Self-adaptive steganography method, system and device for spatial domain image content

Abstract

The invention belongs to the technical field of information security, and discloses a spatial domain image content adaptive steganography method, system and device, and the method comprises the steps: constructing a high-pass quaternion filter, and carrying out the filtering operation of a carrier image; calculating the residual error of the high-pass quaternion filter, and measuring the image texture complexity based on the residual error; constructing a steganography distortion cost function QMP based on quaternion amplitude and phase representation; and embedding image texture information into a distortion cost function to realize content-adaptive image steganography. According to the method, the image texture is analyzed more finely and embedded into the distortion cost function, so that the image information is hidden more effectively, the security of the image is enhanced, and the image steganography performance is remarkably improved. Experimental results show that the steganography distortion cost function provided by the invention can effectively capture the texture features and promote the secret information to be embedded into the texture region of the carrier image, so that the statistical undetectability is improved, and an important technical breakthrough is brought to the field of information security.

Description

Technical Field

[0001] The present invention belongs to the field of information security technology, and in particular relates to a spatial domain image content adaptive steganography method, system and device. Background Art

[0002] Spatial domain image content adaptive steganography is an important information hiding technology that aims to embed information into the carrier image while maintaining the visual quality of the image, while maximizing the security and concealment of the steganography. To achieve these goals, the design of the embedding distortion cost function is one of the key technologies. Existing design schemes mainly include the following three categories:

[0003] (1) Steganography methods based on heuristic manual design; Typical spatial image steganography methods based on heuristic manual design include HUGO (Highly Undetectable stego), WOW (Wavelet Obtained Weights), S-UNIWARD (Spatial Universal Wavelet Relative Distortion), and HiLL (High-pass filter, Low-pass filter, and Low-pass filter) methods. Specifically, the HUGO algorithm first uses the weighted sum of the SPAM (Subtractive Pixel Adjacency Matrix) feature differences between the carrier image and the encrypted image to heuristically design the distortion cost function; the WOW algorithm uses the weighted residual of the Daubechies 8-tap waveletdirection filter bank (DB-8) to adaptively design the distortion cost function; the S-UNIWARD algorithm uses the sum of the relative changes of the wavelet coefficients between the encrypted image and the carrier image in the directional wavelet filter subband to improve the WOW distortion cost function; inspired by WOW, the HiLL algorithm uses a high-pass filter and a low-pass filter to extract the image texture residual based on experience, and then heuristically defines an embedding distortion cost function with higher security. In addition, the HiLL algorithm also applies the cost spreading (CS) criterion to distribute the cost of pixels to adjacent pixels, thereby achieving better security performance than WOW, S-UNIWARD, etc. Such schemes use heuristic methods to construct the embedding distortion cost function. These schemes are simple in design, low in computational cost, and do not rely on specific data sets or hardware facilities, so they are widely used. However, such methods still have room for improvement in their ability to resist detection by steganalyzers, especially for some high-precision steganalyzers.

[0004] (2) Model-driven steganography methods: Typical model-driven spatial image content adaptive steganography methods include MVGG (Multivariate Generalized Gaussian), MiPOD (Minimizing the Power of Optimal Detector), and GMRF (Gaussian Markov Random Field). Specifically, the MVGG algorithm uses an independent Gaussian distribution model to model the pixels of the carrier image, and then calculates the embedding modification probability based on minimizing the KL (Kullback-Leibler) divergence between the carrier image and the carrier image, and reversely derives the embedding distortion cost; the MiPOD algorithm is based on the MVGG algorithm and further introduces the hypothesis testing principle. According to the maximum likelihood estimation algorithm, the distribution of the carrier image pixels is predicted and its corresponding parameters are estimated, thereby obtaining a more appropriate embedding modification probability; the GMRF algorithm introduces a four-element cross-neighborhood Gaussian Markov random field model to capture the dependency between adjacent pixels in the spatial domain, thereby better modeling the image pixel distribution. This type of steganography method can derive the embedding modification probability based on the theoretical model in mathematical statistics, and then reversely derive the distortion cost function, and then embed the secret information into the carrier image based on the minimum distortion framework. Although its theoretical derivation is rigorous and can show good security in some scenarios, its model is usually more complex and has a high computational cost. In addition, the security performance of these schemes is even worse than that of heuristic hand-designed methods under some steganalyzers.

[0005] (3) Image steganography schemes based on deep learning. Typical deep learning-based spatial domain image content adaptive steganography schemes include ASDL-GAN (Automatic Steganographic Distortion Learning with Generative Adversarial Networks), UT-GAN (U-Net, double-Tanh function, Generative Adversarial Networks), and SPARL (Steganographic Pixel-wise Actions and Rewards with Reinforcement Learning) methods. The ASDL-GAN algorithm is the first algorithm that attempts to apply deep learning technology - the generative adversarial network model to the field of image content adaptive steganography. The algorithm consists of a generator for generating the embedding modification probability of the carrier image, a simulated embedding simulator for generating the secret image, and a discriminator for classifying the carrier image and the secret image. When the adversarial optimization of the generator and the discriminator reaches a Nash equilibrium, the generator outputs the final embedding modification probability of the carrier image and converts it into the corresponding embedding distortion cost. Finally, the steganographic embedding is completed based on the minimum distortion framework. In order to solve the instability problem of the simulated embedding simulator in ASDL-GAN, the UT-GAN algorithm proposed a simulated embedding simulator based on the double Tanh activation function, and used a compact and efficient U-Net network structure to design the generator. At the same time, a discriminator network with stronger classification ability was used, which greatly improved the security of the steganographic algorithm compared with the ASDL-GAN, and obtained the best anti-statistical detection ability in the spatial domain image content adaptive steganography method at that time. The SPARL algorithm uses the reinforcement learning mechanism to further improve the simulated embedding simulator based on the UT-GAN algorithm, and achieves better anti-statistical detection ability than UT-GAN. Image content adaptive steganography based on deep learning can use the powerful learning ability of deep networks. For a given carrier image, it can learn a more appropriate embedding modification probability, and then derive the embedding distortion cost, and complete the steganographic embedding of secret information based on the minimum distortion framework. However, this type of method has significant defects: it requires a large amount of training data and computing resources, and the optimal model needs to be retrained each time the embedding load is embedded. The requirements for hardware facilities are high, and the computing cost is significantly increased, which is difficult to meet the needs of resource-constrained scenarios.

