Encryption steganography method for video
The video steganography method that combines RSA encryption and diffusion model solves the problems of insufficient concealment, capacity and robustness in the existing technology, achieves high concealment and high information load that is visually imperceptible, and is adaptable to various video formats.
Patent Information
- Application Number
- CN202510866633.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-26
- Publication Date
- 2025-09-05
AI Technical Summary
Existing video steganography technologies find it difficult to find a balance between improving concealment, capacity, and robustness. LSB-based methods are easily detected, transform domain operations are limited by the capacity-security paradox, and compressed domain schemes lose their versatility.
The RSA encryption algorithm is used to eliminate the statistical characteristics of the target information. Combined with the spatial adaptation block and time domain alignment modules, information is embedded through the least significant bit method, and the diffusion model is used for reverse diffusion recovery. The encrypted information maintains consistency in pixel distribution and inter-frame motion.
It achieves high concealment and high information load that is visually imperceptible, is resistant to noise attacks, adapts to various video formats, avoids traditional detection methods, and improves robustness and information recovery rate.
Smart Images
Figure CN120602670A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing, and in particular to an encryption steganography method for videos. Background Art
[0002] Video steganography can be implemented in a variety of ways. Based on the format of the carrier video, methods can be categorized as either native domain or compressed domain. Native domain methods can be further categorized as spatial domain or transform domain. Least significant bit (LSB) substitution is the primary spatial domain method; discrete wavelet transform (DWT) and discrete cosine transform (DCT) are commonly used transform domain methods; and DCT / DST coefficients, motion vectors, intra-frame prediction, and entropy coding are the primary compressed domain methods.
[0003] Among these methods, spatial-domain LSB substitution is the most common. While most LSB-based methods demonstrate good stealth and high hiding capacity, they are also not robust enough to withstand various attacks and noise. Furthermore, current steganalysis methods easily detect modifications based on LSB substitution because, in LSB-based methods, the modifications are applied directly to the original pixel values of the frame.
[0004] DCT and DWT are two commonly used transform functions in steganography. They can be used to embed information in the transform domain. Many methods improve robustness and security by applying multiple levels of DWT to the carrier medium. However, applying multiple levels of DWT to the carrier medium reduces the information embedding capacity. Summary of the Invention
[0005] In order to improve the noise resistance and concealment performance of video steganography while increasing its versatility, the present invention proposes an encryption steganography method for video, comprising the following steps: S1: Using the RSA encryption algorithm to obtain encrypted information by eliminating the statistical characteristics of the target information, and embedding the encrypted information into the video frame of the original video through the least significant bit method to obtain an encrypted video frame; S2: After adding Gaussian noise to the encrypted video frame based on the diffusion model principle, the original video features are injected by introducing a spatial adaptation block into the denoising network to obtain a preliminary steganographic video with the original video features; S3: Based on the self-attention mechanism and cross-attention mechanism of the temporal alignment module, the preliminary stego video is subjected to inter-frame information interaction and the information correction between the corresponding positions of the original video and the preliminary stego video is performed to obtain the corrected stego video. S4: Recovery of encrypted video frames under reverse diffusion of the corrected steganographic video based on the reversibility of the diffusion model; S5: extracting the least significant bit data from the restored encrypted video frame, restoring the extracted least significant bit data to encrypted information through the least significant bit method, and then decrypting it to the target information through RSA.
[0006] Furthermore, in the step S1, the RSA encryption algorithm is implemented by the following steps: randomly generating a large prime number and and calculate the modulus and the Euler function value , select Crypto Index , calculate the decryption index , based on the decryption index, the target information is encrypted. The encryption formula is as follows: Where c is the encrypted information and m is the target information.
[0007] Furthermore, in the step S1, the embedding operation of the least significant bit method includes: The encrypted information is binarized and divided into several information bit segments. Any color component in the YUV channel of the video frame is randomly selected and the lowest 1-2 bits of the pixel value are replaced with the information bit segment.
