Image steganalysis method based on reversible neural network and electronic equipment

Abstract

The invention discloses an image steganalysis method based on a reversible neural network and electronic equipment. The method comprises the following steps: inputting a carrier image and a secret image into a multilayer reversible neural network, and processing an initial steganographic image output by the multilayer reversible neural network by adopting a residual augmentation strategy to obtain a final steganographic image; the final steganographic image and the carrier image are matched with random Gaussian noise to be input into a multilayer reversible neural network for reverse recovery, and a recovered secret image and a recovered carrier image are output; calculating an overall loss function, and updating parameters of the multilayer reversible neural network; repeating the above steps until the multilayer reversible neural network converges; and inputting the to-be-analyzed image and the random Gaussian noise into the multilayer reversible neural network, outputting a recovered image by the multilayer reversible neural network, and judging whether the to-be-analyzed image is a steganographic image or not according to the recovered image. According to the method, whether the to-be-analyzed image is the steganographic image or not can be judged, hidden information in the steganographic image can be recovered, and generalization ability is improved.

Description

Technical Field

[0001] This invention relates to the fields of information hiding and digital image processing technology, and in particular to an image steganalysis method and electronic device based on a reversible neural network. Background Technology

[0002] Image steganography algorithms aim to embed a secret image into a cover image to generate a stego image. Their core objective is to achieve large-capacity information embedding and highly covert secret communication, thus they are widely used in secure communication, copyright protection, and digital forensics. Existing image steganography algorithms are mainly divided into two categories: methods based on autoencoders (AEs) and methods based on invertible neural networks (INNs). Among them, INN-based image steganography algorithms show superior overall performance in terms of stegograph quality and hidden information recovery.

[0003] As an adversarial technique against image steganography algorithms, image steganalysis can determine whether a given image contains hidden information, which is of significant research importance for ensuring information security. Early image steganalysis methods were mostly based on statistical models, such as detecting steganographic traces through statistical information between adjacent pixels within a region. With the development of deep learning, deep neural networks have been introduced into this field, leading to a series of end-to-end steganalysis models. These methods, leveraging large-scale datasets and transfer learning techniques, have achieved certain breakthroughs in detection accuracy; however, they still suffer from a lack of interpretability and insufficient generalization ability.

[0004] Chinese Patent Publication No. CN117424963A discloses an image steganalysis method based on CNN and Transformer. This method uses a publicly available steganalysis dataset to train an image steganalysis model built on CNN and Transformer, and uses the trained steganalysis model to determine whether a given image is a steganalysis image. This method has the following defects: (1) It can only determine whether a given image is a steganalysis image, but cannot recover or interpret the information hidden in it, resulting in a lack of interpretability in the discrimination process; (2) It relies on the carrier image and pre-generated steganalysis images for training, the training sample size is limited and the diversity of steganalysis methods in the training set is insufficient, resulting in a significant decrease in the detection performance of the trained image steganalysis model in cross-dataset or cross-method scenarios, and insufficient generalization ability. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an image steganalysis method and electronic device based on a reversible neural network. This method can determine whether an image to be analyzed is a steganalyte, recover the hidden information in the steganalyte, and greatly improve the generalization ability across datasets and steganalysis methods through a residual augmentation strategy.

[0006] To solve the above problems, the present invention adopts the following technical solution: The present invention provides an image steganalysis method based on a reversible neural network, comprising the following steps: S1: Transfer the carrier image and secret images Inputting the secret image into a multilayer reversible neural network, the multilayer reversible neural network will process the secret image. Embedded carrier image Output the initial steganalysis image ; S2: Employ a residual augmentation strategy on the initial stegana. The final steganalysis image is obtained through processing. ; S3: Final Steganographic Image Carrier image Each layer of reversible neural network is used to perform inverse reconstruction by inputting random Gaussian noise, and the reversible neural networks output the recovered secret images respectively. Carrier image ; S4: Calculate the overall loss function and update the parameters of the multilayer invertible neural network; S5: Repeat steps S1 to S4 until the multilayer reversible neural network converges and training is complete; S6: Input the image to be analyzed and random Gaussian noise into the trained multilayer reversible neural network. The multilayer reversible neural network outputs the restored image. Based on the restored image, determine whether the image to be analyzed is a steganalysis image.

