Constructive image hiding method based on attribute conversion

Abstract

The invention provides a constructive image hiding method based on attribute conversion, and the method comprises the steps: 1, selecting two images with completely different styles as a reference image and a secret image respectively, deconstructing the reference image and the secret image at a hiding end, generating secret-containing content features through a content feature hiding network, and hiding the secret-containing content features in the hidden end; a secret-containing conversion image is further obtained through a generator; 2, the extraction end deconstructs the secret-containing image, obtains the content features of the secret image through a content feature extraction network, and further recovers the secret image through a generator; and step 3, designing an information hiding and image extraction loss function to train the frame until convergence. According to the method, image hiding can be realized while image conversion is carried out, and the method has important prospects and application values in the field of hidden transmission.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of computer vision and information security, and particularly relates to a constructional image hiding method based on attribute conversion. BACKGROUND

[0002] Image hiding is a technology of hiding secret images into carrier images for covert transmission. Early image hiding hides secret images by modifying carrier images, which makes it possible to be detected by steganalysis algorithms. To solve this problem, the application proposes a constructional image hiding framework. The principle of constructional image hiding is to generate stego images directly according to secret images, which can theoretically resist steganalysis because it does not need to modify images.

[0003] However, the existing constructional image hiding algorithm has the following problems:

[0004] 1. The existing construction algorithm is insufficient, and the quality of the generated stego image is insufficient.

[0005] 2. The existing hiding algorithm is insufficient, and the generated stego image has insufficient steganalysis resistance. SUMMARY

[0006] The technical problem to be solved by the application is the multiple performance deficiencies of the prior art, and the application provides a constructional image hiding method based on attribute conversion, which is realized by the cooperation of a hiding end and an extraction end, and characterized by the following steps: Step 1, select two images with completely different styles as reference images and secret images, respectively, the hiding end decomposes the reference images and the secret images, and generates stego content features by using a content feature hiding network, and further obtains stego conversion images by using a generator; Step 2, the extraction end decomposes the stego image, and obtains the content features of the secret image by using a content feature extraction network, and further restores the secret image by using a generator; Step 3, design information hiding and image extraction loss functions to train the framework to convergence.

[0007] Step 1 includes the following steps: Step 1-1, decompose the reference images and the secret images into corresponding content features and attribute features, respectively; Step 1-2, hide the content features of the secret image into the content features of the reference image by using a content feature hiding network to obtain stego content features; Step 1-3, input the stego content features and the attribute features of the secret image into a generator to obtain the final stego image.

[0008] In step 1-1, to achieve image feature deconstruction, set x and y to two images with completely different styles, where x is the reference image and y is the secret image; and set a shared content space C; Through content encoders and attribute encoders Deconstruct the reference image x and obtain the content features of x and attribute characteristics , through the content encoder and attribute encoders Deconstruct the secret image y and obtain the content features of y and attribute characteristics , all content features belong to the shared content space C, and the formula is: , .

[0009] Steps 1-2 include: building a content feature hidden network , to achieve feature-level hiding, and to transform content features and After inputting into the first reversible block, the first reversible features are obtained respectively and the second reversible feature ,right Use additive transformation to obtain additive reversible features ,right Affine reversible features are obtained using enhanced affine transformation , and then the result and After connecting them, input them into the next reversible block, and repeat the process of steps 1-2. After two or more reversible blocks, the encrypted content features are obtained. : .

[0010] Steps 1-3 include: Utilizing the generator The confidential content feature and the attribute characteristics of the secret image Fusion to obtain encrypted image : .

[0011] Step 2 includes the following steps: Step 2-1, the encrypted image , deconstructed into the recovered encrypted content features and restored attribute characteristics : ; Step 2-2, build content feature extraction network , the confidential content features After inputting into the first reversible block, the additive reversible features are obtained respectively and affine reversible features , then Perform additive inverse transform to restore the first reversible feature ,right Perform enhanced inverse affine transform to restore the second reversible feature ,Will and After connecting, input into the next reversible block, and cycle this process through two or more reversible blocks to obtain the content features of the recovered secret image. Extracting content features of the recovered secret image : ; Step 2-3, combine the recovered secret image content features and restored attribute characteristics , and finally the recovered secret image is obtained : .

[0012] Step 2-1 includes: Using a content encoder and attribute encoders right Deconstruct and recover the encrypted content features and restored attribute characteristics .

[0013] In step 2-2, the content feature extraction network From the recovered encrypted content features Extract the recovered content features of the secret image y .

[0014] Steps 2-3 include: and Input Recovering the secret image .

[0015] Step 3 includes: Step 3-1, calculate the encrypted image loss , as a secret image With converted image The loss between , using the mean absolute error calculation method , This represents the expectation of a random variable: , wherein the conversion image is directly input into the generator by the content feature of the reference image and the attribute feature of the secret image to obtain: ; Step 3-2, calculating the loss between the secret image y and the recovered secret image : : ; Step 3-3, calculating the secret content feature loss : , wherein represents the mean square error calculation method; Step 3-4, calculating the secret content feature loss : ; Step 3-5, combining the losses to obtain the overall loss L of the attribute conversion-based constructed image hiding: , wherein, and are training weights.