[0008] Furthermore, in the step S2, the spatial adaptation block implements original video feature injection through feature modulation, specifically including: Perform convolution to extract mean features and variance characteristics , to add Gaussian noise to the encrypted video frame Calculate its mean and standard deviation , the preliminary steganographic video frame is generated through the feature fusion formula, and the feature fusion formula is expressed as follows: Where, is pixel-by-pixel multiplication, is the video frame index, each video index corresponds to Together they form a preliminary steganographic video .
[0009] Furthermore, in the step S3, the execution process of the time domain alignment module specifically includes: Initial steganographic video The original video G is divided into a time sequence unit consisting of N consecutive frames, which are respectively recorded as and ; Inter-frame information interaction is carried out through the self-attention mechanism, specifically for the initial steganographic video timing unit Perform 3D convolution to generate query matrix Q, key matrix K, value matrix V, and calculate the intermediate steganographic video unit : The spatial-temporal domain correction is performed through the cross-attention mechanism. Perform 3D convolution to generate the key matrix Sum Matrix , and based on the intermediate steganographic video unit Generate query matrix , and calculate the corrected steganographic video : Where, is the dimension of the key vector, and T is the matrix transpose.
[0010] Furthermore, in step S4, the reversibility of the diffusion model is achieved through DDIM reverse diffusion, which specifically includes: Based on noise predictor Construct the denoising function: Where, , is the preset noise scheduling coefficient, t is the diffusion step index and ; Solve for the initial state via an ordinary differential equation: Where, To encrypt video frames, is the diffusion state at the Tth step, and T is the total number of diffusion steps.
[0011] Furthermore, in step S5, the RSA decryption algorithm formula is expressed as: In the formula, c is the encrypted information, m is the target information, To decrypt the index, To generate random large prime numbers and The modulus of .
[0012] Furthermore, in the step S2, before adding Gaussian noise, an information capacity optimization step is also included: dynamically adjusting the amount of embedded encrypted information under a positive correlation relationship based on the texture complexity of the original video.
[0013] Furthermore, in the step S2, the Gaussian noise added obey , is a multidimensional Gaussian distribution, is the noise variance scalar value, is the identity matrix.
[0014] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The present invention proposes an encryption steganography method for video, which uses the RSA encryption algorithm to eliminate the statistical characteristics of the target information and combines it with the feature modulation of the spatial adaptive block to make the pixel distribution of the steganographic video frame maintain statistical consistency with the original video, completely avoiding traditional analysis methods such as chi-square detection. At the same time, the dual attention mechanism of the temporal alignment module ensures the natural continuity of inter-frame motion, eliminates the blocking effect and edge artifacts common in transform domain methods, and achieves visual imperceptibility. (2) A dynamic embedding adjustment strategy based on texture complexity automatically reduces the embedding amount in simple texture areas (reducing visual anomalies) and increases the embedding amount in complex texture areas, achieving global high load; (3) The added Gaussian noise and DDIM reversible recovery form a synergistic anti-interference mechanism, which can still ensure the information recovery rate under noise attack and can be independent of the dependence on specific coding standards. It can adapt to any video format through pure pixel-level operations. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 A method step diagram for an encryption steganography method for video; Figure 2 Schematic diagram of the spatial adaptation block structure; Figure 3 Schematic diagram of the time domain alignment block structure. DETAILED DESCRIPTION
[0016] Faced with the core bottleneck of current video steganography technology that is difficult to balance concealment, capacity and robustness, this paper proposes an innovative solution based on the encryption-diffusion collaborative architecture. In traditional methods, LSB embedding based on the spatial domain is easily cracked by statistical detection, frequency domain operations in the transform domain are limited by the capacity-security paradox, and compressed domain solutions lose their versatility due to their coding dependence. To overcome the above limitations, Figure 1 As shown, the present invention proposes an encryption steganography method for video, comprising the steps of:
[0017] S1: Using the RSA encryption algorithm to obtain encrypted information by eliminating the statistical characteristics of the target information, and embedding the encrypted information into the video frame of the original video through the least significant bit method to obtain an encrypted video frame; S2: After adding Gaussian noise to the encrypted video frame based on the diffusion model principle, the original video features are injected by introducing a spatial adaptation block into the denoising network to obtain a preliminary steganographic video with the original video features; S3: Based on the self-attention mechanism and cross-attention mechanism of the temporal alignment module, the preliminary stego video is subjected to inter-frame information interaction and the information correction between the corresponding positions of the original video and the preliminary stego video is performed to obtain the corrected stego video. S4: Recovery of encrypted video frames under reverse diffusion of the corrected steganographic video based on the reversibility of the diffusion model; S5: extracting the least significant bit data from the restored encrypted video frame, restoring the extracted least significant bit data to encrypted information through the least significant bit method, and then decrypting it to the target information through RSA.
