Video-oriented steganography implementation method

Abstract

The invention discloses a video-oriented steganography implementation method, which comprises the following steps of: based on a deterministic video diffusion model, acquiring an intermediate potential variable after de-noising processing of executing a final time step in a potential space, based on the intermediate potential variable, utilizing a classifier-free guide CFG scale parameter to modulate and generate a standard potential variable and two reference potential variables; and based on the reference potential variable and the two reference potential variables, determining a position area in which a secret message can be embedded in a video by using a multi-mask mechanism, so that a secret message to be written generates a secret-containing video based on the determined position area in which the secret message can be embedded, and video-oriented steganography operation is realized. According to the method, an exclusive steganography channel can be constructed by modulating implicit condition parameters in the diffusion process, a multi-mask mechanism is innovatively introduced to screen a robust region in a potential space, and a key technical guarantee is provided for construction of secure covert communication.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for implementing steganography in video. Background Technology

[0002] In the digital information age, video, due to its high complexity and large data volume, has become an ideal medium for steganography. In other words, with the development of generative artificial intelligence, video generation technologies based on diffusion models (such as Sora, HunyuanVideo, and Wan2.1) have become the mainstream paradigm. According to Zebracat AI statistics, AI-generated videos account for 40% of video content on mainstream social media platforms, providing new application scenarios for steganography. Currently, steganography can be mainly implemented through modification-based steganography and generative steganography.

[0003] Modification steganography typically involves directly modifying the content or encoded elements of the video carrier. For example, common video steganography methods are mainly categorized into three types: spatial domain, transform domain, and compression domain, all of which are modification-based steganography methods. Specifically, spatial domain methods embed information by directly modifying the pixel values ​​of video frames; transform domain methods embed data in the frequency domain using discrete wavelet transform (DWT) and discrete cosine transform (DCT); and compression domain methods modify DCT coefficients, motion vectors, and other elements during video compression. In other words, these methods all require direct modification of the carrier content, making them prone to distortion and drift problems. The resulting distortion accumulates and spreads throughout the frame sequence, leading to a decline in visual quality. Furthermore, the modification operations also destroy the original statistical properties of the encoded elements, making them highly susceptible to detection by deep learning-based steganalysis techniques. This makes them difficult to withstand advanced steganalysis detection techniques, thus compromising their security to meet the needs of complex scenarios.

[0004] Generative steganography, by implicitly embedding secret information during data generation, avoids direct modification of the data carrier and possesses stronger resistance to steganalysis. However, existing research on generative steganography primarily focuses on the image domain. While it can implicitly embed secret information during image generation to avoid direct modification of the data carrier, the significant difference in complexity between images and videos prevents its application in video scenarios. Specifically, extending existing generative steganography to video carriers faces the following insurmountable technical bottlenecks:

[0005] (1) In order to improve reasoning efficiency and content controllability, mainstream video generation models generally adopt deterministic sampling strategies, while existing generative steganography methods rely on random noise sampling, so they cannot be compatible with deterministic diffusion models.

[0006] (2) Social media platforms usually compress videos to save bandwidth. Information loss during compression can also lead to damage to steganographic content. In other words, the robustness of video compression distortion cannot be achieved, and it is difficult to guarantee the reliable extraction of secret information.

[0007] In view of this, the present invention is hereby proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a video steganography implementation method, thereby providing a video steganography scheme with high robustness and high security, and solving the above-mentioned technical problems existing in the prior art.

[0009] The objective of this invention is achieved through the following technical solution:

[0010] A video-oriented steganography implementation method includes:

[0011] Based on a deterministic video diffusion model, intermediate latent variables are obtained after denoising processing in the latent space by performing the final time step. Based on these intermediate latent variables, classifier-free CFG scale parameter modulation is used to generate a baseline latent variable and two reference latent variables. The video generated after decoding by the baseline latent variable and the two reference latent variables is visually identical but has identifiable differences in encoding. The two reference latent variables are used as reference pairs for encoding binary information 0 and 1.