[0006] For spatial domain image content adaptive steganography, the design schemes of embedding distortion cost function mainly include: heuristic manual scheme, model-driven scheme and deep learning scheme. Typical spatial image steganography methods manually designed based on heuristic methods include HUGO, WOW, S-UNIWARD and HiLL. This scheme is widely studied by steganists because of its simple and effective design and is not restricted by data sets, hardware facilities, etc. Typical model-driven spatial image content adaptive steganography includes MVGG, MiPOD and GMRF. Although this scheme can use theoretical models in mathematical statistics to more rigorously derive the embedding distortion cost function, such models are usually more complex, and their anti-steganalyzer statistical detection ability (security) does not achieve the ideal effect. Even for some steganalyzers, the security performance is not as good as the heuristic manual scheme. Typical deep learning-based spatial image content adaptive steganography includes ASDL-GAN, UT-GAN and SPARL. This scheme achieves the best security against steganalysis detection at present, but the computational cost of this method is high. It not only requires a large amount of training data, but also trains an optimal model for each embedding load, and has high requirements for hardware. In view of this, the present invention follows the idea of ​​heuristic manual steganography scheme and designs an embedding distortion cost function with low computational complexity and high security, thereby realizing spatial domain image content adaptive steganography.

[0007] From the perspective of existing technologies, although deep learning-based methods have high security, their computational costs and hardware requirements limit their practical applications. Although model-driven steganography methods are more rigorous in theory, their complexity and limited improvement in security limit their practicality. In contrast, although heuristic manually designed steganography methods are not as good as deep learning methods in anti-detection capabilities, they are simple in design, low in computational cost, and have high practical value. However, existing heuristic methods still have room for improvement in how to adaptively optimize embedding strategies based on image content, especially in areas with high texture complexity.

[0008] The solution objectives of the present invention are as follows: for spatial domain image content adaptive steganography, the heuristic method is not secure enough; the model-driven method has high computational complexity and low security performance; the deep learning solution has a complex and cumbersome training process, and has high computational cost and hardware requirements. The present invention follows the technical idea of ​​heuristic manual design and proposes a spatial domain image content adaptive steganography method with low computational complexity and high security. In view of the problem of limited security improvement in the prior art, a distortion cost function based on quaternion amplitude and phase is designed. By utilizing the image texture complexity to achieve content adaptive optimization embedding, it not only improves the anti-detection ability of the steganography, but also reduces the computational complexity and hardware dependence, meeting the needs of efficient steganography in practical applications.

[0009] The solution to the actual technical problem in this aspect is:

[0010] (1) The problem of inaccurate image texture representation. Traditional methods are prone to bias in identifying complex texture areas, resulting in embedded information leakage. The present invention uses quaternions to represent texture complexity, thereby improving the protection capability for high-risk areas.

[0011] (2) The balance between steganographic efficiency and security. Although the deep learning solution has high security, it has high computational complexity and hardware requirements. Based on the heuristic solution, QMP achieves efficient embedding through lightweight design while ensuring security.

[0012] (3) Insufficient applicability and scalability. The introduction of GQMP solves the problem of flexibly balancing the security of steganographic algorithms in resisting different steganalyzers, and achieves wider coverage of application scenarios. Summary of the invention

[0013] In order to overcome the problems existing in the related art, the disclosed embodiment of the present invention provides a method, system and device for adaptive steganography of spatial domain image content, and the technical solution is as follows:

[0014] The present invention is implemented as follows: a spatial domain image content adaptive steganography method comprising:

[0015] S1, construct a high-pass quaternion filter to perform filtering operations on the carrier image;

[0016] S2, calculates the high-pass quaternion filter residual and measures the image texture complexity based on the residual;

[0017] S3, constructing a steganographic distortion cost function QMP based on quaternion amplitude and phase representation;

[0018] S4, embeds the image texture information into the distortion cost function to achieve content-adaptive image steganography.

[0019] In step S1, a high-pass quaternion filter is constructed, including:

[0020] Given a separable two-dimensional real domain filter F, it is expressed as:

[0021]

[0022] Where, f1, f2 are 1-dimensional real-valued filters, and T is the transpose;

[0023] High-pass quaternion filter F q It is expressed as:

[0024]

[0025] h1=H(f1)

[0026] h2=H(f2)

[0027] In the formula, F (i) is the 2D real domain filter corresponding to the imaginary unit i, i is the imaginary unit, F (j) is the 2D real domain filter corresponding to the imaginary unit j, j is the imaginary unit, F (k) is the 2D real domain filter corresponding to the imaginary unit k, where k is the imaginary unit, is the filter after the transposition of the 1-dimensional real filter f1, h1 is the 1-dimensional real domain filter after the 1-dimensional real filter f1 is subjected to the Hilbert transform, h2 is the 1-dimensional real domain filter after the 1-dimensional real domain filter f2 is subjected to the Hilbert transform, and H() is the Hilbert transform.

[0028] In step S2, the high-pass quaternion filter residual is calculated, and the image texture complexity is measured based on the residual, including:

[0029] Using a high-pass quaternion filter F q Perform convolution operation on the carrier image X, the expression is:

[0030]

[0031] In the formula, R q is the image residual, R is the real image residual, R (i) is the image residual corresponding to the imaginary unit i, R (j) is the image residual corresponding to the imaginary unit j, R (k) is the image residual corresponding to the imaginary unit k, F is a two-dimensional real domain filter, is the convolution operation;

[0032] The element with index (u,v) Then we have:

[0033]

[0034] In the formula, (u,v) is the index subscript, r u,v is the element corresponding to the index subscript (u,v) in the real image residual R, is the image residual R q The element corresponding to the index subscript (u,v) in the is the image residual R (i) The element corresponding to the index subscript (u,v) in the is the image residual R (j) The element corresponding to the index subscript (u,v) in the is the image residual R (k)The element corresponding to the index subscript (u,v);

[0035] The polar coordinate form is:

[0036]

[0037] In the formula, |||| is the element The modulus, θ u,v is the angle corresponding to the index subscript (u,v), μ is the imaginary unit;

[0038] Real part r u,v is decomposed into and cos(θ u,v ), then R q The real part of is decomposed into:

[0039] R=||R q ||ecos(Θ) (6)

[0040] In the formula, ||R q || is R q Each element The amplitude M, e is the Hadamard product, and cos(Θ) is the phase P;

[0041] The absolute value of the image residual R based on the high-pass filter ||R|| is used to represent the image texture complexity C:

[0042] C=|R|=||R q ||e cos(Θ)|=C M e C P (7)

[0043] In the formula, |R| is the absolute value of the image residual R, C M is the complexity of the carrier image determined by the amplitude, C P is the complexity of the carrier image determined by the phase;

[0044] C M and C P are proportional to |R|, and C M Inversely proportional to C P ;

[0045]

[0046] In the formula, For standardized C M , is the standardized |R|, max(C M ) is the amplitude complexity C M The maximum value, min(C M ) is the amplitude complexity C MThe minimum value in |R|, max(|R|) is the maximum value in |R|, and min(|R|) is the minimum value in |R|.

