Image steganography method based on cavity convolution

Abstract

The invention discloses an image steganography method based on cavity convolution, which comprises the following steps: selecting a carrier image, and determining secret information; carrying out cross-scale feature fusion on the carrier image and the secret information through an encoder which introduces a multi-stage cavity convolution module, and generating a secret-containing image; using an improved steganalysis discriminator to carry out adversarial training on the secret-containing image, optimizing statistical distribution of the secret-containing image, and obtaining a probability that the secret-containing image is a real image; and decoding the secret-containing image through a decoder to obtain recovered information. According to the method, the high embedding amount and the high coding picture quality can be ensured, meanwhile, the higher noise attack resistance is achieved, and the problem that the robustness based on a coding-decoding network model is poor is effectively solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information hiding, and particularly relates to an image steganography method based on a cavity convolution. BACKGROUND

[0002] As a core technology in the field of information security, image steganography is committed to realizing the covert transmission of confidential data by minimizing the modification amplitude of the carrier image. The core goal is to embed digital information into ordinary images without trace, ensuring the security of covert communication while only allowing authorized recipients to accurately extract hidden information. This technology has been widely used in digital watermarking, copyright tracing authentication, and sensitive information covert storage scenarios. Compared with traditional encryption technology, steganography hides information without visual perception, which can not only avoid attention risks in the transmission process, but also maintain the apparent authenticity of the carrier, forming a unique "hidden but not secret" security paradigm.

[0003] Early non-adaptive steganography methods ignore image texture features, causing significant damage to the statistical characteristics of the carrier, and are easily detected by steganalysis tools. With in-depth research, adaptive steganography technology effectively improves statistical undetectability by concentrating modifications on image edges or complex texture areas, becoming the mainstream research direction in the past decade. Typical methods are based on additive distortion functions and non-additive distortion models, combining the Syndrome-Trellis Codes (STC) framework to realize distortion minimization embedding. For color images, scholars further explore load allocation strategies under inter-channel correlation constraints, but traditional methods still face the inherent contradiction between steganographic capacity and detection resistance performance. In recent years, the rise of deep learning technology, especially the generative adversarial network (GAN), has provided a new paradigm for steganography - through the adversarial game between the generator and the discriminator, both the distribution of real images can be simulated, and the steganalysis attack can be resisted. However, existing GAN-based steganography schemes have significant bottlenecks in image quality, detection resistance, and computational efficiency, such as the insufficient ability of traditional convolutional structures to capture multi-scale features, making it difficult to effectively disperse steganographic traces. SUMMARY

[0004] To solve the above technical problems, the application provides an image steganography method based on a cavity convolution, which can ensure high embedding capacity, high encoded image quality, and stronger resistance to noise attacks, effectively solving the problem of weak robustness based on the coding-decoding network model.

[0005] To achieve the above purpose, the application provides an image steganography method based on a cavity convolution, comprising:

[0006] Selecting a carrier image and determining secret information;

[0007] The carrier image and the secret information are fused through a multi-level cavity convolution module to generate a stego image;

[0008] The stego image is subjected to adversarial training by using an improved steganalysis discriminator to optimize the statistical distribution of the stego image and obtain a probability that the stego image is a real image;

[0009] The stego image is decoded by a decoder to obtain recovered information.

[0010] Optionally, the execution process of the multi-level cavity convolution module comprises:

[0011] The first-level cavity convolution adopts an expansion rate of 2, the second-level cavity convolution adopts an expansion rate of 3, and the third-level cavity convolution adopts an expansion rate of 1.

[0012] The output of each level is spliced with the secret information and then input into the next convolution.

[0013] Optionally, the generation of the stego image comprises:

[0014]

[0015] wherein p is a feature map of the carrier image after convolution processing, M is a binary secret information tensor, Conv is a convolution operation, Cover is a carrier image, O1 is the output after the first cavity convolution, O2 is the output after the second cavity convolution, O3 is the output after the third cavity convolution, Stego is a stego image, C is an RGB carrier image tensor, D is the number of binary bits that can be embedded in each pixel, and Cat is a feature splicing operation.

