A steganography method without embedded information based on multi semantic fusion of pictures

Abstract

The application discloses a non-embedded steganography method based on picture semantic mixing, and has the characteristics that it comprises the following steps: 1. Through mapping, scaling and other operations, the binary secret message S is converted into independent picture carrier pixels and is shuffled into a carrier image with a cover layer; 2. A reference image with a cover layer is randomly sampled from a reference picture; 3. The carrier image is combined with the cover layer to obtain a semantic mixed image with multiple cover layers; 4. A generator and a discriminator are constructed; 5. A loss function is constructed, and the model is trained using the gradient descent method. The embedding scheme of the application is based on the conversion mapping of the secret message and the pixel, the message and the picture are semantically fused, and the generative adversarial network is trained, so that the application has the advantages of high information embedding rate, strong anti-steganalysis ability and wide practical application scene.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of digital image steganography, and more particularly relates to a no-embedding information steganography method based on picture multi-semantic fusion. BACKGROUND

[0002] Digital steganography technology uses modern digital carriers to embed secret messages, and the receiver can extract and restore the secret messages from the carriers according to a specific mode, while third parties on the network cannot judge the existence of the secret messages, thereby realizing secret communication. Existing mainstream technologies often strive to find effective distortion functions to embed information in the redundant part of digital images. With the advent of adversarial neural networks, the field of digital image steganography has developed to a great extent, especially the no-embedding steganography method using image generation technology has developed greatly, mainly based on different mapping methods of secret messages, training network models, and generating stego images. However, most technologies in this field currently have the problems of low message embedding rate, inaccurate restoration, and poor actual application effect.

[0003] We map and appropriately scale the binary message into pixel values in the image, and perform semantic mixing with the reference message randomly sampled from the reference image. The network model is trained to obtain a stego image with good image quality, high message embedding rate, strong concealment, and good practicality. SUMMARY

[0004] In view of the above problems in the prior art, the no-embedding steganography method based on picture multi-semantic fusion provided by the application solves the problems in the prior art.

[0005] In order to achieve the above-mentioned purposes, the technical scheme adopted by the application is as follows: a no-embedding steganography method based on picture multi-semantic fusion, comprising the following steps:

[0006] S1, mapping the secret message into independent pixel points and shuffling them into a covered message image;

[0007] S2, randomly sampling reference information from a reference image;

[0008] S3, combining the message image obtained in S1 with the reference information obtained in S2 to obtain a semantic mixed image with multiple layers of cover;

[0009] S4, constructing a network model,

[0010] S5, constructing a loss function and training the network model to obtain a stego image.

[0011] Further, the step S1 is specifically defined as follows: the secret message S is in binary format, and each m-bit bit segment is Si , converted into decimal number, and scaled and shifted to get the steganographic pixel p i , summarized as formula

[0012]

[0013] where k is the amplification factor, satisfying 1≤k≤2 8-m , Ψ B→D denotes the conversion operation from binary to decimal, is the positioning item, used to enhance the accuracy of message restoration, and δ is the shift item, satisfying 0≤kΨ B→D ( i )+(k+1)+δ≤255.

[0014] Further, in the step S1, the steganographic pixel p i is scrambled to get the covered message image. Wherein, the image can be filled with the number 0, and for the position of the steganographic pixel, 1 is used to represent in the cover layer M, and 0 is used to represent for the rest. It can be represented as:

[0015]

[0016] In the scrambling operation, Arnold transformation, Fibonacci transformation, Hilbert transformation, magic square transformation and other algorithms can be used for scrambling.

[0017] Further, the step S2 is specifically: random sampling is performed on the reference image I truth , and the rest is filled with 0. For the position of the sampled pixel, 1 is used to represent in the cover layer M ′ , and 0 is used to represent for the rest. It can be represented as

[0018]

[0019] After that, in order to avoid information coverage, M′ needs to be further processed, as shown in the formula

[0020] M′ ij =M′ ij (1-M ij )

[0021] Further, the step S3 is specifically: according to the cover layer M and M′, the message image obtained in S1 and the reference information obtained in S2 are semantically mixed, which can be specifically represented as

[0022]

[0023] The calculation formula is

[0024] p″ij = M ij p ij + M' ij p' ij

[0025] Further, in the step S4, the network model is a generative adversarial network model, which is divided into a generator and a discriminator. The input of the generator is the semantic mixed image I in obtained in the step S3 and the multi-layer coverings M and M', and the output of the generator is the stego image I out to be transmitted. The input of the discriminator is the output I out of the generator and the reference image I truth in the step S2, and the output of the discriminator is a judgment score for the two inputs respectively.