[0047] In step S3, a steganographic distortion cost function QMP based on quaternion amplitude and phase representation is constructed, including:

[0048] The image steganography distortion cost function ρ is inversely proportional to the complexity C, and the expression is:

[0049]

[0050] In the formula, ∞ is the symbol of proportional relationship, ρ M ,ρ P are the image distortion cost based on quaternion amplitude and quaternion phase, respectively. σ and ε are parameters to be determined and are set to constant values. M is the amplitude and P is the phase.

[0051] ρ M C M After taking the reciprocal, use the low-pass filter L for smoothing, the expression is:

[0052]

[0053] In the formula, C M is the amplitude complexity;

[0054] The parameter σ is used to adjust the diffusion effect of the distortion cost as the size of the low-pass filter L changes. P Perform smoothing and embed the distortion cost ρ P Defined as:

[0055]

[0056] Where ε = 10 -10 .

[0057] The embedding distortion cost function QMP is expressed as:

[0058]

[0059] Where cosθ is the cosine of the phase angle θ.

[0060] In step S4, the image texture information is embedded into the distortion cost function to achieve content-adaptive image steganography, including:

[0061] Construct a high-pass quaternion filter F q , used to perform filtering operations on the carrier image;

[0062] Apply the constructed high-pass quaternion filter to the carrier image and calculate the residual R after filteringq ;

[0063] The filtered quaternion residual R q Process and extract the real part And decompose it into two parts, let C M =||R q ||,C p =|cosθ| represents the complexity based on the image quaternion amplitude and phase respectively;

[0064] Calculate the embedding distortion cost ρ based on amplitude M and phase P M and P :Using the complexity priority criterion and the cost value diffusion criterion, we design the image steganography distortion cost function ρ M and P ;

[0065] Calculate the embedding distortion cost corresponding to the carrier image X: Heuristically define the embedding distortion cost function of the spatial domain image as ρ = ρ h eρ P .

[0066] The method also includes constructing a generalized QMP distortion cost function GQMP based on an exponential model, which is expressed as:

[0067]

[0068] By formula Decide, By formula The power exponents α and β are parameters to be determined and their values ​​are non-negative constants.

[0069] The method also includes performing image processing using a generalized QMP distortion cost function GQMP, specifically including:

[0070] Input embedding distortion cost ρ M , embedding distortion cost ρ P and steganalyzer A, and obtain the distortion cost ρ corresponding to the carrier image X;

[0071] Start the outer loop and perform loop iteration operation on the power index α0;

[0072] Start the inner loop and iterate the power index β0; if the condition is met, calculate the distortion cost And use the steganalyzer A to calculate the average detection error rate of the steganalysis scheme determined by ρ Then make a conditional judgment, if If established, execute α=α0, β=β0, otherwise the conditional judgment ends;

[0073] Add 1 to the power exponents α0 and β0 respectively, calculate the power exponents α0 and β0, and end the inner and outer loops;

[0074] Calculate the embedding distortion cost corresponding to the carrier image X

[0075] Returns the distortion cost ρ of the carrier image.

[0076] Another object of the present invention is to provide a spatial domain image content adaptive steganography system, which is used to control the spatial domain image content adaptive steganography method, and the system includes:

[0077] Image filtering module, used to construct a high-pass quaternion filter to perform filtering operations on the carrier image;

[0078] The residual image calculation module is used to apply the constructed quaternion filter to the carrier image, calculate the residual of the high-pass quaternion filter, and measure the image texture complexity based on the residual;

[0079] The information embedding module is used to construct a steganographic distortion cost function QMP based on quaternion amplitude and phase representation, embed the image texture information into the distortion cost function, and realize content-adaptive image steganography.

[0080] The present invention also provides a spatial domain image adaptive steganography device, which includes a processor, a memory, and a spatial domain image adaptive steganography control program stored in the memory and running on the processor. When the spatial domain image adaptive steganography control program is executed by the processor, a spatial domain image content adaptive steganography method is implemented.

[0081] The present invention also provides an information data processing terminal, which is used to implement a spatial domain image content adaptive steganography method.

[0082] Combining all the above technical solutions, the advantages and positive effects of the present invention are as follows:

[0083] First, the technical performance of the present invention is significantly improved. Experimental results show that QMP and GQMP are superior to existing methods in improving steganographic security, embedding efficiency and computational complexity (see experimental data chart for details). Reliability and stability in practical applications. The QMP design avoids dependence on complex models or training data, has higher stability, and can run efficiently in resource-constrained environments. A breakthrough contribution to the field of information security. The present invention not only solves multiple technical problems in the current field of image steganography, but also provides a solid theoretical and practical foundation for the development of future information hiding and protection technologies.

[0084] Second, for content-adaptive image steganography technology, existing solutions face the following challenges: heuristic methods have low computational complexity but insufficient security; although the model-driven method is rigorous in theory, it has high computational complexity; deep learning solutions provide better security, but the training process is complex and cumbersome. Following the design idea of ​​the heuristic solution, the present invention designs a distortion cost function QMP with high security based on quaternion texture representation. The previous design paradigm of content-adaptive steganography distortion cost function under the minimum distortion embedding framework shows that an efficient embedding distortion cost function depends to a large extent on the representation of image texture complexity. The present invention has taken a step forward by using quaternion representation, and the image texture complexity is decomposed into two parts: the complexity based on quaternion amplitude and the complexity based on quaternion phase. On this basis, the present invention empirically defines the embedding distortion cost ρ based on quaternion amplitude M and the embedding distortion cost ρ based on the quaternion phase P , and use ρ M and ρ P The Hadamard product between them obtains a new distortion cost function QMP.

[0085] Third, the present invention achieves more effective hiding of image information and enhances image security by more sophisticated analysis of image texture and embedding it into the distortion cost function. Secondly, the present invention uses image texture features represented by quaternions to improve the efficiency and accuracy of image processing, making the steganographic process more stable and reliable. Most importantly, this method achieves a significant improvement in image steganographic performance while maintaining image quality, bringing important technological breakthroughs to the field of information security. The implementation of the present invention will provide strong support for the future development of image steganography technology and have a far-reaching impact in the field of information security.

[0086] Fourth, the present invention realizes efficient and highly secure image steganography technology, reduces the demand for computing resources, and can be widely used in fields such as information transmission, digital copyright protection, and privacy protection, providing technical support for enterprises to occupy a leading position in the information security market. The present invention has a simple design and strong adaptability, and can be seamlessly integrated into existing image processing and communication systems, which helps enterprises to quickly deploy steganography technology and shorten the product launch cycle. Compared with deep learning-driven steganography technology, this solution has the advantage of low computational complexity, reduces hardware investment and operating costs, and provides affordable high-security solutions for small and medium-sized enterprises.