[0018] As can be seen, this invention deeply integrates cryptographic preprocessing with a generative model enhancement mechanism: First, RSA encryption eliminates statistical information features, while a texture-adaptive dynamic embedding strategy maximizes information payload. Next, the spatial-temporal modulation capability of the diffusion model is leveraged to construct an interference-resistant stegoscopy carrier while injecting original video features. Finally, the strict reversibility of the denoising diffusion implicit model (DDIM) enables lossless information recovery. The following, combined with specific examples, details the implementation path and technical advantages of this solution in addressing issues such as fragile concealment, limited capacity, and insufficient robustness.
[0019] For target information that needs information hiding, the present invention first randomly generates two large prime numbers and , calculate the modulus And the corresponding Euler function value , select the encryption index , so that it satisfies And with Coprime (i.e. ), and then calculate the decryption index The target information m (such as text or binary stream) to be hidden is processed into blocks according to the formula:
[0020] , The ciphertext c is generated by encryption block by block. This process completely destroys the statistical characteristics of the target information m (such as character frequency distribution), making the ciphertext c statistically uniformly random, fundamentally circumventing traditional steganalysis such as chi-square detection.
[0021] Next, based on the texture complexity D of the original video frame, the per-frame embedding amount L (in bits) is dynamically determined according to the proportional relationship L = k × D, where the scaling factor k is set based on the application scenario. Based on the dynamically adjusted per-frame embedding amount, the ciphertext c is binarized and split into bit segments of length L. A component in the YUV color space of the video frame is randomly selected as the embedding channel. Within the pixel positions of the selected component, L pixels are pseudo-randomly selected, and their lowest 1-2 bits (determined by the embedding amount L) are replaced with the ciphertext bits, thus generating an encrypted video frame.
[0022] Compared with the fixed replacement of the lowest bit with the LSB, the present invention uses texture-driven dynamic embedding and YUV channel randomization in this link to increase information capacity while maintaining visual concealment. The random switching of channels makes it impossible for attackers to locate the tampered area, thereby ensuring the stability of the encrypted information.
[0023] In order to further enhance the concealment of the steganographic video and preserve the visual features of the original video, the present invention also performs noise processing and feature injection on the generated encrypted video frame. Specifically, the forward diffusion process is first performed on the encrypted video frame, and Gaussian noise is added to the RGB channel of each pixel to simulate the channel interference in the real transmission environment (such as wireless transmission packet loss) by destroying its statistical characteristics, making it closer to the noise distribution of natural video. Must obey multidimensional normal distribution ,in is a multidimensional Gaussian distribution, is the noise variance scalar value, is the identity matrix. By adjusting The value of can dynamically balance the robustness and concealment of the steganographic information, for example, increasing the noise intensity in areas with complex textures to improve the information embedding capacity.
[0024] Then, if Figure 2 As shown in Figure 2, a spatial adaptation block is introduced into the denoising network to perform the injection of original video features. Perform two-way 2D convolution (Conv2D, convolution kernel 3×3, step size 1) and output mean features respectively and variance characteristics , in order to capture the texture structure and illumination distribution. At the same time, the noisy encrypted video frame is calculated through normalization (Norm) processing The frame-by-frame mean and standard deviation , and apply the feature fusion formula to generate preliminary steganographic video frames :
[0025] , in, is pixel-by-pixel multiplication, is the video frame index. Through this formula, the local features of the original video ( ) Inject the steganographic frame, through the preliminary steganographic video frame under each video frame index Together they form the required preliminary steganographic video This operation makes the output of the preliminary steganographic video frame Video frames that are similar to the original video in terms of edge details and color distribution Maintain high consistency and completely avoid visual inspection based on residual analysis.