[0012] Based on the aforementioned one baseline latent variable and two reference latent variables, a multi-mask mechanism is used to determine the location region in the video where a secret message can be embedded, and the secret message to be written is used to generate a first secret latent variable based on the determined location region where the secret message can be embedded.

[0013] By using the first hidden latent variables, a hidden video is generated through the decoder of the deterministic video diffusion model, thus realizing video-oriented steganography.

[0014] The CFG scale parameters include three different CFG scale parameters used as weighting parameters. These weighting parameters are parameters applied in the calculation of noise prediction using a deterministic video diffusion model. The three different CFG scale parameters include standard CFG scale parameters and first and second CFG scale parameters. The process of generating one baseline latent variable and two reference latent variables includes:

[0015] In the final time step of performing denoising in the latent space, a baseline latent variable is generated using the standard CFG scale parameters, and two reference latent variables are generated using the first and second CFG scale parameters, respectively.

[0016] The process of determining the location region in the video where a secret message can be embedded using a multi-mask mechanism includes:

[0017] An invariant mask is generated based on the standard CFG scale parameters, indicating the location region least sensitive to encoding / decoding and noise; and a discriminative mask is generated based on the two reference latent variables, indicating the location region most easily decoded correctly by the receiver.

[0018] A combined mask is generated based on the invariant mask and the distinguishable mask, and the combined mask indicates the location region where the secret message can be embedded.

[0019] The process of generating an invariant mask based on standard CFG scale parameters includes:

[0020] The baseline latent variable is generated using the standard CFG scale parameters, and the video obtained by decoding the baseline latent variable is encoded back into the latent space to obtain the first latent variable;

[0021] Calculate the distance between the baseline latent variable and the first latent variable at each location, and select the location with the smallest predetermined distance as 1, and mark the other locations as 0, to form the invariance mask.

[0022] The process of generating a discriminative mask based on the two reference latent variables includes:

[0023] Calculate the distance between the two reference latent variables at each location, and select the location with the largest predetermined distance as 1, and the remaining locations as 0, to form the distinguishability mask.

[0024] The process of generating a combined mask based on the invariant mask and the distinguishable mask includes:

[0025] The invariant mask and the distinguishable mask are subjected to a point product operation, and the result is used as the combined mask. The positions with a value of 1 in the combined mask are used as the positions where secret messages can be embedded.

[0026] The process of generating the first dense latent variable includes:

[0027] For the location region in the combined mask where a secret message can be embedded, for message bit 0 in the binary message composed of the secret message to be written, the first reference latent variable of the two reference latent variables is written at the corresponding position; for message bit 1 in the binary message, the second reference latent variable of the two reference latent variables is written at the corresponding position.

[0028] For the location regions in the combined mask that cannot embed secret messages, the corresponding locations are written into the baseline latent variables.

[0029] At the receiver of the encrypted video, the method further includes:

[0030] The receiver encodes the received video to obtain the second latent variable;

[0031] The secret message in the video is extracted based on the first reference latent variable, the second reference latent variable, and the combined mask generated by the receiver itself.

[0032] The process of extracting the secret message includes:

[0033] For each position marked as 1 in the combined mask, calculate the distance between the value of the second dense latent variable at that position and the corresponding values ​​of the first reference latent variable and the second reference latent variable, and:

[0034] If the second hidden latent variable is closer to the first reference latent variable, then the extracted message bits are 0;

[0035] If the second hidden latent variable is closer to the second reference latent variable, then the extracted message bit is 1;

[0036] After traversing all locations, the complete secret message is obtained.