[0087] Fifth, the present invention introduces the joint representation of quaternion amplitude and phase in the field of image steganography for the first time, and proposes a distortion cost function QMP based on the Hadamard product. This innovative design is the first of its kind in the world, and provides a new perspective for the expression and processing of image texture complexity. Solve the problem of insufficient representation of texture features. The expression of texture complexity in the prior art generally relies on a single method, resulting in inaccuracy in the information embedding process. The present invention refines texture complexity into two dimensions, amplitude and phase, through quaternion representation, significantly improving the accuracy of texture feature capture. The traditional method has long had a bottleneck in the trade-off between performance and complexity. The present invention, based on the heuristic manual method, has broken through this bottleneck through technological innovation, opening up a new direction for the development of the industry.

[0088] Sixth, for a long time, it has been difficult for steganography to improve security and reduce computational complexity. The present invention designs the distortion cost function QMP by introducing the complexity based on quaternion amplitude and the complexity based on quaternion phase, successfully achieving a balance between security and complexity, and setting a new benchmark for image steganography. How to accurately embed embedded information into the texture area to maximize undetectability is a difficult problem in the industry. The present invention successfully achieves this goal by combining texture analysis of quaternion phase and amplitude, effectively improving the anti-detection ability of the steganography process. The present invention is based on heuristic design, does not rely on deep learning or highly complex statistical models, and makes up for the limitations of previous solutions in performance improvement through technical optimization. The present invention breaks through the limitations of traditional technical paths: traditional steganographic technology paths usually focus on the representation of a single texture dimension and ignore the potential of phase information. The present invention breaks through this prejudice through quaternion joint representation, showing new possibilities for complex texture information processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0090] Figure 1 is a flow chart of a spatial domain image content adaptive steganography method provided by an embodiment of the present invention;

[0091] Figure 2 C and C provided in the embodiment of the present invention M Schematic diagram of the statistical relationship between them; in each sub-graph, the circle represents the sampling point and the curve represents the corresponding fitting curve;

[0092] Figure 3 C and C provided in the embodiment of the present invention P Schematic diagram of the statistical relationship between them; in each sub-graph, the circle represents the sampling point and the curve represents the corresponding fitting curve;

[0093] Figure 4 C provided by the embodiment of the present invention M and C P Schematic diagram of the statistical relationship between them; in each sub-graph, the circle represents the sampling point and the curve represents the corresponding fitting curve;

[0094] Figure 5 is a flow chart of a QMP algorithm provided by an embodiment of the present invention;

[0095] Figure 6 is a flow chart of a GQMP algorithm provided by an embodiment of the present invention;

[0096] Figure 7 is a schematic diagram of a carrier image cover provided by an embodiment of the present invention;

[0097] Figure 8 is an embedding modification probability histogram corresponding to the cover provided by an embodiment of the present invention;

[0098] Fig. 9 is an embedding modification probability map corresponding to the carrier image cover provided by an embodiment of the present invention;

[0099] Fig.10 3 is a schematic diagram of embedding modification corresponding to the carrier image cover provided in an embodiment of the present invention; wherein a black dot represents -1 and a white dot represents +1. DETAILED DESCRIPTION

[0100] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0101] For content-adaptive image steganography, in view of the problems that the existing heuristic schemes have low computational complexity but low security, the model-driven methods are rigorous in theory but have high computational complexity, and the deep learning schemes have good security but cumbersome training, the present invention, based on the design ideas of the heuristic schemes, uses the quaternion representation of the residual of the carrier image filter, uses the amplitude value and phase based on the quaternion to characterize the image texture, and further uses it to evaluate the complexity of the image texture, and proposes a new distortion cost function method QMP. In addition, in order to further improve the security of the algorithm, the generalized form of QMP, namely GQMP, is proposed using the exponential model. The innovations of the present invention are mainly reflected in the following aspects:

[0102] (1) Image texture representation based on quaternions. This paper introduces quaternions into the texture representation of spatial image content adaptive steganography for the first time, and represents the complexity of image texture through the amplitude and phase of quaternions. Compared with traditional methods, this technology can more accurately capture and characterize the complexity of image texture, and improve the pertinence and accuracy of steganographic algorithms for embedding information in highly complex areas.

[0103] (2) Innovative design of distortion cost function QMP. This paper proposes a distortion cost function QMP based on quaternion amplitude and phase, which decomposes the complexity of image texture into two dimensions: amplitude complexity and phase complexity for quantitative evaluation. A comprehensive distortion cost function is formed by fusing amplitude and phase information through Hadamard product, which makes it easier to embed secret information into statistically difficult-to-detect areas, thereby enhancing the security of steganography.

[0104] (3) The proposal of generalized QMP (GQMP). Based on the basic framework of QMP and further combined with the exponential model, a generalized form of GQMP is proposed, which makes the expression of the distortion cost function more flexible and extensible. GQMP can adapt to different types of image carriers and steganographic requirements through adjustable parameters, thereby improving the universality of the method.

[0105] (4) Significantly improve the security of spatial domain image steganography. Experiments show that QMP and GQMP have significant advantages in improving the statistical undetectability (security) of image steganography and maintaining image quality. Compared with traditional heuristic distortion cost function methods (such as HUGO, WOW, etc.), the QMP designed in this invention achieves an innovative improvement between security and embedding efficiency.

[0106] Embodiment 1, in order to design content-adaptive image steganography with low computational complexity and high performance, the present invention follows the design idea of ​​heuristic steganography method, and focuses on how to define the embedding distortion cost function according to the image texture under the minimum distortion framework. To this end, the present invention uses the quaternion representation of the carrier image filter residual, uses the amplitude value and phase based on the quaternion to characterize the image texture, and further uses it to evaluate the complexity of the image texture, and proposes a new distortion cost function method QMP. In addition, in order to further improve the security of the algorithm, a generalized form of QMP, namely GQMP, is proposed using an exponential model.

[0107] The present invention also relates to a content-adaptive image steganography, which solves the challenges of image texture analysis and information hiding in image processing by combining the image processing process of embedding the image texture definition into the distortion cost function with practical problems. Image texture plays a vital role in image processing, but traditional methods have certain limitations in processing image texture, especially in the field of information hiding. The present invention uses the quaternion representation of the carrier image filter residual to describe and evaluate the complexity of image texture in detail with the amplitude value and phase of the quaternion. Based on this new image texture analysis method, the present invention proposes a QMP distortion cost function method, embeds the image texture information into the distortion cost function, and realizes content-adaptive image steganography technology. This innovative method has achieved remarkable results. First, the present invention realizes more effective hiding of image information and enhances the security of the image by more sophisticated analysis of the image texture and embedding it into the distortion cost function. Secondly, the present invention uses the image texture features represented by quaternions to improve the efficiency and accuracy of image processing, making the steganography process more stable and reliable. Most importantly, this method achieves a significant improvement in image steganography performance while maintaining image quality, bringing important technical breakthroughs to the field of information security. The implementation of this invention will provide strong support for the future development of image steganography technology and have a profound impact in the field of information security.