[0007] Preferably, step S1 includes the following steps: Carrier image Secret Images Discrete wavelet transform is performed on the carrier image respectively. Secret Images The images are decomposed into low-frequency and high-frequency subbands, then input into a multi-layer reversible neural network. The multi-layer reversible neural network then processes the secret image. Embedded carrier image The output of the multilayer reversible neural network is transformed by inverse discrete wavelet transform to obtain the initial steganalysis image. .

[0008] Preferably, the final steganalysis image The calculation formula is as follows: , in, The coefficients are randomly sampled from a uniform distribution [0,1]. It represents the Hadamardi (or Hadama) stack.

[0009] Preferably, step S3 includes the following steps: The final steganographic image After performing discrete wavelet transform and random Gaussian noise The input is combined with a multilayer reversible neural network for inverse reconstruction. The multilayer reversible neural network outputs the reconstruction result, which is then subjected to inverse discrete wavelet transform to obtain the secret image. ; Carrier image After performing discrete wavelet transform and random Gaussian noise The images are input into a multilayer reversible neural network for inverse reconstruction. The network outputs the reconstruction result, which is then subjected to inverse discrete wavelet transform to obtain the carrier image. .

[0010] Preferably, the random Gaussian noise Random Gaussian noise These are different noises sampled from the same Gaussian distribution.

[0011] Preferably, when inputting an image and random Gaussian noise into a multilayer invertible neural network for reverse recovery, the iterative formula of the i-th layer invertible neural network is as follows: , , in, The noise is the inverse output of the i-th layer of the reversible neural network. The noise is the inverse input to the i-th layer of the reversible neural network. The image is the reverse output of the i-th layer of the reversible neural network. The image is the inverse input of the i-th layer of the reversible neural network. , , For a multilayer invertible neural network, it is a nonlinear function. This is the Sigmoid activation function.

[0012] Preferably, the formula for the overall loss function is as follows: , , , , , in, For the overall loss function, , , , For hyperparameters, Indicates mean square error. This indicates the low-frequency wavelet subband extraction operation.

[0013] Preferably, the method for determining whether the image to be analyzed is a steganalysis image based on the restored image is as follows: calculate the similarity between the restored image and the image to be analyzed; if the similarity is less than or equal to a set threshold, it indicates that the image to be analyzed is a steganalysis image; if the similarity is greater than the set threshold, it indicates that the image to be analyzed is not a steganalysis image.

[0014] Preferably, the similarity is expressed using peak signal-to-noise ratio or structural similarity index.

[0015] An electronic device according to the present invention includes a memory and a processor, wherein executable code is stored in the memory, and when the executable code is executed by the processor, the above-described method is performed.

[0016] The beneficial effects of this invention are: (1) It can output a restored image based on the input image to be analyzed, and determine whether the image to be analyzed is a steganalyte based on the similarity between the image to be analyzed and the restored image. If it is a steganalyte, the restored image can partially restore the hidden secret image, making the analysis results more intuitive and credible, and improving the interpretability of the steganalysis process. (2) It uses a multi-layer reversible neural network to realize the secret image. Embedded carrier image This generates the initial steganalysis image. Then, the initial stegana was augmented using a residual augmentation strategy. With carrier image The differences are randomly perturbed to simulate diverse steganalysis generation methods, thereby improving the diversity of steganalysis image samples and significantly improving the generalization of this method in cross-dataset and cross-steganography scenarios. (3) There is no need to construct large-scale steganalysis image-carrier image pairs during the training phase, which simplifies the data preparation process and reduces the dependence on the training dataset. Attached Figure Description

[0017] Figure 1 This is a flowchart of an embodiment; Figure 2 This is a schematic diagram of the images used for training in this embodiment; Figure 3This is a schematic diagram illustrating the effect of the method in this embodiment on steganalyzing a stegated image and outputting a restored image; Figure 4 This is a schematic diagram illustrating the effect of the method in this embodiment of the example on the output of the restored image through steganalysis of stegated images obtained by different steganalysis methods. Detailed Implementation

[0018] The technical solution of the present invention will be further described in detail below through embodiments and in conjunction with the accompanying drawings.