[0026] Compared with traditional steganography that only modifies pixel values, the encryption optimization method combined with forward diffusion proposed in this invention realizes generative feature reconstruction through spatial adaptation blocks. During the forward noisification process of the diffusion model, the spatiotemporal features of the original video are encoded into the steganographic carrier, making the generated preliminary steganographic video visually indistinguishable.
[0027] Although the above operation removes the statistical characteristics of the original video and improves the concealment of the steganographic information, as a temporal carrier, the dynamic correlation between adjacent frames of the video still has a key impact on visual perception and steganographic robustness. Figure 3 As shown in the figure, the present invention also introduces a self-attention mechanism and a cross-attention mechanism based on time-domain alignment blocks. Through inter-frame information interaction and spatiotemporal correction between the original video and the steganographic video, the temporal discontinuity that may be caused by the steganographic operation is further eliminated, thereby achieving a steganographic effect with higher concealment and robustness in dynamic scenarios.
[0028] First, the initial steganographic video The original video G is divided into a time sequence unit consisting of N consecutive frames ( and ), where N is the preset time window. This segmentation strategy only captures the local temporal dynamic features of the video while reducing the computational complexity. In the self-attention mechanism, the initial steganographic video temporal unit Perform 3D convolution operations to generate query matrix Q, key matrix K and value matrix V respectively. Then, calculate the intermediate steganographic video unit through the self-attention formula :
[0029] .
[0030] The process dynamically aggregates the inter-frame information in the value matrix V by matching the similarity between the query matrix Q and the key matrix K. While preserving preliminary stegographic information, it also enhances the spatiotemporal consistency between adjacent frames. For example, in fast-moving scenes, the self-attention mechanism can suppress the inter-frame flickering effect caused by the embedding of encrypted information, thereby preventing the stegographic traces from being detected by visual perception or statistical analysis methods.
[0031] Furthermore, in order to correct the deviation of the temporal and spatial characteristics between the steganographic video and the original video, the original video timing unit Also perform 3D convolution operation to generate the key matrix Sum Matrix , while based on the intermediate steganographic video unit after self-attention processing Generate query matrix Obtain the corrected steganographic video unit by calculating the cross attention formula :
[0032] .
[0033] The core of this step is to use the timing unit of the original video The global temporal features of the intermediate steganographic video unit Perform spatial-temporal correction to eliminate local feature distortion caused by embedding encrypted information and adding noise. For example, in static backgrounds or slow-motion scenes, the cross-attention mechanism can restore the texture details of the stego video through the reference frame of the original video, ensuring that the corrected stego video The visual quality is highly consistent with the original video. In addition, the mechanism dynamically adjusts the weight distribution to prioritize the features of high-texture complexity areas in the original video, thereby improving concealment while also ensuring the robustness of information embedding.
[0034] In this regard, the present invention introduces a temporal alignment module, using a cascaded design of self-attention and cross-attention to achieve dual optimization of the stego video in the temporal dimension: on the one hand, the self-attention mechanism enhances the dynamic coherence of the video through inter-frame information interaction; on the other hand, the cross-attention mechanism corrects the spatiotemporal features of the original video, further bridging the local differences introduced by the stego operation. This design not only significantly reduces the risk of feature loss caused by temporal inconsistencies during video encoding, transmission, or compression, but also enhances the resistance of the stego information to conventional steganalysis algorithms by preserving the original video's motion vectors and inter-frame dependencies.
[0035] After completing inter-frame information exchange and positional information correction to obtain the corrected steganographic video, the present invention enters the critical stage of information recovery. In this stage, the inherent reversibility of the diffusion model is exploited to recover the video frames originally embedded with the encrypted information from the corrected steganographic video processed through the aforementioned steps. This process is essentially the reverse of the diffusion process (noising).