[0108] Application of technical solutions in different scenarios:

[0109] (1) Data steganography scenario: In the confidentiality scenario of information transmission, the QMP distortion cost function proposed in the present invention can embed secret information into the high-texture area of ​​the image, thereby effectively evading the statistical detection of the steganalyzer.

[0110] (2) Sensitive data protection. The present invention is applicable to application scenarios where sensitive data needs to be hidden, such as the storage and transmission of private data. In the protection of medical imaging data, QMP can ensure that patient privacy information is embedded in complex areas of the image while ensuring the high quality of the original image.

[0111] (3) Digital watermark technology: The present invention can also be applied to copyright protection, and the embedding strength of the watermark and the anti-tampering ability of the image are enhanced through the generalized distortion cost function GQMP.

[0112] like Figure 1 As shown, the spatial domain image content adaptive steganography method provided by the present invention comprises the following steps:

[0113] S1, construct a high-pass quaternion filter to perform filtering operations on the carrier image;

[0114] S2, calculates the high-pass quaternion filter residual and measures the image texture complexity based on the residual;

[0115] S3, constructing a steganographic distortion cost function QMP based on quaternion amplitude and phase representation;

[0116] S4, embeds the image texture information into the distortion cost function to achieve content-adaptive image steganography.

[0117] One of the key points of the present invention is how to construct a high-pass quaternion filter F q In recent years, in the field of image processing, many studies based on image quaternion representation of image features have shown that compared with filters in the two-dimensional real domain, two-dimensional quaternion filters can better represent image textures because they can also obtain additional image phase features that can provide a lot of image texture structure information. For a given separable two-dimensional real domain filter F, that is, if F can be expressed as:

[0118]

[0119] Where, f1, f2 are 1-dimensional real-valued filters, and T is the transpose;

[0120] High-pass quaternion filter F q It is expressed as:

[0121]

[0122] h1=H(f1)

[0123] h2=H(f2)

[0124] In the formula, F (i) is the 2D real domain filter corresponding to the imaginary unit i, i is the imaginary unit, F (j) is the 2D real domain filter corresponding to the imaginary unit j, j is the imaginary unit, F (k) is the 2D real domain filter corresponding to the imaginary unit k, where k is the imaginary unit, is the filter after the transposition of the 1-dimensional real filter f1, h1 is the 1-dimensional real domain filter after the 1-dimensional real filter f1 is subjected to the Hilbert transform, h2 is the 1-dimensional real domain filter after the 1-dimensional real domain filter f2 is subjected to the Hilbert transform, and H() is the Hilbert transform.

[0125] The second key point of the present invention is a novel measurement scheme for image texture complexity based on quaternion filter residual. The present invention uses quaternion filter F q To obtain the image residual R q , that is, using the quaternion filter F q Perform convolution operation on the carrier image X, that is:

[0126]

[0127] In the formula, R q is the image residual, R is the real image residual, R (i) is the image residual corresponding to the imaginary unit i, R (j) is the image residual corresponding to the imaginary unit j, R (k) is the image residual corresponding to the imaginary unit k, F is a two-dimensional real domain filter, is the convolution operation;

[0128] The element with index (u,v) Then we have:

[0129]

[0130] In the formula, (u,v) is the index subscript, r u,v is the element corresponding to the index subscript (u,v) in the real image residual R, is the image residual R q The element corresponding to the index subscript (u,v) in the is the image residual R (i) The element corresponding to the index subscript (u,v) in the is the image residual R (j) The element corresponding to the index subscript (u,v) in the is the image residual R (k) The element corresponding to the index subscript (u,v);

[0131] The polar coordinate form is:

[0132]

[0133] In the formula, |||| is the element The modulus, θ u,v is the angle corresponding to the index subscript (u,v), μ is the imaginary unit;

[0134] Real part r u,v is decomposed into and cos(θ u,v ), then R q The real part of is decomposed into:

[0135] R=||R q ||e cos(Θ) (6)

[0136] In the formula, ||R q || is R q Each element The amplitude M, e is the Hadamard product, and cos(Θ) is the phase R;

[0137] The present invention uses the absolute value of the image residual R based on the high-pass filter, namely |R|, to represent the image texture complexity C, namely:

[0138] C=|R|=||R q ||e cos(Θ)|=C M e C P (7)

[0139] In the formula, |R| is the absolute value of the image residual R, C M is the complexity of the carrier image determined by the amplitude, C P is the complexity of the carrier image determined by the phase;

[0140] C M and C P The rich texture content of the carrier image is characterized from the perspectives of amplitude and structure. From the above formula, the present invention can derive two inferences:

[0141] i) Given C M In this case, C P Proportional to C (i.e. ||R||), and vice versa;

[0142] ii) Given C, C M Inversely proportional to C P In order to verify the above findings, the present invention randomly selects 2000 images from the database BOSSBaseV1.0 for statistical analysis of |R|, C M and C P The relationship between Figure 2-Figure 4 As shown. Figure 2 C, C M Schematic diagram of the statistical relationship between is the standardized |R| (image texture complexity C), the ordinate C M is the amplitude complexity. In each sub-figure, the circles represent the sampling points and the curves represent the corresponding fitting curves. Figure 3 C, C P Schematic diagram of the statistical relationship between is the standardized |R| (image texture complexity C), the ordinate C P is the phase complexity. In each sub-figure, the circles represent the sampling points and the curves represent the corresponding fitting curves. Figure 4 In, C M and C P Schematic diagram of the statistical relationship between is the standardized image texture complexity C M , vertical coordinate C Pis the phase complexity. In each sub-figure, the circles represent the sampling points and the curves represent the corresponding fitting curves. It can be observed that C M and C P are proportional to |R|, and C M Inversely proportional to C P ,in, Represents the standardized C M and |R|. Therefore, the present invention believes that C M and C P It can indeed effectively present the rich texture and edge features of the image at the micro level. For standardized C M , is the standardized |R|, max(C M ) is the amplitude complexity C M The maximum value, min(C M ) is the amplitude complexity C M The minimum value in |R|, max(|R|) is the maximum value in |R|, and min(|R|) is the minimum value in |R|.

[0143] The present invention firstly proposes a new measure of image texture complexity based on quaternion amplitude and phase, namely, the complexity C based on image quaternion amplitude. M and the phase-based complexity C P ; Amplitude complexity C M and phase complexity C P It can effectively present the rich texture and edge features of the image at the micro level.