[0026] Further, the generator in the step S4 is a U-shaped convolutional neural network, which includes 8 groups of down-sampling modules and 8 groups of up-sampling modules respectively. The down-sampling modules are sequentially connected, the up-sampling modules are sequentially connected and additionally connected to the outputs of the corresponding levels of the down-sampling modules. In the down-sampling module, a convolutional layer, a batch normalization layer and an activation layer are sequentially connected, wherein the convolution kernel size of the convolutional layer is set to 3*3 and the step is 2, and the activation layer adopts a leaky rectified linear unit function as an activation function. In the up-sampling module, a deconvolutional layer, a batch normalization layer and an activation layer are sequentially connected, and similarly, the convolution kernel size of the convolutional layer is set to 3*3 and the step is 2, but the activation layer adopts a linear rectifier function as an activation function, and in particular, the hyperbolic tangent function is adopted as the activation function in the last layer of the up-sampling module.

[0027] Further, the discriminator in the step S4 is a sequential structure, which sequentially connects 6 groups of down-sampling modules and a fully connected layer. The down-sampling modules have the same structure as the down-sampling modules in the generator, which sequentially connect a convolutional layer, a batch normalization layer and an activation layer. Among them, the convolution kernel size of the convolutional layer is set to 3*3 and the step is 2, and the activation layer adopts a leaky rectified linear unit function as an activation function.

[0028] Further, the loss function in the step S5 is specifically:

[0029] L total = λ adv L adv + λ msg L msg + λ img L img + λ tv L tv

[0030] L advdenotes an adversarial loss function, L msg denotes a message loss function, L img denotes an image loss function, L tv denotes a total variation loss function, λ adv , λ msg , λ img , λ tv denote L adv , L msg , L img , L tv the weight of the loss function L total .

[0031] Further, the adversarial loss function L adv is specifically

[0032] L adv = log D (I truth ) + log (1-D (G (I in ) ) ).

[0033] where I truth is a reference image, and I in is a semantic mixed image.

[0034] The message loss function L msg is specifically:

[0035]

[0036] where M denotes a mask layer of a message image, N M denotes the number of valid pixels in the message image, and is calculated as N M =∑M, I in denotes a mixed semantic image input to the network model, I out denotes a steganographic image output by the network model, and ||*||1 denotes a 1-norm.

[0037] The image loss function L img is specifically:

[0038]

[0039] Similarly, where M 2′ denotes a mask layer of a reference image, N M′ denotes the number of valid pixels in the reference image, and is calculated as N M′ =∑M′.

[0040] The total variation loss function L tv is specifically:

[0041]

[0042] wherein, p i,j represents the pixel point in the steganographic image I out output by the network model. By default, β is a value of 2.

[0043] Further, in the step S5, the training method of the network model is to train the generative adversarial network according to the loss function and by the gradient descent method. First, the generator is fixed, the network model is run, the loss function is calculated according to the output result of the discriminator, and the discriminator is updated. Then, the discriminator is fixed, the network model is run, and the loss function is calculated, and the generator is updated. The above two steps are alternately performed, so that the generator and the discriminator are iterated with each other until the steganographic image output by the generator meets the embedding requirement.

[0044] Compared with the prior art, the beneficial effects of the technical scheme of the present application are:

[0045] (1) The present application provides a new message and pixel conversion method, which is novel and highly secret.

[0046] By scaling and translation, the accuracy of message restoration is enhanced, and the free conversion of secret messages in different carriers is realized.

[0047] (2) The present application uses multiple cover layers to make the network model learn the features of two semantics, and constrains them through the loss function, so that the secret message is hidden, the reference message is explicitly represented, and the network model generates high-quality natural images through adversarial training.