[0037] Compared with existing technologies, the present invention provides a video steganography implementation method, namely, a training-free robust video steganography method based on a deterministic latent diffusion model, abbreviated as LD-RoViS, thus providing a training-free robust video steganography framework. It not only achieves high-capacity and high-precision embedding of secret information within a deterministic latent diffusion model, but also significantly improves robustness and concealment under distortion conditions such as compression and noise through a multi-mask mechanism, thereby possessing broad application prospects. In summary, the steganography implementation method provided by the present invention has the advantages of requiring no additional training and possessing high capacity, high robustness, and high security. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 A general overview diagram of the LD-RoViS framework provided for embodiments of the present invention;

[0040] Figure 2 A detailed flowchart of the LD-RoViS framework provided in the embodiments of the present invention;

[0041] Figure 3 This is a schematic diagram of the LD-RoViS algorithm implementation provided in an embodiment of the present invention;

[0042] Figure 4 This is a visual comparison diagram of the embodiments of the present invention with other baseline methods. Detailed Implementation

[0043] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the specific content of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments, which do not constitute a limitation of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0044] First, the following explanations are provided for the terms that may be used in this article:

[0045] The term "and / or" means that either or both can be achieved simultaneously. For example, X and / or Y means that it includes both "X" or "Y" as well as the three cases of "X and Y".

[0046] The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0047] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0048] Unless otherwise explicitly specified or limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this document according to the specific circumstances.

[0049] When concentration, temperature, pressure, size, or other parameters are expressed as numerical ranges, such ranges should be understood to specifically disclose all ranges formed by any pairing of upper limits, lower limits, or preferred values ​​within that range, regardless of whether the range is explicitly stated; for example, if the numerical range "2 to 8" is stated, then that range should be interpreted to include ranges such as "2 to 7", "2 to 6", "5 to 7", "3 to 4 and 6 to 7", "3 to 5 and 7", "2 and 5 to 7", etc. Unless otherwise stated, the numerical ranges described herein include both their endpoints and all integers and fractions within that range.

[0050] The terms “center,” “longitudinal,” “lateral,” “length,” “width,” “thickness,” “upper,” “lower,” “front,” “back,” “left,” “right,” “vertical,” “horizontal,” “top,” “bottom,” “inner,” “outer,” “clockwise,” and “counterclockwise” indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience and simplification of description and do not imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this document.

[0051] The video steganography implementation method provided in this invention mainly relates to the interdisciplinary fields of digital content security and generative artificial intelligence. Specifically, it is a training-free robust video steganography method for deterministic potential diffusion video generation models, which can be applied to scenarios where secret information needs to be transmitted covertly through video carriers.

[0052] Specifically, the implementation of this invention addresses the problems of weak compression resistance, susceptibility to steganalysis and detection, and the need for additional training in existing video steganography techniques. It proposes a training-free robust video steganography framework for deterministic latent diffusion models. This framework constructs a dedicated steganography channel by modulating implicit conditional parameters during the diffusion process and innovatively introduces a multi-mask mechanism to filter robust regions in the latent space, effectively solving the compression distortion problem and providing key technical support for the construction of secure and covert communication.

[0053] To address the problems existing in the prior art, this invention provides a video steganography implementation method. This method is a training-free robust video steganography approach based on a deterministic latent diffusion model. This scheme constructs a steganography channel through implicit conditional parameter modulation during the diffusion process and introduces a multi-mask mechanism to enhance resistance to video compression and post-processing. It can provide a highly secure and robust protection scheme for covert video communication over the internet. Experimental verification shows that the technical solution provided by this invention can embed approximately 12,000 bits of secret information into a 5-second video segment with an extraction accuracy exceeding 99%, and exhibits high visual quality and strong concealment.

[0054] In the specific implementation process, the embodiments of the present invention mainly consider using at least the following technical means, including:

[0055] (1) By utilizing the deterministic sampling characteristics of the diffusion model, the classifier-independent guidance (CFG) parameters (i.e. implicit conditional parameters) are modulated at the final time step of the diffusion process to construct a steganalysis channel in the latent space, embedding secret information into the implicit representation of the generated video without retraining the video generation model.

[0056] Specifically, the "classifier-free guidance (CFG)" mechanism in the generation process of a deterministic latent diffusion video model can be utilized by using three different CFG scale parameters in the final step of the denoising process. ), from the same intermediate latent variables Generate three semantically highly consistent but subtly different latent variable outputs. ),in and These form a reference pair used to encode binary information "0" and "1".

