, and the original carrier image obtained by the preprocessing based on the semantic information, further comprising:
preprocessing the information to be hidden by using an information preprocessing model to obtain the preprocessed information to be hidden;
the information preprocessing model comprises the following steps: a linear connection layer, a reconstruction layer, and a convolution module; the linear connection layer is used for expanding the information to be hidden to obtain information with the length of N, the reconstruction layer is used for reconstructing the information with the length of N to obtain vectors with the same number of dimensions as the original carrier image, the convolution module is used for carrying out convolution processing on the vectors output by the reconstruction layer to obtain the preprocessed information to be hidden, and N is an integer larger than 1.
3. The information hiding method according to claim 2, wherein extracting the model includes: deconvolution module, convolution module, and linear connection layer, the method further comprising:
performing up-sampling processing on the target carrier image by using a deconvolution module of the extraction model to obtain second carrier image information;
performing downsampling processing on the second carrier image information by using a convolution module of the extraction model to obtain third carrier image information;
processing the third carrier image information by using a linear connection layer of the extraction model to obtain extraction information; the extracted information is used for optimizing the embedded model.
4. An information hiding method according to any one of claims 1-3, characterized in that,
the semantic information includes at least one of: edge information, gradient information, texture information.
5. An information hiding method according to any one of claims 1-3, characterized in that,
the semantic information comprises multi-scale semantic information of the original carrier image.
6. An information hiding method according to any one of claims 1-3, characterized in that,
the information to be hidden comprises information of K different tasks, wherein K is an integer greater than 1.
7. An information hiding method according to claim 3, characterized in that the method further comprises:
based on the original carrier image and the target carrier image, acquiring difference information of the original carrier image and the target carrier image by utilizing a discrimination model; the difference information is represented using probabilities.
8. The information hiding method according to claim 7, further comprising:
optimizing at least one model based on the error information;
the at least one model includes at least one of: the system comprises an embedded model, an information preprocessing model, a judging model and an information extracting model;
the error information includes at least one of: error between original carrier image and target carrier image, error between information to be hidden and extracted information, cross entropy of discrimination model and error between semantic information and difference image of original carrier image; the difference image is a difference image between the original carrier image and the target carrier image.
9. An information hiding method according to any one of claims 1-3, characterized in that,
the embedded model is trained based on training data, the training data comprising: information to be hidden, position of information to be hidden, carrier image.
10. An information hiding apparatus, characterized by comprising:
the acquisition module is used for acquiring semantic information of the original carrier image;
the processing module is used for embedding the information to be hidden into the original carrier image to obtain a target carrier image by utilizing the trained embedding model based on the semantic information;
the semantic information is used for indicating the embedding position of the information to be hidden in the original carrier image;
the embedded model comprises a deconvolution module, an overlapping layer and a convolution module, and the processing module is specifically used for:
performing up-sampling processing on the original carrier image by using a deconvolution module of the embedded model based on the semantic information to obtain first carrier image information; the first carrier image information has the same size as the information to be hidden obtained by preprocessing;
superposing the information to be hidden, the first carrier image information and the original carrier image obtained by preprocessing on the basis of the semantic information, and carrying out downsampling on the superposed information by utilizing a convolution module of the embedded model on the basis of the semantic information to obtain the target carrier image; the semantic information is used for guiding the embedding position of the information to be hidden.
11. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the information hiding method according to any one of claims 1 to 9 when executing the program.
12. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the information hiding method according to any one of claims 1 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210583727.1A CN115086503B (en) | 2022-05-25 | 2022-05-25 | Information hiding method, device, equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210583727.1A CN115086503B (en) | 2022-05-25 | 2022-05-25 | Information hiding method, device, equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115086503A CN115086503A (en) | 2022-09-20 |
| CN115086503B true CN115086503B (en) | 2023-09-22 |
Family
ID=83249718
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210583727.1A Active CN115086503B (en) | 2022-05-25 | 2022-05-25 | Information hiding method, device, equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115086503B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598713A (en) * | 2019-08-06 | 2019-12-20 | 厦门大学 | Intelligent image automatic description method based on deep neural network |
| CN112966230A (en) * | 2019-11-28 | 2021-06-15 | 阿里巴巴集团控股有限公司 | Information steganography and extraction method, device and equipment |
| CN114339258A (en) * | 2021-12-28 | 2022-04-12 | 中国人民武装警察部队工程大学 | Information steganography method and device based on video carrier |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108229479B (en) * | 2017-08-01 | 2019-12-31 | 北京市商汤科技开发有限公司 | Training method and device of semantic segmentation model, electronic equipment and storage medium |
-
2022
- 2022-05-25 CN CN202210583727.1A patent/CN115086503B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110598713A (en) * | 2019-08-06 | 2019-12-20 | 厦门大学 | Intelligent image automatic description method based on deep neural network |
| CN112966230A (en) * | 2019-11-28 | 2021-06-15 | 阿里巴巴集团控股有限公司 | Information steganography and extraction method, device and equipment |
| CN114339258A (en) * | 2021-12-28 | 2022-04-12 | 中国人民武装警察部队工程大学 | Information steganography method and device based on video carrier |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115086503A (en) | 2022-09-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378844B (en) | Image blind motion blur removing method based on cyclic multi-scale generation countermeasure network | |
| CN111275637B (en) | Attention model-based non-uniform motion blurred image self-adaptive restoration method | |
| CN110458758B (en) | Image super-resolution reconstruction method and system and computer storage medium | |
| CN109118495B (en) | Retinal vessel segmentation method and device | |
| CN111709883B (en) | Image detection method, device and equipment | |
| CN107240061A (en) | A kind of watermark insertion, extracting method and device based on Dynamic BP neural | |
| CN111711821B (en) | Information hiding method based on deep learning | |
| CN112395943A (en) | Detection method for counterfeiting face video based on deep learning | |
| CN115809953A (en) | Attention mechanism-based multi-size image robust watermarking method and system | |
| CN113538246A (en) | Remote sensing image super-resolution reconstruction method based on unsupervised multi-stage fusion network | |
| CN115984979A (en) | Unknown-countermeasure-attack-oriented face counterfeiting identification method and device | |
| CN116152061A (en) | Super-resolution reconstruction method based on fuzzy core estimation | |
| CN113538616B (en) | Magnetic resonance image reconstruction method combining PUGAN with improved U-net | |
| CN115086503B (en) | Information hiding method, device, equipment and storage medium | |
| CN114283058A (en) | Image super-resolution reconstruction method based on countermeasure network and maximum mutual information optimization | |
| CN116912074A (en) | Image steganography method based on residual improvement MWCNN network | |
| CN115272131B (en) | Image mole pattern removing system and method based on self-adaptive multispectral coding | |
| CN114359009B (en) | Watermark embedding method, watermark embedding network construction method, system and storage medium for robust image based on visual perception | |
| Nam et al. | WAN: Watermarking attack network | |
| CN113269167B (en) | Face counterfeiting detection method based on image blocking and disordering | |
| CN115358910A (en) | Digital watermark attack method and system based on convolutional neural network denoising algorithm | |
| CN116029887A (en) | Image high-capacity robust watermarking method based on wavelet neural network | |
| Rijati | Nested block based double self-embedding fragile image watermarking with super-resolution recovery | |
| CN114936977A (en) | Image deblurring method based on channel attention and cross-scale feature fusion | |
| CN113592697A (en) | Digital watermark attack method and system based on convolutional neural network denoising idea |
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 |