[0016] Beneficial effects: the present application proposes an attribute feature conversion-based constructed image hiding framework, decomposes an image into content features and attribute features, generates images of different styles by changing the attribute features, and performs image hiding, which can solve the problems of insufficient quality of the secret image and insufficient anti-steganalysis performance in the current image hiding field. The present application uses image attribute conversion for image hiding, provides a new idea for image hiding, and has good application prospects in the field of covert communication. It can be applied to the fields of privacy protection and secure communication, digital copyright protection and traceability. BRIEF DESCRIPTION OF DRAWINGS

[0017] The above and other aspects of the present application will become more apparent from the following detailed description, taken in conjunction with the accompanying drawings.

[0018] Figure 1 is a secret image generated by the framework.

[0019] Figure 2 is a workflow diagram for generating a secret conversion image.

[0020] Figure 3It is a workflow diagram for extracting secret images. DETAILED DESCRIPTION

[0021] The embodiment of the present invention provides a constructive image hiding method based on attribute conversion. The resolution of the input image is usually 512×512 (variable). Figure 2 lower left corner image) and a Van Gogh painting (secret image, such as Figure 2 The upper left corner image) is taken as input, and the method secretes the content features of the image Hiding the content features of the reference image through the content feature hiding network Get the secret content features , confidential content characteristics Attribute characteristics of the secret image Combined to generate secret transformation image ; At the extraction end, the encrypted converted image Deconstruction and recovery of encrypted content features , and extract the content features of the recovered secret image through the content feature extraction network , and finally the attribute features of the recovered secret image Combine to get the recovered secret image The encrypted image generated by the present invention has good visual effects and artistry, and can directly hide an image with the same size as the reference image while generating a style-transformed image with artistic value.

[0022] The method specifically includes: using image attribute conversion at the hidden end to change the image attributes and hide the secret image to generate high-quality, diverse secret conversion images At the extraction end, the secret image can be extracted from the secret-containing transformed image. The specific hiding and extraction process is as follows: Step 1: On the hidden side, set x and y to two images of different styles (e.g., x is a natural photo and y is a Van Gogh painting), where x is the reference image and y is the secret image. Figure 2 The complete process of generating a secret transformed image is shown. 、 as well as Respectively Figure 2 The content encoder, attribute encoder and generator in are assumed to have a shared content space C.

[0023] Step 101: The reference image x is passed through the content encoder. and attribute encoders Deconstruction into the content characteristics of x and attribute characteristics , the secret image y is converted into content features of y by a content encoder and attribute encoder , and attribute features of y , , ; Step 102, the content features of the secret image y are hidden into the content features of the reference image x by a content feature hiding network to obtain the stego content features : , wherein, is the reversible block-based hiding network proposed in the present application. The inputs of the network are two content features and , after inputting them into the first reversible block, the first reversible feature and the second reversible feature are obtained, the addition transformation is used on to obtain the addition reversible feature , the enhanced affine transformation is used on to obtain the affine reversible feature , then the obtained results are connected and input into the next reversible block, and this process is repeated for two or more reversible blocks to obtain the fusion feature. Since the number of channels of the input and output of the reversible block is consistent, but two content features are fused in the hiding process, the number of channels of the obtained fusion feature is doubled. Therefore, the original reversible network is improved, the fusion feature is divided according to the number of channels, the first half of the channel features is taken as the final stego content feature , and the second half is taken as the feature matrix r (which needs to be extracted). The specific calculation process of the forward propagation of the reversible block can be represented by the following formula, wherein, , and take the default value of 1. exp is the exponential function, is the dot product operation: , , Step 103, the stego content features and the attribute features of y are input into a generator to generate the stego image , as shown in Figure 1 . The calculation formula is as follows: ; Step 2: At the extraction end, convert the image The deconstruction of can recover the attribute characteristics of the secret image and confidential content characteristics , and then recover the encrypted content features Input to the extraction network of the present invention Get the content features of the recovered secret image Finally, the recovered secret image is combined with the content features and attribute characteristics The recovered secret image can be obtained , Figure 3 This is the entire process of extracting secret images.

[0024] Step 201: Use content encoder and attribute encoders right Deconstruct and recover the encrypted content features and the attribute characteristics of the secret image : , in ; Step 202: Extract network content features From the recovered encrypted content features Extract the content features of the recovered secret image , the formula is: , The specific process is to restore the encrypted content features Combined with the feature matrix r to form a fusion feature, and then input into a reversible block to recover the additive reversible feature and affine reversible features Then Perform additive inverse transform to restore the first reversible feature ,right Perform enhanced inverse affine transform to restore the second reversible feature ,Will and After being connected, they are input into the next reversible block. After repeating this process for two or more reversible blocks, the content features of the reconstructed and recovered secret image can be obtained. The specific calculation process of the reversible block reverse propagation can be expressed by the following formula: , , Step 203: Get the content features of the secret image and attribute characteristics Input to the generator Restoring the secret image : Step 3, In order to generate high-quality stego-converted images and restore high-quality secret images, it is necessary to design hidden and extraction loss functions to train the framework to convergence. This method designs four kinds of loss functions to train the proposed framework.