[0036] Specifically, this embodiment uses the Denoising Diffusion Implicit Models (DDIM) to implement the reverse diffusion process to efficiently and accurately restore the original encrypted video frame. This process relies on the noise predictor model trained in the previous step. , taking the corrected stego video as the starting state of the reverse diffusion process, which corresponds to the T-step diffusion state in the diffusion model theory .
[0037] The core of DDIM back diffusion is to define a key denoising function , which is a noise predictor trained to convergence Constructed of: , In this formula, t represents the current back-diffusion step index (step by step from the starting step T to the target step 0), is a preset cumulative noise scheduling coefficient result ( ,in is a value in the predefined sequence of forward noise scheduling coefficients).
[0038] The entire recovery process is completed by solving an ordinary differential equation (ODE). The ODE is based on the initial state (i.e., correcting the representation state of the stego video), trained noise predictor , total diffusion steps T and target step 0 as input parameters:
[0039] , The result of solving the above equation is This means the status has been successfully restored. The state corresponds to the original encrypted video frame in step S1 of the present invention.
[0040] Solved This is the final restored encrypted video frame. This encrypted video frame clearly and completely retains the encryption information originally embedded using the least significant bit (LSB) method, and its pixel structure is ready for subsequent extraction of the least significant bit information.
[0041] Based on the recovered encrypted video frames, the present invention employs a least significant bit (LSB) extraction algorithm to separate the binary data of the encrypted information from the recovered video frames. Specifically, the system iterates through each video frame, locates the YUV channel color component (e.g., the Y component) selected during the embedding phase, and extracts the least significant bit values of the 1-2 bits pixel by pixel. Because the structure of the original video frame has been restored through inverse diffusion, the embedded least significant bit data can be fully read with high fidelity. The extracted binary data segments are reassembled in the order in which they were segmented during embedding to form the complete encrypted information bitstream.
[0042] Subsequently, the extracted encrypted information is decrypted using the RSA decryption algorithm to restore the original target information. According to the encryption parameter settings of the data RSA encryption step, the decryption process uses the pre-calculated decryption exponent d and the modulus 𝑛 to perform the following modular exponentiation operation:
[0043] , Where 𝑐 is the integer value converted from the encrypted information binary stream extracted from the LSB, and m is the target information after decryption. To ensure the correctness of decryption, the system verifies the decryption exponent before decryption. and the Euler function value The mutual primacy of the two parameters prevents decryption failures due to parameter errors. The decrypted data is converted to an encoding (such as UTF-8) and finally output as readable plaintext information.
[0044] In summary, the present invention proposes an encrypted steganography method for videos. It uses the RSA encryption algorithm to eliminate the statistical characteristics of the target information. Combined with the feature modulation of the spatial adaptive block, it ensures that the pixel distribution of the steganographic video frame is statistically consistent with that of the original video, completely circumventing traditional analysis methods such as the chi-square test. At the same time, the dual attention mechanism of the temporal alignment module ensures the natural continuity of inter-frame motion, eliminates the blocking effect and edge artifacts common in transform domain methods, and achieves visual imperceptibility.
[0045] A dynamic embedding adjustment strategy based on texture complexity automatically reduces embedding in simple texture areas (reducing visual anomalies) and increases embedding in complex texture areas, achieving high global load. The addition of Gaussian noise and DDIM reversible recovery form a synergistic anti-interference mechanism, ensuring information recovery under noise attacks. It also breaks away from reliance on specific coding standards and adapts to any video format through pure pixel-level operations.
[0046] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.
[0047] In addition, in the present invention, descriptions such as "first," "second," and "one" are for descriptive purposes only and should not be understood to indicate or imply their relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.
[0048] In the present invention, unless otherwise specified or limited, the terms "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can mean fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two elements or interaction between two elements, unless otherwise specified. Those skilled in the art will be able to understand the specific meanings of the above terms in the present invention based on specific circumstances.
[0049] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.