[0016] Optionally, the improved steganalysis discriminator comprises an RGB input adaptation unit, a residual feature enhancement unit, and a multi-scale feature fusion unit.

[0017] The RGB input adaptation unit is configured to process a color carrier image through a Gray operation followed by SRM and KV high-pass filtering.

[0018] The residual feature enhancement unit is configured to extract noise residuals in the image by using a high-pass filter set built in the XuNet.

[0019] The multi-scale feature fusion unit is configured to capture steganographic trace features at different scales by using a hierarchical convolution and pooling structure.

[0020] Optionally, the adversarial training of the stego image by using the improved steganalysis discriminator comprises:

[0021] The parameters of the encoder and the decoder are frozen, and the parameters of the discriminator are updated preferentially.

[0022] Alternately optimize the encoder / decoder and the discriminator, and perform two discriminator updates after each encoder update.

[0023] Optionally, obtaining the probability that the stego image is a real image comprises:

[0024]

[0025] wherein S g is the image after gray processing; Gray is the RGB-to-gray operation; f0 is the feature map extracted after the filtering operation; f1 is the activation output; f2 is the first pooling output; f3 is the second pooling output; p is the steganography probability; SRM is the Spatial Rich Models filter; JV is the Kim-Van filter; Tanh is the activation function; ABS is the absolute value activation function; and AvgPool is the average pooling; is the full connection operation; Conv is the convolution operation; Flatten is the operation of converting multi-dimensional features into 1-dimensional features; and sigma is the Sigmoid function, outputting the classification probability.

[0026] Optionally, the stego image is decoded by the decoder to obtain the recovered information, comprising:

[0027]

[0028] wherein O1 is the output after the first dilated convolution; O2 is the output after the second dilated convolution; O3 is the output after the third dilated convolution; Conv is the convolution operation; and Stego is the steganography image. is the extracted secret message; and sigma is the Sigmoid function, outputting the classification probability.

[0029] Optionally, the secret information is adjusted to match the spatial size of the carrier image by padding or repeating operation before embedding.

[0030] Technical effects of the present application: The present application discloses an image steganography method based on dilated convolution, which uses dilated convolution to deeply extract image features, optimizes the loss function to improve the network training performance, and makes the model after adding the noise layer have stronger decoding capability. The present application can ensure high embedding capacity and high coding picture quality while having stronger resistance to noise attack capability, effectively solving the problem of weak robustness of the coding-decoding network model. Future work can focus on mixed training of various noises to adapt to the application scenario where the steganography image is subjected to multiple noise attacks in real scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0031] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of this application, and of the description of these embodiments, are used to explain the application and are not intended to limit the application in any way. In the drawings:

[0032] Figure 1 A flowchart of an image steganography method based on a cavity convolutional network according to an embodiment of the present application;

[0033] Figure 2 A comparison diagram of a carrier image and a stego image of different objects according to an embodiment of the present application, wherein (a) is a carrier image and (b) is a stego image. DETAILED DESCRIPTION

[0034] It should be noted that the embodiments and features of the embodiments in the present application can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0035] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0036] As shown in the accompanying drawings, Figure 1 The present embodiment provides an image steganography method based on a cavity convolutional network, which comprises:

[0037] Selecting a carrier image and determining secret information;

[0038] Performing cross-scale feature fusion on the carrier image and the secret information by introducing a multi-level cavity convolutional module encoder to generate a stego image;

[0039] Performing adversarial training on the stego image using an improved steganalysis discriminator to optimize the statistical distribution of the stego image and obtain the probability that the stego image is a real image;

[0040] Decoding the stego image through a decoder to obtain recovered information.