[0048] (3) The present application adopts a multi-semantic hybrid method of secret messages and reference pictures, which can specify the content of the generated image, has strong applicability, is more flexible and natural in actual application, and can effectively hide secret messages. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A flowchart of a multi-semantic fusion-based non-embedding steganography method according to the present application.

[0050] Figure 2 A message and pixel conversion diagram according to the present application.

[0051] Figure 3 A generator diagram according to the present application.

[0052] Figure 4 A discriminator diagram according to the present application.

[0053] Figure 5 A down-sampling module diagram according to the present application.

[0054] Figure 6 An up-sampling module diagram according to the present application. Detailed Implementation

[0055] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0056] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0057] like Figure 1 As shown, a method for steganography without embedding information based on multi-semantic fusion of images includes the following steps:

[0058] S1, such as Figure 2 As shown, first according to the mapping formula Map the secret message S to individual secret-carrying pixels, where message fragment S i It contains m bits of information, where m is an integer, typically set to 3-8. k is the amplification factor, which can range from 1-32 depending on the value of m. Next, the encrypted pixels are sequentially arranged and scrambled into a masked message image. Scrambling algorithms such as Arnold transform, Fibonacci transform, Hilbert transform, and magic square transform can be used. Blank areas can be filled with zeros or random noise. Finally, a masking layer M1 is used to mark the encrypted pixels in the message image; pixels at encrypted positions are marked with a value of 1, while those not classified as encrypted pixels are marked with a value of 0.

[0059] S2. Extract reference images from the database and randomly sample reference pixels from the reference images. Similarly, use the masking layer M2 to mark the reference pixels in the reference images. For pixels that are reference pixels, set the mark to 1; for pixels that are not reference pixels, set the mark to 0. Then, apply the formula to the masking layer M2. Update.

[0060] S3, according to the formula Based on the masking layers M1 and M2, semantic mixing is performed on the message image obtained from S1 and the reference image obtained from S2 to obtain a semantically mixed image with multiple masking layers. After normalization, it is used as the input of the model.

[0061] S4. Construct the network model, mainly including the generator and discriminator. The generator has a U-shaped structure, such as... Figure 3As shown, there are 8 sets of downsampling and upsampling modules. The input is the semantic mixture image obtained in S3, and the output is the generated target's density-carrying image. The downsampling modules are connected sequentially, and the upsampling modules are sequentially connected to, and additionally connected to, the outputs of the corresponding level's downsampling modules. The output of the previous level's upsampling module is concatenated with the output of the corresponding level's downsampling module, and the result is used as the input of that module. Figure 5 As shown, in the downsampling module, a convolutional layer, a batch normalization layer, and an activation layer are connected sequentially. The convolutional layer has a 3x3 kernel size and a stride of 2, and the activation layer uses a linear unit function with leakage correction as its activation function. In the upsampling module, as shown... Figure 6 As shown, a deconvolutional layer, a batch normalization layer, and an activation layer are connected sequentially. Similarly, the kernel size of the convolutional layer is set to 3x3 with a stride of 2. However, the activation layer uses a linear rectified function as its activation function. Specifically, in the last upsampling module, a hyperbolic tangent function is used as the activation function. The discriminator has a sequential structure, as shown... Figure 4 As shown, six downsampling modules are sequentially connected to one or two fully connected layers. The downsampling modules have the same structure as the downsampling modules in the generator, as shown below. Figure 5 As shown, a convolutional layer, a batch normalization layer, and an activation layer are connected sequentially. The convolutional kernel size is set to 3x3 with a stride of 2, and the activation layer uses a linear unit function with leakage correction as its activation function. In the fully connected layer, the Flatten function is first used to expand the feature map of the upper layer, followed by the Dense function with a dimension of 1 to obtain a decision score.

[0062] S5. Construct the loss function. The total loss function is as follows:

[0063] L total =λ adv L adv +λ msg L msg +λ img L img +λ tv L tv

[0064] Where L adv L represents the adversarial loss function. msg L represents the message loss function. img L represents the image loss function. tv λ represents the total variational loss function. adv , λ msg , λ img , λ tv L respectively adv L msg Limg L tv In the loss function L totai The weight it occupies in the middle.

[0065] The adversarial loss function L adv Specifically

[0066] L adv =ogD(I truth )+log(1-(G(I in )))

[0067] Among them, I truth For reference image, I in This is a semantically mixed image.