[0057] (2) Before embedding, noise is added by encoding and decoding to simulate the compression distortion of social platform videos, calculate the robust mask and discriminative mask of the latent representation, and combine the two to obtain a multi-mask mechanism, automatically select robust and discriminative positions to enhance the stability of hidden information in transmission and compression;

[0058] In order to enhance the resistance to channel distortions such as video compression and noise, a novel multi-mask mechanism was designed and implemented to identify the most robust regions in the potential space that are best suited for embedding information.

[0059] (3) Based on the embedding positions obtained from multiple masks, the secret information (or secret message) is constructed into a steganography matrix, and the three CFG parameters (i.e., standard CFG scale parameters) are applied. and the first and second CFG scale parameters The generated latent variables (i.e., one baseline latent variable and two reference latent variables) are mixed to obtain a latent representation carrying the steganographic message, and then decoded to generate a steganographic video; the receiving end reproduces the latent variables using the same cue words, seeds and parameters, and can accurately extract the secret information by comparing the distances;

[0060] Specifically, on the sender side, based on the combined mask The position is selectively determined based on the binary message bits (0 or 1) to be embedded. or The values ​​are filled into the corresponding latent variables. In this process, the final dense latent variables are formed. And decoded into a encrypted video. On the receiving end, a shared key is used (text prompt, random seed, ...). Regenerate baseline latent variables , and mask The received video is decoded back into the latent space by comparing it with the mask location. or To accurately extract secret information from a distance.

[0061] In summary, the implementation of this invention provides a training-free robust video steganography framework, which can not only achieve high-capacity and high-precision embedding of secret information in deterministic latent diffusion models, but also significantly improve robustness and concealment under distortion conditions such as compression and noise through a multi-mask mechanism, thus giving it broad application prospects.

[0062] To facilitate understanding of the embodiments of the present invention, the specific implementation methods will be described in detail below with reference to the accompanying drawings.

[0063] Reference Figure 1 The LD-RoViS framework provided by this invention consists of a sender and a receiver. Both parties share the same set of parameters, including: a text prompt, a random seed, and three classifier-free guided (CFG) scale values. The corresponding implementation processes on the sender and receiver sides may include:

[0064] (a) Implementation process on the sending side

[0065] Use prompt, seed, and standard CFG scales. Drive a deterministic video diffusion model, perform T-1 steps of denoising in the latent space, and obtain intermediate latent variables.

[0066] In the final denoising step, corresponding parameter modulation and multi-mask mechanisms are also utilized, based on the first secret message to be embedded. This generates the final dense latent variables. ;

[0067] The dense latent variables The first encrypted video is generated through the model's decoder (D). ;

[0068] The first encrypted video The second encrypted video is sent to the recipient via public channels such as social networks. The recipient then receives the encrypted video, which is denoted as... .

[0069] (ii) Implementation process on the receiving side

[0070] Using the shared prompt, seed, and CFG scale parameters, perform the same operations as the sender to regenerate the latent variables and masks required for the decoding baseline;

[0071] The received second encrypted video Transform back into the second latent variable using the model's encoder (E). ;

[0072] Using the second latent variable The second secret message is extracted by comparing it with the regenerated baseline within the area specified by the mask. .

[0073] It should be noted that the aforementioned first confidential video The second confidential video is recorded as The description does not imply that the two encrypted videos are different; the description here is only to distinguish between the encrypted video sent by the sender and the encrypted video received by the receiver. Similarly, the corresponding encrypted latent variables... Second latent variable and First Secret Message Second Secret Message The same situation applies.

[0074] The following will refer to Figure 2 The provided LD-RoViS framework flowchart details the three main processing parts—preprocessing, sending, and receiving—during the implementation of this invention.

[0075] (a) Pretreatment process

[0076] The corresponding preprocessing process consists of a steganography channel construction process based on parameter modulation and a robust region filtering process based on a multi-mask mechanism. Both processes need to be executed at both the sender and receiver in order to write and read the encrypted video.