[0144] The third key point of the present invention: the construction of the steganographic distortion cost function QMP based on the quaternion amplitude and phase representation. The design of the image steganographic distortion cost function ρ should follow the complexity priority principle, that is, the area with high complexity in the carrier image should have a higher embedding priority, and then the pixels in this area should be assigned a lower distortion cost, that is, ρ and C should be inversely proportional. In this way, the present invention has:

[0145]

[0146] In the formula, ∞ is the symbol of proportional relationship, ρ M ,ρ P are the image distortion cost based on quaternion amplitude and quaternion phase, respectively. σ and ε are parameters to be determined and are set to constant values. M is the amplitude and P is the phase.

[0147] In the content-adaptive steganography scheme, considering the cost value diffusion criterion can further improve the anti-statistical detection ability of the steganography scheme, such as HiLL, MiPOD and GMRF. To this end, the present invention also applies the same cost value diffusion criterion to ρ M and ρ P Specifically, ρ M First, C M Take the reciprocal and then use the low-pass filter L for smoothing, that is:

[0148]

[0149] In the formula, C M is the amplitude complexity;

[0150] The parameter σ is used to adjust the diffusion effect of the distortion cost as the size of the low-pass filter L changes. P Smoothing is performed, and the design is slightly different. The present invention directly performs smoothing on C P Smoothing is performed instead of taking its inverse first, because C P With C M Therefore, the embedding distortion cost ρ P Defined as:

[0151]

[0152] Where ε = 10 -10 To prevent the denominator from being 0. The kernel size of the low-pass filter L is closely related to the effect of distortion cost diffusion, and the appropriate size of the low-pass filter L kernel involved should be fully determined.

[0153] In view of this, the present invention proposes a new embedding distortion cost function, namely, an efficient spatial domain image steganography embedding distortion cost function QMP based on quaternion representation. Specifically, a high-pass filter is first used to construct F q , then according to the F q The filtered image residual is calculated as C M and C P According to the design paradigm of content-adaptive steganographic distortion cost function, the distortion ρ based on the image quaternion amplitude is defined M and the distortion ρ based on the image quaternion phase P In the present invention, ρ M Heuristically defined as C M The reciprocal of each element in, because statistics show that C M is proportional to |R|. In addition, ρ M is smoothed with a low-pass filter L to diffuse the lower distortion cost to its neighboring pixels. M , ρ PIt is also heuristically defined as C P It should be noted that the low-pass filter L is used to smooth the C P Instead of P , because it is obtained experimentally that C P With C M In this way, the present invention can use ρ represented by quaternion M and ρ P The Hadamard product between them is used to define a new embedding distortion cost function QMP:

[0154]

[0155] Where cosθ is the cosine of the phase angle θ.

[0156] The present invention proposes a new embedding distortion cost function called QMP, which is an embedding distortion cost function based on quaternion amplitude and phase. QMP can improve the security of steganography schemes in resisting steganalyzer detection by enhancing content adaptability. The corresponding pseudo code algorithm table and algorithm flow chart of QMP are shown in Table 1 and Table 2. Figure 5 shown.

[0157] Table 1 Pseudocode algorithm table of QMP

[0158]

[0159] In combination with the step description of Algorithm 1, the present invention obtains the process of Algorithm QMP as follows Figure 5 The following are the specific steps for image processing using QMP:

[0160] 1. Construct quaternion filter: Design a specific quaternion filter F q , used to perform filtering operations on the carrier image in subsequent processing.

[0161] 2. Calculate the quaternion filter residual: Apply the constructed quaternion filter to the carrier image and calculate the residual R after filtering q , which is the difference before and after filtering.

[0162] 3. Real part decomposition of image quaternion residual: The quaternion residual R obtained after filtering q Process and extract its real number part And decompose it into two parts, among which, let C M =||R q ||,C p =|cosθ| represents the complexity based on the image quaternion amplitude and phase respectively.

[0163] 4. Calculate the embedding distortion cost ρ based on amplitude M and phase PM and P :Under the design paradigm of content-adaptive steganography scheme, the design of image steganography distortion cost function ρ is designed by using complexity priority criterion and cost value diffusion criterion M and P .

[0164] 5. Calculate the embedding distortion cost corresponding to the carrier image X: Heuristically define the embedding distortion cost function of the spatial domain image as ρ = ρ M eρ P .

[0165] The image texture information is embedded into the distortion cost function QMP to achieve content-adaptive image steganography, that is, the distortion metrics of each element of the carrier image are associated with the image content, and the secret information is adaptively embedded into the complex texture area of ​​the image that is not easily detected by the steganalysis detector. The experimental results show that the QMP proposed in this paper can effectively capture texture features and enable the secret information to be embedded into the texture area of ​​the carrier image (see Figure 5-Figure 8 ), thereby improving statistical undetectability (see Tables 3-5).

[0166] The fourth key point of the present invention is the design of a generalized QMP (GQMP) distortion cost function based on an exponential model. In order to flexibly balance the security of QMP in resisting steganalyzer SRM and maxSRMd2 detection, the present invention further introduces an exponential model and obtains a generalized form of QMP, which the present invention calls GQMP, namely:

[0167]

[0168] By formula Decide, By formula The power exponents α and β are parameters to be determined and their values ​​are non-negative constants.

[0169] In order to flexibly balance the security of the steganographic algorithm in resisting different steganalyzers, the present invention proposes an improved QMP based on an exponential model, namely, GQMP.

[0170] The corresponding pseudo code algorithm table and algorithm flow chart of GQMP are shown in Table 2 and Figure 6 shown.

[0171] Table 2 Pseudocode algorithm table of GQMP

[0172]

[0173]

[0174] In combination with the step description of Algorithm 2, the present invention obtains the process of the algorithm GQMP, as shown in Figure 6 The following are the specific steps for image processing using GQMP:

[0175] 1. Input: Input embedding distortion cost ρ M , embedding distortion cost ρ P and steganalyzerA.

[0176] 2. Output: Get the distortion cost ρ corresponding to the carrier image X.

[0177] 3. The outer loop starts: the power index α0 is iterated in a loop.

[0178] 4. Inner loop starts: loop iterates the power index β0. If the condition is met, calculate the distortion cost And use the steganalyzer A to calculate the average detection error rate of the steganalysis scheme determined by ρ Then make a conditional judgment, if If established, execute α=α0, β=β0, otherwise the conditional judgment ends;

[0179] 5. Calculate the power exponents α0 and β0: add 1 to the power exponents α0 and β0 respectively.

[0180] 6. End condition: end the inner loop.

[0181] 7. End condition: end the outer loop.