[0068] The message loss function L msg Specifically:

[0069]

[0070] Where M represents the overlay layer of the message image, N M N represents the number of valid pixels in the message image. M =∑M,I in I represents the mixed semantic image as input to the network model. out This represents the dense image output by the network model, and ||*||1 represents the 1-norm.

[0071] The image loss function L img Specifically:

[0072]

[0073] Similarly, M 2′ N represents the masking layer of the reference image. M′ N represents the number of valid pixels in the reference image. M′ =∑M ′ .

[0074] The total variational loss function L tv Specifically:

[0075]

[0076] Where, p i,j The dense image I represents the output of the network model. out The number of pixels in the array. β defaults to the value 2.

[0077] The network model is trained using a loss function and gradient descent. First, the generator is fixed, the network model is run, and the loss function is calculated based on the discriminator's output, updating the discriminator. Then, the discriminator is fixed again, the network model is run, the loss function is calculated, and the generator is updated. These two steps are repeated alternately, allowing the generator and discriminator to iterate until the image output by the generator meets the embedding requirements.

Claims

1. An embedding-free steganography method based on image multi-semantic fusion, characterized in that, Based on the transformation mapping between secret messages and pixels, semantic fusion of messages and images is performed, and a generative adversarial network is trained. The training process includes the following steps: S1: Map the secret message to individual pixels and scramble it into a masked message image; S2: Obtain reference information by randomly sampling from the reference image; S3: Combine the message image obtained in S1 with the reference information obtained in S2 to obtain a semantically hybrid image with multiple layers of masking; S4: Construct the network model; S5: Construct the loss function and train the network model to obtain the image with high density; In step S1, the secret message S is defined as binary format, and fragments are divided and converted according to the conversion method to obtain the secret-carrying pixel p. i ; then for the dense pixel p i Scramble the image to obtain the masked message image; The aforementioned dense pixel p i During the scrambling process, algorithms such as Arnold transform, Fibonacci transform, Hilbert transform, and magic square transform can be used for scrambling operations, and the covered parts can be filled with the number 0; In step S2, the reference image I truth Perform random sampling, and fill the rest with 0. Random sampling can be performed through random overlay layers. In step S3, according to the different masking layers obtained in the above steps, the message image obtained in S1 and the reference information obtained in S2 need to be semantically mixed. In step S4, the network model is a generative adversarial network model, which is divided into a generator and a discriminator. The generator in step S4 is a U-shaped convolutional neural network, comprising 8 sets of downsampling modules and 8 sets of upsampling modules. The downsampling modules are connected sequentially, and the upsampling modules are connected sequentially and additionally to the outputs of the corresponding layers of downsampling modules. In the downsampling module, a convolutional layer, a batch normalization layer, and an activation layer are connected sequentially. The convolutional kernel size of the convolutional layer is set to 3*3 with a stride of 2, and the activation layer uses a linear unit function with leakage correction as the activation function. In the upsampling module, a deconvolutional layer, a batch normalization layer, and an activation layer are connected sequentially. Similarly, the convolutional kernel size of the convolutional layer is set to 3*3 with a stride of 2, but the activation layer uses a linear rectified function as the activation function. In particular, in the last layer of the upsampling module, a hyperbolic tangent function is used as the activation function. The discriminator in step S4 has a sequential structure, connecting 6 groups of downsampling modules and a fully connected layer in sequence; wherein, the downsampling module has the same structure as the downsampling module in the generator, connecting a convolutional layer, a batch normalization layer and an activation layer in sequence; wherein, the convolutional kernel size of the convolutional layer is set to 3*3 and the stride is 2, and the activation layer uses a linear unit function with leakage correction as the activation function. The loss function in step S5 includes four sub-loss functions: adversarial loss function, message loss function, image loss function, and total variational loss function. The training method for the network model in step S5 is to train the generative adversarial network according to the loss function and gradient descent. First, fix the generator, run the network model, calculate the loss function according to the output of the discriminator, and update the discriminator. Then, fix the discriminator, run the network model, calculate the loss function, and update the generator. The above two steps are performed alternately, so that the generator and discriminator are trained iteratively.