Claims
1. A video encryption steganography method, characterized in that: Including steps: S1: Using the RSA encryption algorithm to obtain encrypted information by eliminating the statistical characteristics of the target information, and embedding the encrypted information into the video frame of the original video through the least significant bit method to obtain an encrypted video frame; S2: After adding Gaussian noise to the encrypted video frame based on the diffusion model principle, the original video features are injected by introducing a spatial adaptation block into the denoising network to obtain a preliminary steganographic video with the original video features; S3: Based on the self-attention mechanism and cross-attention mechanism of the temporal alignment module, the preliminary stego video is subjected to inter-frame information interaction and the information correction between the corresponding positions of the original video and the preliminary stego video is performed to obtain the corrected stego video. S4: Recovery of encrypted video frames under reverse diffusion of the corrected steganographic video based on the reversibility of the diffusion model; S5: extracting the least significant bit data from the restored encrypted video frame, restoring the extracted least significant bit data to encrypted information through the least significant bit method, and then decrypting it to the target information through RSA.
2. The video encryption steganography method according to claim 1, wherein: In the step S1, the RSA encryption algorithm is implemented by the following steps: randomly generate a large prime number and And calculate the modulus and the Euler function value , select Crypto Index , calculate the decryption index , based on the decryption index, the target information is encrypted. The encryption formula is as follows: Where, c is the encrypted information and m is the target information.
3. The video encryption steganography method according to claim 1, wherein: In the step S1, the embedding operation of the least significant bit method includes: The encrypted information is binarized and divided into several information bit segments. Any color component in the YUV channel of the video frame is randomly selected and the lowest 1-2 bits of the pixel value are replaced with the information bit segment.
4. The video encryption steganography method according to claim 1, wherein: In the step S2, the spatial adaptation block implements the original video feature injection through feature modulation, specifically including: Perform convolution to extract mean features and variance characteristics , to add Gaussian noise to the encrypted video frame Calculate its mean and standard deviation , the preliminary steganographic video frame is generated through the feature fusion formula, and the feature fusion formula is expressed as follows: Where, is pixel-by-pixel multiplication, is the video frame index, each video index corresponds to Together they form a preliminary steganographic video .
5. The video encryption steganography method according to claim 1, wherein: In the step S3, the execution process of the time domain alignment module specifically includes: Initial steganographic video The original video G is divided into a time sequence unit consisting of N consecutive frames, which are respectively recorded as and ; Inter-frame information interaction is carried out through the self-attention mechanism, specifically for the initial steganographic video timing unit Perform 3D convolution to generate query matrix Q, key matrix K, value matrix V, and calculate the intermediate steganographic video unit : ， The spatial-temporal domain correction is performed through the cross-attention mechanism. Perform 3D convolution to generate the key matrix Sum Matrix , and based on the intermediate steganographic video unit Generate query matrix , and calculate the corrected steganographic video : Where, is the dimension of the key vector, and T is the matrix transpose.
6. The video encryption steganography method according to claim 1, wherein: In step S4, the reversibility of the diffusion model is achieved through DDIM reverse diffusion, which specifically includes: Based on noise predictor Construct the denoising function: Where, , is the preset noise scheduling coefficient, t is the diffusion step index and ; Solve for the initial state via an ordinary differential equation: Where, To encrypt video frames, is the diffusion state at the Tth step, and T is the total number of diffusion steps.
7. The video encryption steganography method according to claim 1, wherein: In step S5, the RSA decryption algorithm formula is expressed as: Where, c is the encrypted information, m is the target information, To decrypt the index, To generate random large prime numbers and The modulus of .
8. The video encryption steganography method according to claim 1, wherein: In the step S2, before adding Gaussian noise, an information capacity optimization step is also included: dynamically adjusting the amount of embedded encrypted information under a positive correlation relationship based on the texture complexity of the original video.
9. The video encryption steganography method according to claim 1, wherein: In the S2 step, Gaussian noise is added obey , is a multidimensional Gaussian distribution, is the noise variance scalar value, is the identity matrix.