[0041] The present application enhances feature representation capability by introducing a multi-scale cavity convolutional module, combines an improved XuNet steganalysis network to construct a discriminator, and realizes global diffusion and high security embedding of steganographic traces.

[0042] Further, the encoder network is denoted as Enc, the input includes an RGB cover image tensor C with shape 3 x W x H and a binary secret message tensor M e {0, 1}, where the secret message needs to be adjusted to the dimension D x W x H by padding or repeating operation, where D represents the number of binary bits that can be embedded in each pixel as preset by the model, so as to ensure that the secret data can be evenly distributed in each pixel position of the image and completely matched with the cover image in the spatial structure. The process of message embedding of the encoder is shown in formula (1),

[0043]

[0044] wherein p is the feature map of the cover image after convolution processing; M is the binary secret message tensor; Conv is the convolution operation; Cover is the cover image; O1 is the output after the first dilated convolution; O2 is the output after the second dilated convolution; O3 is the output after the third dilated convolution; Stego is the stego image; C is the RGB cover image tensor; D is the number of binary bits that can be embedded in each pixel; and Cat is the feature concatenation operation.

[0045] Further, the decoder network is denoted as Dec, which receives the stego image Stego generated by the encoder and gradually recovers the hidden binary message M through multi-level feature extraction and fusion. The process of message extraction of the decoder is shown in formula (2),

[0046]

[0047] wherein O1 is the output after the first dilated convolution; O2 is the output after the second dilated convolution; O3 is the output after the third dilated convolution; Conv is the convolution operation; Stego is the stego image with stego traces dispersed; and Cat is the feature concatenation operation. is the extracted secret message; and σ is the Sigmoid function, outputting the classification probability.

[0048] Further, the discriminator network is denoted as Dis, which receives the input image I (which can be a stego image or a natural image), extracts discriminative features through multi-layer convolution and feature compression operation, and finally outputs a scalar value p e [0, 1], representing the probability that the input image is a real image. The discrimination process of the discriminator is shown in formula (3),

[0049]

[0050] wherein S gis the image after the grayscale processing; Gray is the RGB-to-grayscale operation; f0 is the feature map extracted after the filtering operation; f1 is the activation output; f2 is the first pooling output; f3 is the second pooling output; p is the steganography probability; SRM is the Spatial Rich Models filter; KV is the Kim-Van filter; Tanh is the activation function; ABS is the absolute value activation function; and AvgPool is the average pooling. is the full connection operation; Conv is the convolution operation; Flatten is the conversion of multi-dimensional features into one-dimensional features; and σ is the Sigmoid function, which outputs the classification probability.

[0051] Specifically, SRM and KV represent the convolution operations of 30 SRM filter kernels and 12 KV kernels, respectively, and the output channel number is 42. ABS represents the absolute value activation function, which enhances the contrast of steganographic noise. AvgPool is the average pooling. σ is the Sigmoid function, which outputs the classification probability.

[0052] By jointly optimizing the loss function, a balance between high steganographic capacity, high information recovery accuracy, and strong anti-detection ability is achieved. The joint loss function of the encoder and the decoder can be subdivided into: similarity loss The purpose is to ensure that the steganographic image and the original carrier are highly consistent at the pixel level and avoid visual distortion. By constraining the modification range of the encoder on the carrier, the natural statistical characteristics of the image are maintained, as shown in equation (4); accuracy loss The purpose is to maximize the consistency between the recovered secret information of the decoder and the original information, and to measure the bit-level error through binary cross-entropy, thereby optimizing the reliability of information extraction, as shown in equation (5); security loss The purpose is to counter steganalysis tools and make the steganographic image approach the original carrier in statistical distribution. The encoder improves the anti-detection ability by minimizing the discrimination probability of the discriminator, as shown in equation (6).

[0053]

[0054] Therefore, the loss function when the encoder and the decoder are jointly optimized is as shown in equation (7):

[0055]

[0056] where λ1, λ2, and λ3 represent weight hyperparameters.