[0025] Step 301, First, in order to generate high-quality stego-converted images, a stego image loss is proposed to calculate the loss between the stego-converted image and the converted image, using the mean absolute error calculation method represents the expectation of a random variable: where represents the converted image, which is directly input into the generator by the content features of the reference image and the attribute features of the secret image : Step 302, Calculate the loss between the secret image and the restored secret image Step 303, Calculate the stego content feature loss . Where, represents the mean square error calculation method: Step 304, Calculate the secret content feature loss Step 305, Finally, the above four kinds of loss are combined to get the overall loss L of the attribute-converted stego image construction based on: where and are training weights, the default values are all 1, which can be adjusted according to actual conditions.

[0026] This method can be used in the fields of privacy protection and secure communication, digital copyright protection and traceability, etc.

[0027] ​​​​​​​​​​​The application provides a constructional image hiding method based on attribute conversion. The above description is only the preferred embodiment of the application, and it should be pointed out that, for the ordinary skilled in the art, several improvements and refinements can be made without departing from the principle of the application, and these improvements and refinements should also be regarded as the protection scope of the application. The components not explicitly described in the embodiment can be realized by the prior art.

Claims

1. An attribute transformation based constructional image hiding method, the method is realized by the hiding end and the extraction end working together, characterized in that, The method comprises the following steps: Step 1, selecting two images of completely different styles as a reference image and a secret image respectively, deconstructing the reference image and the secret image, generating a content feature containing a secret by using a content feature hiding network, and further obtaining a conversion image containing a secret by using a generator; Step 2, deconstructing the image containing the secret, obtaining the content feature of the secret image by using a content feature extraction network, and further recovering the secret image by using a generator; Step 3, designing an information hiding and image extraction loss function to train the framework to converge.

2. The method of claim 1, wherein, Step 1 comprises the following steps: Step 1-1, deconstructing the reference image and the secret image into corresponding content features and attribute features respectively; Step 1-2, hiding the content feature of the secret image into the content feature of the reference image by using a content feature hiding network to obtain a content feature containing a secret; Steps 1-3, input the secret image and the attribute features of the secret image into the generator to obtain the final stego image.

3. The method of claim 2, wherein, In step 1-1, to realize the feature deconstruction of the image, x and y are set as two images of completely different styles, wherein x is a reference image and y is a secret image; and a shared content space C is set; By content encoder and attribute encoder deconstruct the reference image x to get the content features of x and attribute features , by content encoder and attribute encoder deconstruct the secret image y to get the content features of y and attribute features , all content features belong to the shared content space C, formula: ， 。 4. The method of claim 3, wherein, Steps 1-2 include: building a content feature hidden network , to achieve feature-level hiding, and to transform content features and After inputting into the first reversible block, the first reversible features are obtained respectively and the second reversible feature ,right Use additive transformation to obtain additive reversible features ,right Affine reversible features are obtained using enhanced affine transformation , and then the result and After connecting them, input them into the next reversible block, and repeat the process of steps 1-2. After two or more reversible blocks, the encrypted content features are obtained. : 。 5. The method of claim 4, wherein, Steps 1-3 include: using a generator secret image and attribute features of the secret image fusion to obtain a stego image : 。 6. The method of claim 5, wherein, Step 2 comprises the following steps: Step 2-1, to the steg-image , deconstructing into recovered stego-content features and recovered attribute features : ； Step 2-2: Build a content feature extraction network , the confidential content features After inputting into the first reversible block, the additive reversible features are obtained respectively and affine reversible features , then Perform additive inverse transform to restore the first reversible feature ,right Perform enhanced inverse affine transform to restore the second reversible feature ,Will and After connecting, input into the next reversible block, and cycle this process through two or more reversible blocks to obtain the content features of the recovered secret image. Extracting content features of the recovered secret image : ； Step 2-3, combining recovered secret image content features and recovered attribute features resulting in a recovered secret image : 。 7. The method of claim 6, wherein, Step 2-1 includes using a content encoder and an attribute encoder to deconstruct the recovered content feature and the recovered attribute feature .

8. The method of claim 7, wherein, In Step 2-2, the content feature is extracted by a content feature extraction network From the recovered content feature with the secret The recovered content feature of the secret image y is extracted from the recovered content feature with the secret .

9. The method of claim 8, wherein, Step 2-3 includes: recovering the secret image and from the input to Step 2-4 includes: outputting the secret image .

10. The method of claim 9, wherein, Step 3 comprises: Step 3-1, calculating stego image loss , as a loss between the stego image and the converted image , using the mean absolute error calculation method , denotes the expectation over the random variable: ， wherein the converted image is derived from the content features of the reference image and the attribute features of the secret image are directly input into the generator to obtain: ； Step 3-2, calculating the loss between the secret image y and the recovered secret image :​ ； Step 3-3, compute loss of content-embedded feature : ， wherein denotes the mean square error calculation method; Step 3-4, compute secret content feature loss : ； Step 3-5, combining the losses to obtain a total loss L of the constructed image hiding based on the attribute conversion: ， wherein, with are the training weights.

Images