[0019] Example: This example illustrates an image steganalysis method based on a reversible neural network, such as... Figure 1 As shown, it includes the following steps: S1: Transfer the carrier image Secret Images Discrete wavelet transform is performed on the carrier image respectively. Secret Images The images are decomposed into low-frequency and high-frequency subbands, then input into a multi-layer reversible neural network. The multi-layer reversible neural network then processes the secret image. Embedded carrier image The output of the multilayer reversible neural network is transformed by inverse discrete wavelet transform to obtain the initial steganalysis image. .

[0020] Taking the i-th layer of a reversible neural network as an example, the input-output mapping relationship of its single layer can be expressed as: , , in, The image is the carrier image input to the i-th layer of the reversible neural network. Let be the secret image input to the i-th layer of the invertible neural network. The image represents the carrier image output by the i-th layer of the reversible neural network. Let be the secret image output by the i-th layer of the invertible neural network. It represents the Hadamah accumulation. , , For a multilayer invertible neural network, it is a nonlinear function. This is the Sigmoid activation function.

[0021] S2: Employ a residual augmentation strategy on the initial stegana. The final steganalysis image is obtained through processing. ; Final Stegated Image The calculation formula is as follows: , in, The coefficients are randomly sampled from a uniform distribution [0,1]. It represents the Hadamardi (or Hadama) stack.

[0022] The initial stegana was augmented using a residual augmentation strategy. With carrier image The differences are randomly perturbed to simulate the steganalysis patterns of different steganalysis methods, so that the generated final steganalysis image has a more diverse feature distribution during training, thereby improving the steganalysis performance in cross-dataset and cross-steganalysis method scenarios and significantly improving the generalization ability.

[0023] S3: Final Steganographic Image After performing discrete wavelet transform and random Gaussian noise The input is combined with a multilayer reversible neural network for inverse reconstruction. The multilayer reversible neural network outputs the reconstruction result, which is then subjected to inverse discrete wavelet transform to obtain the secret image. ; Taking the i-th layer of the invertible neural network as an example, its iterative formula is: , , in, The noise is the inverse output of the i-th layer of the reversible neural network. The noise is the inverse input to the i-th layer of the reversible neural network. The image is the reverse output of the i-th layer of the reversible neural network. The image is the inverse input to the i-th layer of the reversible neural network. Carrier image After performing discrete wavelet transform and random Gaussian noise The images are input into a multilayer reversible neural network for inverse reconstruction. The network outputs the reconstruction result, which is then subjected to inverse discrete wavelet transform to obtain the carrier image. Random Gaussian noise Random Gaussian noise They are different noises sampled from the same Gaussian distribution; Taking the i-th layer of the invertible neural network as an example, its iterative formula is: , , in, The noise is the inverse output of the i-th layer of the reversible neural network. The noise is the inverse input to the i-th layer of the reversible neural network. The image is the reverse output of the i-th layer of the reversible neural network. The image is the inverse input of the i-th layer of the reversible neural network.

[0024] S4: Calculate the overall loss function of the multilayer invertible neural network and update the parameters of the multilayer invertible neural network based on gradient descent (e.g., Adam optimization algorithm).

[0025] The formula for the overall loss function is as follows: , , , , , in, For the overall loss function, , , , For hyperparameters, Indicates mean square error. This indicates the low-frequency wavelet subband extraction operation.

[0026] Represents the final steganalyte image With carrier image Loss of similarity between them Represents the final steganalyte image Carrier image Low-frequency subband consistency loss after wavelet transform This indicates the loss in secret image recovery. This represents the loss in carrier image recovery. During training, a balanced optimization was performed on three aspects: steganalyte quality, secret map recovery, and carrier image recovery. This allows the method in this embodiment to achieve highly covert image hiding while effectively recovering steganalyte information and accurately performing steganalysis.