[0182] 8. Calculate the embedding distortion cost: Calculate the embedding distortion cost corresponding to the carrier image X

[0183] 9. Return result: Returns the distortion cost ρ of the carrier image.

[0184] The above-mentioned distortion cost function GQMP processes the algorithm flow of the image, and the algorithm can more effectively realize content-adaptive image steganography.

[0185] Embodiment 2, the embodiment of the present invention further provides a spatial domain image content adaptive steganography system, the system is used to control the spatial domain image content adaptive steganography method, the system comprises:

[0186] Image filtering module, used to construct a high-pass quaternion filter to perform filtering operations on the carrier image;

[0187] The residual image calculation module is used to apply the constructed quaternion filter to the carrier image, calculate the residual of the high-pass quaternion filter, and measure the image texture complexity based on the residual;

[0188] The information embedding module is used to construct a steganographic distortion cost function based on quaternion amplitude and phase representation, embed the image texture information into the distortion cost function, and realize content-adaptive image steganography.

[0189] Embodiment 3, the embodiment of the present invention further provides a spatial domain image adaptive steganography device, the device includes a processor, a memory, and a spatial domain image adaptive steganography control program stored in the memory and running on the processor, the spatial domain image adaptive steganography control program when executed by the processor implements the following steps:

[0190] Construct a high-pass quaternion filter to perform filtering operations on the carrier image;

[0191] Calculate the high-pass quaternion filter residual and measure the image texture complexity based on the residual;

[0192] Construct a steganographic distortion cost function QMP based on quaternion amplitude and phase representation;

[0193] Image texture information is embedded into the distortion cost function to achieve content-adaptive image steganography.

[0194] Embodiment 4, the embodiment of the present invention also provides an information data processing terminal, which is used to implement the spatial domain image content adaptive steganography system. The system captures the image texture complexity more accurately. Through the residual of the high-pass quaternion filter, the texture complexity of the image is accurately quantified, so as to identify the distribution of complex and simple areas in the image. Improve the security and concealment of spatial domain image steganography. Design a distortion cost function based on the amplitude and phase texture complexity, embed information in the complex texture area, make the spatial domain image steganography more difficult to detect, and improve the concealment of the spatial domain image steganography. Reduce the visual impact of the image after embedding secret information. Embedding information in high-complexity areas can effectively reduce the damage to the overall quality of the image and ensure the visual consistency between the carrier image and the secret image.

[0195] This technology for spatial domain image content adaptive steganography system solves the following problems in image steganography: Differences in texture characteristics of different complex areas. Different areas of the image have different texture complexities, and direct steganography may cause obvious artifacts in areas with low texture complexity. Calculating texture complexity through a high-pass quaternion filter can locate areas suitable for steganography. The risk of easy detection of stego information. Embedding stego information in areas with low texture complexity is easy to be found as anomaly by statistical detection tools. This patent improves anti-detection capabilities by embedding information in areas with high texture complexity as much as possible. Impaired image visual quality. This system minimizes the impact of information embedding on image quality through texture adaptive operations.

[0196] To further demonstrate the positive effects of the above embodiments, the present invention conducts the following experiments based on the above technical solution. See Table 3 to Table 6 for details.

[0197] For content-adaptive image steganography technology, existing solutions face the following challenges: heuristic methods have low computational complexity but insufficient security; model-driven methods are theoretically rigorous but have high computational complexity; deep learning solutions provide better security, but the training process is complex and cumbersome. This paper follows the design idea of ​​heuristic solutions and designs a distortion cost function QMP with high security based on quaternion texture representation.

[0198] Previous design paradigms for content-adaptive steganographic distortion cost functions in the minimum distortion embedding framework show that an efficient embedding distortion cost function relies heavily on the representation of image texture complexity. This paper takes a step forward by using quaternion representation, where the image texture complexity is decomposed into two parts: the complexity based on quaternion amplitude and the complexity based on quaternion phase. On this basis, we empirically define the embedding distortion cost ρ based on quaternion amplitude M and the embedding distortion cost ρ based on the quaternion phase P , and use ρ M and ρ P The Hadamard product between them obtains a new distortion cost function QMP.

[0199] The results show that the proposed QMP can effectively capture texture features and embed secret information into the texture area of ​​the carrier image, thereby improving statistical undetectability. Figure 7-10 The results in are verified: in areas with rich textures, the embedding probability of the proposed QMP is generally high, which means that in the QMP scheme proposed in this chapter, more pixels in areas with complex image textures are used for embedding modification. In addition, a generalized form of QMP based on the exponential model, namely GQMP, is developed to further improve the security performance. Extensive experiments show that the QMP and GQMP proposed in this invention have superior security performance in resisting the steganalysis of steganalyzers SRM, maxSRMd2 and SRNet than the most advanced methods such as S-UNIWARD, HiLL and MiPOD, see Tables 3-5.

[0200] Table 3 Average detection error probability of the security performance of the QMP and GQMP proposed in the present invention and S-UNIWARD, HiLL and MiPOD in resisting SRM detection Compare

[0201]

[0202] Table 4 Security performance of QMP and GQMP proposed in the present invention and S-UNIWARD, HiLL and MiPOD in resisting maxSRMd2 detection Average detection error probability Compare

[0203]

[0204]

[0205] Table 5 Average detection error probability of the security performance of QMP and GQMP proposed in this invention and S-UNIWARD, HiLL and MiPOD in resisting SRNet detection Compare

[0206]

[0207] In addition, compared with deep learning solutions, for secure payloads (i.e. The proposed method has almost no difference in security with SPAR-RL-v2 and is significantly higher than the UT-GAN algorithm, as shown in Table 6. In summary, compared with the SOTA heuristic method and model-driven method, QMP has higher security and its computational complexity is much lower than that of the model-driven method; compared with the deep learning method, QMP achieves comparable security and even has higher security at lower embedding loads. In addition, the generalized form of QMP GQMP based on the exponential model can further improve the security performance.

[0208] Table 6 Average detection error probability of the security performance of QMP proposed in the present invention, UT-GAN, and SPAR-RL-V2 in resisting steganalyzer SRM detection Compare

[0209]

[0210] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A spatial domain image content adaptive steganography method, characterized in that: The method comprises the steps of: S1, construct a high-pass quaternion filter to perform filtering operations on the carrier image; S2, calculates the high-pass quaternion filter residual and measures the image texture complexity based on the residual; S3, constructing a steganographic distortion cost function QMP based on quaternion amplitude and phase representation; S4, embeds the image texture information into the distortion cost function to achieve content-adaptive image steganography.