[0057] Discriminator loss Based on the Wasserstein distance, the discriminator is driven to distinguish between steganographic images and real carriers, while providing an adversarial gradient for the encoder, forming a dynamic game, as shown in equation (8),

[0058]

[0059] Further, the model of the present application is a ternary collaborative optimization model composed of an encoder, a decoder and a discriminator, and the training of the model adopts a dynamic alternating optimization strategy to realize multi-objective balance through the parameter update of the encoder, the decoder and the discriminator. In the training process, the encoder and the decoder are first frozen, and the parameter of the discriminator is updated, and batch real carrier images and stego images generated by the encoder are used as input to maximize the Wasserstein distance loss to improve the ability of the discriminator to distinguish stego images and real carriers; then the parameters of the discriminator are frozen, and the encoder and the decoder are optimized, and the total loss of the encoder is minimized to simultaneously improve the visual fidelity, information recovery accuracy and anti-detection ability of the stego image. To balance the stability of the adversarial training, the discriminator is updated twice after each encoder / decoder update to prevent the model from collapsing due to the over-strength of the discriminator, and L2 regularization (weight decay coefficient 5x10-4 5x10-4) is applied to the parameters of the discriminator to suppress the risk of overfitting. Through the alternating optimization and constraint mechanism, the model can still maintain low detection rate and high image quality under high steganographic capacity.

[0060] One specific application example of the present application:

[0061] This embodiment is built based on the PyTorch 2.0.0 deep learning framework, the hardware platform is configured with NVIDIA RTX4090 GPU, the CUDA version is upgraded to 12.2 to fully release the hardware acceleration potential, and the software environment relies on Python 3.10 and cuDNN 8.8.0. The experimental data is selected from the MS COCO 2017 dataset, 2,000 high-resolution images are randomly selected from it to construct the training set, 500 images are used as the test set, and a stratified sampling strategy is used to ensure the balanced class distribution. All input images are preprocessed by standardization: first, the original image is uniformly scaled to 512x512 resolution, then random horizontal flipping (probability 0.5), color jittering (brightness = 0.2, contrast = 0.15, saturation = 0.1) and normalization (mean = [0.485, 0.456, 0.406], standard deviation = [0.229, 0.224, 0.225]) are applied.

[0062] The imperceptibility of steganography as its core performance indicator embodies the balance ability of covert communication system between information hiding and carrier fidelity. This feature requires that the carrier image (i.e. stego image) after steganography processing should not produce perceptible abnormal changes in its physical characteristics while maintaining the integrity of the original visual semantics. Specifically, the realization of imperceptibility needs to meet double constraints: on the one hand, at the subjective visual level, human observers cannot identify the difference between the stego image and the original carrier by naked eye; on the other hand, at the objective quality level, steganography operation should not lead to significant degradation of quantifiable indicators such as image structural features and statistical distribution. Therefore, the invention evaluates the imperceptibility of the proposed method from both qualitative and quantitative angles.

[0063] Qualitative evaluation: As shown in Fig. 1, (a) is the carrier image, and (b) is the stego image, and the obtained stego image has no obvious perceptible difference in vision. Figure 2

[0064] Quantitative test: In order to further verify the quality of the stego image, the Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity (SSIM) are used to quantitatively evaluate the imperceptibility of the stego image.

[0065] The PSNR is used to measure the distortion degree of the stego image, and the formula is as follows:

[0066]

[0067] In the formula, X and Y are two images with the size of CxWxH, and MAX I is the maximum possible pixel value of the image. The larger the PSNR value, the smaller the distortion.

[0068] The Structural Similarity (SSIM) is based on three comparison metrics between two samples: brightness, contrast and structure. The SSIM value of X and Y of two images is calculated as follows:

[0069]

[0070] In the formula, μ X , μ Y represent the average value of X and Y, respectively, σ X , σ Y represent the standard deviation of X and Y, respectively, σ XY is the variance of X and Y, and c1 and c2 are constants, usually c2 is set to 0.01 by default, and c2 is set to 0.03 to avoid the case that the denominator is 0.