[0027] S5: Repeat steps S1 to S4 until the multilayer reversible neural network converges and training is complete.

[0028] S6: The image to be analyzed is subjected to discrete wavelet transform and then trained with random Gaussian noise input to form a multilayer reversible neural network. The multilayer reversible neural network outputs the recovery result, which is then subjected to inverse discrete wavelet transform to obtain the recovered image. The recovered image is used to determine whether the image to be analyzed is a steganalysis image.

[0029] When an image or random Gaussian noise is input into a multilayer invertible neural network for inverse reconstruction, the iterative formula for the i-th layer of the invertible neural network is as follows: , , in, The noise is the inverse output of the i-th layer of the reversible neural network. The noise is the inverse input to the i-th layer of the reversible neural network. The image is the reverse output of the i-th layer of the reversible neural network. The image is the inverse input of the i-th layer of the reversible neural network.

[0030] The method for determining whether an image to be analyzed is a steganalyte based on the restored image is as follows: Calculate the similarity between the restored image and the image to be analyzed. If the similarity is less than or equal to a set threshold, the image to be analyzed is a steganalyte; if the similarity is greater than the set threshold, the image to be analyzed is not a steganalyte. The similarity is expressed using peak signal-to-noise ratio or structural similarity index. The discriminant function is: , in, It is a binary discriminant function. For similarity measurement function, This represents a reverse invertible neural network, where X is the image to be analyzed and D is random Gaussian noise. The threshold is used. When the input image to be analyzed is a carrier image, the similarity between the restored image and the image to be analyzed is high; when the input image to be analyzed is a steganalyte, the similarity is significantly reduced.

[0031] In this scheme, a multi-layer reversible neural network is first used to process the secret image. Embedded carrier image This generates the initial steganalysis image. Then, the initial stegana was augmented using a residual augmentation strategy. With carrier image The differences are randomly perturbed to simulate the steganalysis patterns of different steganalysis methods, so that the generated final stegana image has a more diverse feature distribution during training. Then the final stegana image... Carrier image The method employs a multi-layer reversible neural network with random Gaussian noise input for reverse recovery. During training, a total loss function is used to balance and optimize the steganalysis image quality, secret image recovery, and carrier image recovery. This allows the method in this embodiment to effectively recover steganalytic information and accurately perform steganalysis while achieving highly concealed image hiding.

[0032] This method not only determines whether the input image to be analyzed is a steganalyte, but also partially reconstructs the hidden image, making the analysis results more intuitive and reliable, and improving the interpretability of the steganalysis process. This method eliminates the need to construct large-scale steganalyte-carrier image pairs during the training phase, simplifying the data preparation process and reducing dependence on the training dataset.

[0033] For example, in the training phase of a multilayer reversible neural network, the following methods are used: Figure 2 The animal type images shown are used as carrier images and secret images for training to obtain a trained multilayer reversible neural network; Figure 3 The secret image in the middle is the information hidden by the steganographic image above it. Figure 3 Three steganalytes are input into a multilayer reversible neural network for steganalysis. The multilayer reversible neural network outputs the corresponding restored images. It can be seen that the restored images have low similarity to the corresponding steganalytes, thus it can be determined that the images input into the multilayer reversible neural network are steganalytes. It can also be seen that the restored images partially restore the information of the secret images hidden in the steganalytes, verifying that the steganalysis results of this embodiment are intuitive and reliable, improving the interpretability of the steganalysis process, and the detection performance remains effective in cross-dataset and cross-method scenarios.