2. The embedding-free steganography method based on image multi-semantic fusion according to claim 1, characterized in that: The conversion process in step S1 can be represented as follows: Where S i Here, k is a message fragment, and k is the amplification factor, satisfying 1 ≤ k ≤ 2. 8-m Ψ B→D This represents the conversion operation from binary to decimal. The positioning term is used to enhance the accuracy of message reconstruction, and δ is the shift term, satisfying 0≤kΨ. B→D (S i )+(k+1)+δ≤255.

3. The image-based multi-semantic fusion-based non-embedding steganography method according to claim 1, characterized in that: In step S1, the dense pixel p i The message image is obtained by scrambling; the position of the dense pixel is represented by 1 in the masking layer M, and the rest is represented by 0, which can be expressed as:

4. The embedding-free steganography method based on image multi-semantic fusion according to claim 1, characterized in that: Step S2 involves randomly sampling the image and filling the remaining parts with the number 0; wherein the position of the sampled pixel is represented by 1 in the masking layer M′, and by 0 for the rest, which can be represented as follows: Next, to avoid information overwriting, M′ needs further processing, as shown in the formula: M′ ij =M′ ij (1-M ij )。 5. The image-based multi-semantic fusion-based non-embedding steganography method according to claim 1, characterized in that: In step S3, semantic mixing is performed on the message image obtained in S1 and the reference information obtained in S2 according to the masking layers M and M′, which can be specifically represented as follows: The calculation formula is: p″ ij =M ij p ij +M′ ij p′ ij 。 6. The image-based multi-semantic fusion-based non-embedding steganography method according to claim 1, characterized in that: The generator described in step S4 is a U-shaped convolutional neural network, comprising 8 sets of downsampling modules and 8 sets of upsampling modules. The downsampling modules are connected sequentially, and the upsampling modules are connected sequentially and additionally to the outputs of the corresponding layers of downsampling modules. In the downsampling module, a convolutional layer, a batch normalization layer, and an activation layer are connected sequentially. The convolutional kernel size of the convolutional layer is set to 3*3 with a stride of 2, and the activation layer uses a linear unit function with leakage correction as the activation function. In the upsampling module, a deconvolutional layer, a batch normalization layer, and an activation layer are connected sequentially. Similarly, the convolutional kernel size of the convolutional layer is set to 3*3 with a stride of 2, but the activation layer uses a linear rectified function as the activation function. In particular, in the last layer of the upsampling module, a hyperbolic tangent function is used as the activation function.

7. The embedding-free steganography method based on image multi-semantic fusion according to claim 1, characterized in that: The discriminator described in step S4 has a sequential structure, connecting 6 groups of downsampling modules and a fully connected layer in sequence; wherein, the downsampling module has the same structure as the downsampling module in the generator, and is connected in sequence to a convolutional layer, a batch normalization layer and an activation layer; wherein, the convolutional kernel size of the convolutional layer is set to 3*3 and the stride is 2, and the activation layer uses a linear unit function with leakage correction as the activation function.

8. The image-based multi-semantic fusion-based non-embedding steganography method according to claim 1, characterized in that: The loss function mentioned in step S5 is specifically as follows: L total =λ adv L adv +λ msg L msg +λ img L img +λ tv L tv Where L adv L represents the adversarial loss function. msg L represents the message loss function. img L represents the image loss function. tv λ represents the total variational loss function; adv , λ msg , λ img , λ tv L adv L msg L img L tv In the loss function L total The weight it occupies in the middle; The adversarial loss function L adv Specifically: L adv =logD(I truth )+log(1-D(G(I in ))) Among them, I truth For reference image, I in A semantically mixed image; The message loss function L msg Specifically: Where M represents the overlay layer of the message image, N M N represents the number of valid pixels in the message image. M =∑M,I in I represents the mixed semantic image as input to the network model. out This represents the dense image output by the network model, where ||*||1 represents the 1-norm; The image loss function L img Specifically: Similarly, M 2′ N represents the masking layer of the reference image. M′ N represents the number of valid pixels in the reference image. M′ =∑M′; The total variational loss function L tv Specifically: Where, p i，j The image I represents the data structure output by the network model. out The number of pixels in the array; β defaults to the value 2.

Images