[0071] ​The visual quantitative test result of the scheme of the present application is that when the embedding amount is 1bpp, the PSNR value is 39.1DB, and the SSIM is 0.988, which is obviously improved compared with the existing scheme.

[0072] The steganographic images generated by the present application are detected by using mainstream steganalysis tools based on deep neural networks (SRNet and YeNet). The test set contains 2,000 steganographic images from the COCO dataset and an equal amount of original images, and the anti-detection performance is evaluated by the detection error rate (PE) and the area under the ROC curve (AUC). As in the case of embedding capacity of 1bpp, the method of the present application is superior to SteganoGAN and its variant scheme.

[0073] In order to explore the effectiveness of the added cavity convolution and the steganalysis discriminator based on XuNet in the scheme of the present application, an ablation experiment is carried out by changing the configuration in the proposed scheme. Then four different versions of the scheme are obtained, as follows: (1) Scheme one: remove the cavity convolution and the steganalysis discriminator based on XuNet; (2) Scheme two: remove the cavity convolution; (3) Scheme three: remove the steganalysis discriminator based on XuNet; (4) Scheme four: the original scheme, that is, the steganographic scheme with cavity convolution and the steganalysis discriminator based on XuNet. Each different version is trained with an embedding capacity of 1bpp, and their performance in image quality and extraction accuracy is tested. Table 1 shows the ablation experiment results, and the two modules of cavity convolution and steganalysis discriminator based on XuNet in the scheme of the present application play a very important role in improving the overall steganographic performance.

[0074] Table 1

[0075]

[0076] In view of the deficiencies of existing image steganography methods based on generative adversarial network (GAN) in security, image quality and embedding capacity, and in particular the problem that the limited receptive field of traditional convolutional structure leads to insufficient diffusion of steganographic traces and difficulty in effectively resisting advanced steganalysis, the present application proposes and realizes the core innovation. One of the core innovations is the fundamental enhancement of the encoder network structure. The present application innovatively introduces dilated convolution into the encoder design to solve the problem that the limited receptive field of traditional convolutional neural network (CNN) makes it difficult to effectively capture the information of complex texture and edge regions widely distributed in the image, resulting in local concentration of secret information embedding traces and easy detection. By constructing multi-level dilated convolution modules and combining different dilation rates (such as dil = 1, 2, 3), the receptive field of feature extraction is significantly expanded without increasing the parameter amount or resolution, realizing cross-scale context perception. This design enables secret information to be more effectively "diffused" to the more extensive and perceptually insensitive areas of the carrier image (especially the texture-rich and edge regions), significantly improving the encoder's ability to embed information in complex regions, the concealment of embedding operations and the resistance to statistical analysis, laying a key structural foundation for subsequent anti-steganalysis.

[0077] Another core innovation is the specialized reconstruction of the discriminator to improve the ability to resist advanced steganalysis. In view of the problem that ordinary discriminators are difficult to capture extremely weak statistical anomalies and noise residual characteristics introduced by steganography, the present application directly integrates and improves an advanced steganalysis network, XuNet, as the discriminator (Dis) in the GAN framework. Key improvements include: adapting to RGB input (optimizing the front-end processing flow by using Gray operation followed by SRM and KV high-pass filtering to effectively process color carrier images), strengthening residual feature extraction (making full use of the rich high-pass filter set (SRM, KV) built-in XuNet to extract subtle noise residuals in the image, which is a key clue for detecting steganographic traces), and multi-scale feature fusion (using its hierarchical convolution and pooling structure to capture steganographic trace features at different scales). This successfully builds a highly specialized and extremely sensitive discriminator for steganographic trace detection. Its powerful multi-scale noise residual feature capture capability provides the generator (encoder) with more accurate and challenging counter signals, forcing the encoder to generate steganographic images with statistical characteristics closer to natural images, thereby significantly improving the security of the overall scheme against advanced steganalysis such as SRNet and YeNet.