2. According to claim 1, the spatial domain image content adaptive steganography method is characterized by: In step S1, a high-pass quaternion filter is constructed, including: The two-dimensional real domain filter F is expressed as: Where, f1, f2 are 1-dimensional real-valued filters, and T is the transpose; High-pass quaternion filter F q It is expressed as: h1=H(f1) h2=H(f2) In the formula, F (i) is the 2D real domain filter corresponding to the imaginary unit i, F (j) is the 2D real domain filter corresponding to the imaginary unit j, F (k) is the 2D real domain filter corresponding to the imaginary unit k, i, j, k are all imaginary units, is the filter after the transposition of the 1-dimensional real filter f1, h1 is the 1-dimensional real domain filter after the 1-dimensional real filter f1 is subjected to the Hilbert transform, h2 is the 1-dimensional real domain filter after the 1-dimensional real domain filter f2 is subjected to the Hilbert transform, and H() is the Hilbert transform.

3. The spatial domain image content adaptive steganography method according to claim 2, characterized in that: In step S2, the high-pass quaternion filter residual is calculated, and the image texture complexity is measured based on the residual, including: Using a high-pass quaternion filter F q Perform convolution operation on the carrier image X, the expression is: In the formula, R q is the image residual, R is the real image residual, R (i) is the image residual corresponding to the imaginary unit i, R (j) is the image residual corresponding to the imaginary unit j, R (k) is the image residual corresponding to the imaginary unit k, F is a two-dimensional real domain filter, is the convolution operation, X is the carrier image; The element with index (u, v) Then we have: In the formula, (u, v) is the index subscript, r u，v is the element with index (u, v) in the real image residual R, is the image residual R q The element corresponding to the index subscript (u, v) in the is the image residual R (i) The element corresponding to the index subscript (u, v) in the is the image residual R (j) The element corresponding to the index subscript (u, v) in the is the image residual R (k) The element corresponding to the index subscript (u, v) in the The polar coordinate form is: In the formula, |||| is the element The modulus, θ u，v is the angle corresponding to the index subscript (u, v), μ is the imaginary unit; Real part r u，v is decomposed into and cos(θ u，v ), then R q The real part of is decomposed into: R=||R q ||e cos(Θ) (6) In the formula, ||R q || is R q Each element The amplitude M, e is the Hadamard product, cos(Θ) is the phase P, and R is the real image residual; The absolute value of the image residual R based on the high-pass filter ||R|| is used to represent the image texture complexity C: C=|R|=|||R q ||and cos(Θ)|=C M and C P (7) In the formula, |R| is the absolute value of the image residual R, C M is the complexity of the carrier image determined by the amplitude, C P is the complexity of the carrier image determined by the phase; C M and C P are proportional to |R|, and C M Inversely proportional to C P ; In the formula, For standardized C M , is the standardized |R|, max(C M ) is the amplitude complexity C M The maximum value, min(C M ) is the amplitude complexity C M The minimum value in |R|, max(|R|) is the maximum value in |R|, and min(|R|) is the minimum value in |R|.

4. The spatial domain image content adaptive steganography method according to claim 1, characterized in that: In step S3, a steganographic distortion cost function QMP based on quaternion amplitude and phase representation is constructed, including: The image steganography distortion cost function ρ is inversely proportional to the complexity C, and the expression is: In the formula, ∞ is the symbol of proportional relationship, ρ M , ρ P are the image distortion cost based on quaternion amplitude and quaternion phase, respectively. σ and ε are parameters to be determined and are set to constant values. M is the amplitude and P is the phase. ρ M C M After taking the reciprocal, use the low-pass filter L for smoothing, the expression is: In the formula, C M is the amplitude complexity; The parameter σ is used to adjust the diffusion effect of the distortion cost as the size of the low-pass filter L changes. P Perform smoothing and embed the distortion cost ρ P Defined as: Where ε = 10 -10 ; The embedding distortion cost function QMP is expressed as: Where cosθ is the cosine of the phase angle θ.

5. The spatial domain image content adaptive steganography method according to claim 1, characterized in that: In step S4, the image texture information is embedded into the distortion cost function to achieve content-adaptive image steganography, including: Construct a high-pass quaternion filter F q , used to perform filtering operations on the carrier image; Apply the constructed high-pass quaternion filter to the carrier image and calculate the residual R after filtering q ; The filtered quaternion residual R q Process and extract the real part And decompose it into two parts, let C M =||R q ||,C p =|cosθ| represents the complexity based on the image quaternion amplitude and phase respectively; Calculate the embedding distortion cost ρ based on amplitude M and phase P M and P :Using the complexity priority criterion and the cost value diffusion criterion, we design the image steganography distortion cost function ρ M and P ; Calculate the embedding distortion cost corresponding to the carrier image X: Heuristically define the embedding distortion cost function of the spatial domain image as ρ = ρ M e ρ P .

6. The spatial domain image content adaptive steganography method according to claim 1, characterized in that: The method also includes constructing a generalized QMP distortion cost function GQMP based on an exponential model, which is expressed as: By formula Decide, By formula The power exponents α and β are parameters to be determined and their values ​​are non-negative constants.

7. The spatial domain image content adaptive steganography method according to claim 6, characterized in that: The method also includes performing image processing using a generalized QMP distortion cost function GQMP, specifically including: Input embedding distortion cost ρ M , embedding distortion cost ρ P and steganalyzer A, and obtain the distortion cost ρ corresponding to the carrier image X; Start the outer loop and perform loop iteration operation on the power index α0; Start the inner loop and iterate the power index β0; if the condition is met, calculate the distortion cost And use the steganalyzer A to calculate the average detection error rate of the steganalysis scheme determined by ρ Then make a conditional judgment, if If established, execute α=α0, β=β0, otherwise end this condition determination; Add 1 to the power exponents α0 and β0 respectively, calculate the power exponents α0 and β0, and end the inner and outer loops; Calculate the embedding distortion cost corresponding to the carrier image X Returns the distortion cost ρ of the carrier image.

8. A spatial domain image content adaptive steganography system, characterized in that: The system is used to control the spatial domain image content adaptive steganography method according to any one of claims 1 to 7, and the system comprises: Image filtering module, used to construct a high-pass quaternion filter to perform filtering operations on the carrier image; The residual image calculation module is used to apply the constructed quaternion filter to the carrier image, calculate the residual of the high-pass quaternion filter, and measure the image texture complexity based on the residual; The information embedding module is used to construct a steganographic distortion cost function QMP based on quaternion amplitude and phase representation, embed the image texture information into the distortion cost function, and realize content-adaptive image steganography.

9. A spatial domain image adaptive steganography device, characterized in that: The device comprises a processor, a memory and a spatial domain image adaptive steganography control program stored in the memory and running on the processor. When the spatial domain image adaptive steganography control program is executed by the processor, a spatial domain image content adaptive steganography method is implemented.

10. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the spatial domain image content adaptive steganography method as described in claim 8.

Images