[0077] (1) Steg channel construction process based on implicit conditional parameter modulation

[0078] The core of this process lies in modulating the trajectory of the diffusion process using CFG scale parameters, which can be specifically implemented as follows:

[0079] In each step of denoising in the latent space, the model predicts the noise. It consists of a weighted average of conditional and unconditional predictions, with the CFG scale serving as the weighting coefficient; the corresponding formula for calculating the noise in the model predictions can be:

[0080] ;

[0081] in, For the model at time step Predicted noise, This is referred to as "conditional noise" (guided by prompt and related to video semantics). This is "unconditional noise" (not guided by prompt). The scale controls the weight ratio of the two. For the model at time step The latent variables obtained;

[0082] At the final time step (i.e. from) arrive (Steps), by setting three different Scale parameters These are the standard CFG scale parameters. and the first and second CFG scale parameters This is used to generate three semantically consistent but different latent variables, and the corresponding calculation formula can be:

[0083] (in , ∈ );

[0084] In the formula, G is a deterministic latent diffusion video model, and Diffuse (・) represents the denoising process function. The intermediate latent variable after the first T-1 steps of denoising; As a baseline latent (i.e., a baseline latent variable). As reference latent variables (i.e., two reference latent variables, hereinafter referred to as the first and second reference latent variables), these three variables form the basis of the secret information embedding; after decoding, these three latent variables generate visually almost indistinguishable videos (which can be understood as visually identical videos), but they have subtle differences in the latent space that can be used for encoding (i.e., there are identifiable differences); among them, and This forms a "discrimination pair", which are used to represent binary information "0" and "1" respectively;

[0085] Specifically, the process of generating the three latent variables mentioned above includes:

[0086] Using the standard CFG scale parameters Generate the corresponding baseline latent variables ,Right now:

[0087] ;

[0088] Using the first and second CFG scale parameters Generate two reference latent variables and ,Right now:

[0089] ;

[0090] .

[0091] (2) Robust region selection process based on multi-mask mechanism

[0092] This processing is primarily designed to combat video distortion caused by compression and other operations during transmission. The corresponding process, based on a multi-mask mechanism to locate the most robust embedding region, may include:

[0093] (21) Invariant mask Generation:

[0094] First, channel distortion is simulated through a "preprocessing" step to apply standard CFG scale parameters. Generated baseline latent variables Decode it into video, then apply analog compression or noise reduction, and finally encode it back into the latent space to obtain... ;

[0095] Then, calculate and At each position Distance, and select the one with the smallest distance. (This parameter value can be preset or set as needed, such as presetting the corresponding minimum amount.) The position marked is 1, and the other positions are marked 0, forming an invariant mask. The corresponding positions represent the areas least sensitive to encoding / decoding and noise.

[0096] (22) Distinguishing mask Generation: Calculate two reference latent variables as discriminant pairs. and At each position Distance, and select the one with the largest distance. (This parameter value can be preset or set as needed, such as presetting the corresponding maximum predetermined amount.) The position marked is 1, and the other positions are marked as 0, forming a distinguishability mask. It indicates the location area that is most easily decoded correctly by the receiver. That is, the corresponding location represents the area where the difference is most significant after embedding "0" and "1" and is most easily decoded correctly by the receiver.

[0097] (23) Combined mask Generation: and Perform a bitwise product (logical AND) operation to obtain the final combined mask. It indicates the location area where the secret message can be embedded, that is, only when... Only positions with a median value of 1 are used to embed secret information.

[0098] (ii) The sender's transmission process

[0099] The sender of the video needs to perform the corresponding video information embedding process, which mainly involves combining masks. The determined embedding location, and the binary message of the first secret message to be embedded. And can construct dense latent variables according to the following rules. The corresponding rules may include:

[0100] If a certain location is If the value in the middle is 0 (non-embedded region), then The value at that position is equal to the baseline latent variable. The corresponding value;

[0101] If a certain location is The middle part is 1 (embedding region), and the corresponding message bit ,but The value at that position is equal to the first reference latent variable. The corresponding value;

[0102] If a certain location is The middle part is 1 (embedding region), and the corresponding message bit ,but The value at that position is equal to the second reference latent variable. The corresponding value;

[0103] Constructing dense latent variables Then, the constructed dense latent variables can be... Input to decoder to generate the corresponding first encrypted video. .