[0078] The above merely provides the preferred embodiments of the present application, and the protection scope of the present application is not limited thereto, and any changes or substitutions within the technical scope disclosed by the present application should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An image steganography method based on dilated convolution, characterized in that, include: Select a carrier image to determine the secret information; An encoder that incorporates multi-level dilated convolution modules performs cross-scale feature fusion of the carrier image and secret information to generate a secret image; An improved steganalysis discriminator is used to perform adversarial training on the dense image, optimize the statistical distribution of the dense image, and obtain the probability that the dense image is a real image; The encrypted image is decoded using a decoder to obtain the recovered information.

2. The image steganography method based on dilated convolution as described in claim 1, characterized in that, The execution process of the multi-level dilated convolution module includes: The first-level dilated convolution uses a dilation rate of 2, the second-level dilated convolution uses a dilation rate of 3, and the third-level dilated convolution uses a dilation rate of 1. The output of each level is concatenated with the secret information and then input into the next level of convolution.

3. The image steganography method based on dilated convolution as described in claim 1, characterized in that, Generating a dense image includes: Where p is the feature map of the carrier image after convolution; M is the binary secret information tensor; Conv is the convolution operation; Cover is the carrier image; O1 is the output after the first dilated convolution; O2 is the output after the second dilated convolution; O3 is the output after the third dilated convolution; Stego is the dense image; C is the RGB carrier image tensor; D is the number of binary bits that can be embedded in each pixel; and Cat is the feature concatenation operation.

4. The image steganography method based on dilated convolution as described in claim 1, characterized in that, The improved steganalysis discriminator includes: an RGB input adaptation unit, an enhanced residual feature extraction unit, and a multi-scale feature fusion unit; The adaptive RGB input unit is used to process the color carrier image by SRM and KV high-pass filtering after Gray operation; The enhanced residual feature extraction unit is used to extract noise residuals in the image using the high-pass filter bank built into XuNet; The multi-scale feature fusion unit is used to capture steganalytic trace features at different scales using hierarchical convolution and pooling structures.

5. The image steganography method based on dilated convolution as described in claim 1, characterized in that, Adversarial training of the dense image using an improved steganalysis discriminator includes: Freeze encoder and decoder parameters, and prioritize updating discriminator parameters; The encoder / decoder and discriminator are optimized alternately, and the discriminator is updated twice after each encoder update.

6. The image steganography method based on dilated convolution as described in claim 1, characterized in that, The probability of obtaining a dense image as a real image includes: Among them, S g f0 is the image after grayscale conversion; f1 is the feature map extracted after filtering; f2 is the first pooling output; f3 is the second pooling output; p is the steganalysis probability; SRM is the Spatial Rich Models filter; KV is the Kim-Van filter; Tanh is the activation function; ABS is the absolute value activation function; AvgPool is the average pooling function. is a fully connected operation; Conv is a convolution operation; Flatten converts multi-dimensional features into 1-dimensional features; σ is the Sigmoid function, which outputs the classification probability.

7. The image steganography method based on dilated convolution as described in claim 1, characterized in that, The encrypted image is decoded using a decoder to obtain the recovered information, including: Where O1 is the output after the first dilated convolution; O2 is the output after the second dilated convolution; O3 is the output after the third dilated convolution; Conv is the convolution operation; Stego is the steganographic image; and Cat is the feature concatenation operation. σ represents the extracted secret message; σ is the Sigmoid function, which outputs the classification probability.

8. The image steganography method based on dilated convolution as described in claim 1, characterized in that, The secret information is adjusted to match the spatial size of the carrier image through padding or repetition operations before embedding.

Images