[0034] For example: The training dataset contains 30,000 images, from which image pairs are randomly sampled, each image pair consisting of one carrier image. And 1 secret image for hiding Carrier image Secret Images The resolution of all images is 512×512. These images are used to train a multi-layer reversible neural network, with the following settings: The value is 1.0. The value is 10.0. The value is 5.0. The value is 5.0. The multilayer reversible neural network is trained using steps S1 to S5 of this embodiment to obtain the trained multilayer reversible neural network. Step S6 is then used to perform steganalysis on images from the public datasets DIV2K, COCO, and ImageNet, respectively. Similarity is represented by peak signal-to-noise ratio, and the threshold is... The accuracy is 25.0, and the entire process is denoted as the ZSIIS method. Its steganalysis accuracy compared to other existing steganalysis methods on the public datasets DIV2K, COCO, and ImageNet is shown in Tables 1 to 3 below: Table 1. Steganalysis accuracy on the public dataset DIV2K

[0035] Table 2. Steganalysis accuracy on the public dataset COCO

[0036] Table 3. Steganography accuracy on the public ImageNet dataset

[0037] The Weng, HiNet, LiDiNet, StegFormert, and StegMamba methods in Tables 1 to 3 are existing single-image steganography methods, which hide one secret image into one carrier image. The DeepMIH and StegFormer methods are existing multi-image steganography methods. For example, the DeepMIH method hides three secret images into one carrier image, and the StegFormer method hides five secret images into one carrier image. The carrier image and secret image are taken from the corresponding public datasets.

[0038] The ZSIIS method is the method used in this embodiment. The XuNet, YeNet, SRNet, StegNet, ZhuNet, and SiaStegNet methods are existing steganalysis methods, which are trained using the same training samples. The training samples are obtained as follows: images randomly sampled during the training of the ZSIIS method are used to perform image steganalysis on the multilayer reversible neural network trained in the ZSIIS method. The initial steganalyst image output by the trained multilayer reversible neural network and its corresponding carrier image constitute the training samples.

[0039] As can be seen from Tables 1 to 3, the method in this embodiment achieves significantly higher steganalysis accuracy for stegated images generated by various existing image steganography methods than other existing steganalysis methods. For example, as shown in Table 3, on the public ImageNet dataset, for stegated images generated using the HiNet method, the highest accuracy of other existing steganalysis methods is only 63.06% for the StegNet method, while the steganalysis accuracy of the method in this embodiment reaches 87.28%, significantly higher than existing steganalysis methods. Tables 1 to 3 also show that the method in this embodiment maintains significantly higher steganalysis accuracy in multi-image hiding scenarios than other existing steganalysis methods, demonstrating excellent steganalysis performance.

[0040] The difference between ZSIIS w / o RA and ZSIIS is that ZSIIS does not use a residual augmentation strategy; otherwise, it is the same as ZSIIS. Specifically, the last row in Tables 1, 2, and 3 shows the steganalysis accuracy of the method in this embodiment without using a residual augmentation strategy. It can be seen that the generalization performance (especially in cross-method scenarios) is significantly reduced.

[0041] The XuNet, YeNet, SRNet, StegNet, ZhuNet, and SiaStegNet methods all require the pre-construction of large-scale training samples (i.e., stegimage-carrier image pairs) during the training phase. However, the method in this embodiment does not require the construction of training samples (i.e., zero samples) during the training phase, which simplifies the data preparation process and reduces the dependence on the training dataset.

[0042] The method in this embodiment demonstrates the effectiveness of steganalysis and output of restored images from stegographic images obtained using the Weng method, HiNet method, and LiDiNet method. Figure 4 As shown. Figure 4 In the image, the left side shows the carrier image and the secret image. The first row of three images on the right are steganalyst images obtained by embedding the secret image into the carrier image using the Weng method, HiNet method, and LiDiNet method. The second row of three images on the right are the recovered images output by steganalysis of the first row of three steganalyst images using the method of this embodiment. The recovery results show that, in a zero-shot scenario, the method of this embodiment can successfully recover part of the secret image from the steganalyst image, thus achieving zero-shot interpretable image steganalysis.

[0043] An electronic device according to this embodiment includes a memory and a processor. The memory stores executable code, and when the executable code is executed by the processor, the above-described method is performed.