[0104] (III) Receiving and processing procedures of the recipient

[0105] The receiving party of the video needs to perform the corresponding information extraction and processing. In this process, the receiving party first uses the shared parameters to fully execute the above preprocessing process and regenerate the discrimination pair. , and combined masks Then, the received second encrypted video... Encoded as a second dense latent variable For For each position where the bid is 1, the receiver calculates the second hidden latent variable. The value at that position and , Corresponding value Calculate the distance and perform the following processing:

[0106] if and If the distance is closer, the extracted message bits will be 0;

[0107] if and If the distance is closer, the extracted message bit is 1; after traversing all embedding positions, the complete second secret message extracted by the receiver can be recovered. The second secret message With the first secret message The same information was obtained, thus completing the extraction of the secret message.

[0108] In the implementation of the above embodiments of the present invention, the corresponding message embedding and extraction algorithms can be implemented as follows: Figure 3 The method shown is used to implement this, and the specific implementation process is described above and will not be repeated here.

[0109] To further verify the effectiveness of the technical solutions provided in the embodiments of the present invention, the following corresponding experimental procedures were also explained.

[0110] I. Experimental Environment and Parameter Settings

[0111] Basic model selection: The T2V-1.3B model in the Wan2.1 video generation model is adopted as the basic framework. This model is based on the deterministic sampling strategy of Flow Matching, which can efficiently generate high-quality videos, has low memory requirements, supports single GPU operation, and facilitates technology implementation.

[0112] Dataset configuration: 100 prompts were randomly selected from the VidProM text-video cue dataset (a large-scale, diverse text-video pairing dataset). One video sample was generated for each prompt. The video parameters were uniformly set as follows: duration 5 seconds, resolution 480×832, frame rate 16fps, and each video contained 81 frames to ensure the diversity and consistency of the experimental samples.

[0113] Core parameter settings: random seed = 99 (to ensure reproducibility of the experiment), baseline Scale parameters =5.0, mask threshold , hyperparameters ;

[0114] Hardware and software environment: The hardware uses 4 NVIDIA RTX A6000 GPUs (48GB of video memory per card), and the software relies on the PyTorch deep learning framework. The video encoding / decoding tool uses FFmpeg to ensure the efficient operation of the experiment.

[0115] II. Performance Evaluation Indicators

[0116] To comprehensively evaluate the performance of the technical solutions provided in the embodiments of the present invention, the following five core indicators were specifically adopted:

[0117] Extraction accuracy (acc, %): The proportion of correctly extracted secret bits to the total number of embedded bits, reflecting the method's robustness against distortion; the higher the value, the better.

[0118] Peak Signal-to-Noise Ratio (PSNR, dB): The frame-average PSNR of the encrypted video and the original video (the reference video without embedded secret information), reflecting the degree of visual distortion. The higher the value, the better the visual quality.

[0119] BRISQUE (Blind Vision Assessment No-Reference Image Spatial Quality Evaluator) Score: A no-reference image quality assessment metric that evaluates the visual quality of loaded video without requiring the original video. The lower the score, the better the visual quality.

[0120] Embedding capacity (bits): The total number of bits of secret information that can be embedded in a single video, reflecting the information transmission capability of the method; the higher the value, the better.

[0121] Detection error rate ( , %): The false positive rate of the steganalysis model for "secret video" and "original video" reflects the method's anti-detection security. The closer the value is to 50% (random guessing level), the better.

[0122] III. Comparative Experiments and Result Analysis

[0123] 1. Comprehensive performance comparison experiment

[0124] Three mainstream video steganography methods were selected as the benchmark for comparison: the traditional transform domain method AQIM, MEC_AQIM, and the generative method RoGVSN. The corresponding experimental results are shown in Table 1 (the results are the mean ± standard deviation of 100 video samples).