Claims

1. An image steganalysis method based on a reversible neural network, characterized in that, Includes the following steps: S1: Transfer the carrier image and secret images Inputting the secret image into a multilayer reversible neural network, the multilayer reversible neural network will process the secret image. Embedded carrier image Output the initial steganalysis image ; S2: Employ a residual augmentation strategy on the initial stegana. The final steganalysis image is obtained through processing. ; S3: Final Steganographic Image Carrier image Each layer of reversible neural network is used to perform inverse reconstruction by inputting random Gaussian noise, and the reversible neural networks output the recovered secret images respectively. Carrier image ; S4: Calculate the overall loss function and update the parameters of the multilayer invertible neural network; S5: Repeat steps S1 to S4 until the multilayer reversible neural network converges and training is complete; S6: Input the image to be analyzed and random Gaussian noise into the trained multilayer reversible neural network. The multilayer reversible neural network outputs the restored image. Based on the restored image, determine whether the image to be analyzed is a steganalysis image.

2. The image steganalysis method based on a reversible neural network according to claim 1, characterized in that, Step S1 includes the following steps: Carrier image Secret Images Discrete wavelet transform is performed on the carrier image respectively. Secret Images The images are decomposed into low-frequency and high-frequency subbands, then input into a multi-layer reversible neural network. The multi-layer reversible neural network then processes the secret image. Embedded carrier image The output of the multilayer reversible neural network is transformed by inverse discrete wavelet transform to obtain the initial steganalysis image. .

3. The image steganalysis method based on a reversible neural network according to claim 1, characterized in that, The final steg image The calculation formula is as follows: ， in, The coefficients are randomly sampled from a uniform distribution [0,1]. It represents the Hadamardi (or Hadama) stack.

4. The image steganalysis method based on a reversible neural network according to claim 1, characterized in that, Step S3 includes the following steps: The final steganographic image After performing discrete wavelet transform and random Gaussian noise The input is combined with a multilayer reversible neural network for inverse reconstruction. The multilayer reversible neural network outputs the reconstruction result, which is then subjected to inverse discrete wavelet transform to obtain the secret image. ; Carrier image After performing discrete wavelet transform and random Gaussian noise The images are input into a multilayer reversible neural network for inverse reconstruction. The network outputs the reconstruction result, which is then subjected to inverse discrete wavelet transform to obtain the carrier image. .

5. The image steganalysis method based on a reversible neural network according to claim 4, characterized in that, The random Gaussian noise Random Gaussian noise These are different noises sampled from the same Gaussian distribution.

6. The image steganalysis method based on a reversible neural network according to claim 1, characterized in that, When an image or random Gaussian noise is input into a multilayer invertible neural network for inverse reconstruction, the iterative formula for the i-th layer of the invertible neural network is as follows: ， ， in, The noise is the inverse output of the i-th layer of the reversible neural network. The noise is the inverse input to the i-th layer of the reversible neural network. The image is the reverse output of the i-th layer of the reversible neural network. The image is the inverse input of the i-th layer of the reversible neural network. , , For a multilayer invertible neural network, it is a nonlinear function. This is the Sigmoid activation function.

7. The image steganalysis method based on a reversible neural network according to claim 1, characterized in that, The formula for the overall loss function is as follows: ， ， ， ， ， in, For the overall loss function, , , , For hyperparameters, This represents the mean square error. This indicates the low-frequency wavelet subband extraction operation.

8. The image steganalysis method based on a reversible neural network according to claim 1, characterized in that, The method for determining whether an image to be analyzed is a steganalysis image based on the restored image is as follows: calculate the similarity between the restored image and the image to be analyzed. If the similarity is less than or equal to a set threshold, it indicates that the image to be analyzed is a steganalysis image. If the similarity is greater than the set threshold, it indicates that the image to be analyzed is not a steganalysis image.

9. The image steganalysis method based on a reversible neural network according to claim 8, characterized in that, The similarity is expressed using peak signal-to-noise ratio or structural similarity index.

10. An electronic device, characterized in that, It includes a memory and a processor, wherein executable code is stored in the memory, and when the executable code is executed by the processor, the method described in any one of claims 1-9 is performed.

Images