[0125] Table 1

[0126]

[0127] As shown in Table 1 above, the technical solution provided by the embodiments of the present invention is significantly superior to the comparison methods in terms of visual quality, with a PSNR of 41.66dB and a BRISQUE score of 28.90, both of which are better than the traditional transform domain method and the generative method RoGVSN. The extraction accuracy remains at a high level of over 99%, comparable to AQIM and RoGVSN, and far superior to MEC_AQIM. The embedding capacity reaches 11983 bits, far exceeding RoGVSN (729 bits), and higher than the fixed capacity of 10000 bits of the traditional method, achieving a balance of "high visual quality, high robustness, and high capacity".

[0128] Visual comparison diagram of the encrypted video and the original video as shown below. Figure 4 As shown in the figure, there are four sets of comparisons: AQIM, MEC_AQIM, RoGVSN, and the technical solution provided in this embodiment of the invention. The top row of each set is the middle frame (frame 41) of the encrypted video, and the bottom row is the pixel difference map between the encrypted frame and the original frame (brightness magnified 10 times to highlight the difference). It can be observed that the difference maps of traditional methods (AQIM, MEC_AQIM) have obvious block effects, RoGVSN has blurring distortion in the facial area, while the difference map of the technical solution provided in this embodiment of the invention only has slight differences in areas with complex textures, with minimal visual distortion.

[0129] 2. Steganography Resistance Security Experiment

[0130] Security was evaluated using three mainstream steganalysis models: the video steganalysis feature extractor SUPERB (combined with a linear classifier), and the image steganalysis models CovNet and LWENet. The experiment used 100 pairs of "original video-secretly-loaded video" samples, decoding to obtain 8100 pairs of frame samples, which were then divided into training, validation, and test sets at a ratio of 4000:600:3500. The experimental results of the corresponding steganalysis error rates are shown in Table 2. To detect the error rate, the higher the value, the better.

[0131] Table 2

[0132]

[0133] As shown in Table 2 above, traditional methods AQIM, MEC_AQIM, and generative methods RoGVSN are easily detected by steganalysis models, and CovNet and LWENet have detection error rates close to 0. However, the technical solution provided in this embodiment of the invention has a detection error rate of close to 50% under the three steganalysis models, reaching the level of random guessing, proving that it has extremely strong anti-detection security and fundamentally avoids the detection risk caused by directly modifying the carrier.

[0134] 3. Compression and noise robustness test

[0135] The extraction accuracy under different distortion scenarios were simulated using H.264 video compression (CRF=18, 23, 27, the higher the CRF value, the higher the compression rate and the greater the distortion), Gaussian noise (standard deviation 0.05), salt and pepper noise (probability 0.01), and brightness adjustment (HSV spatial brightness increased by 0.1). The experimental results of the accuracy under different distortions are shown in Table 3 below.

[0136] Table 3

[0137]

[0138] As shown in Table 3 above, the technical solution provided by the embodiments of the present invention maintains a high extraction accuracy under various distortion scenarios—the accuracy still reaches 91.67% under H.264 compression (CRF=27) and 92.82% under Gaussian noise. In particular, it has excellent robustness to salt-and-pepper noise and brightness adjustment, with accuracy rates of 98.72% and 99.02%, respectively. Although the accuracy is slightly lower than RoGVSN under high compression rates (CRF=23, 27), the embedding capacity of RoGVSN is only 729 bits, which is far lower than the 12,000-bit capacity of the technical solution provided by the embodiments of the present invention. It can be seen that the technical solution provided by the embodiments of the present invention performs better in the balance between "robustness and capacity".

[0139] 4. Multi-mask ablation experiment

[0140] To verify the multi-mask mechanism (Invariant Mask) and The necessity of (discriminative mask) is emphasized, and three variant models are designed for comparison: Variant 1 (using only...) ,none Variant 2 (only) ,none Variant 3 (Complete Model) + The experimental results comparing the performance of the different variants are shown in Table 4 below.

[0141] Table 4

[0142]

[0143] As shown in Table 4 above: using only At that time, due to the lack of anti-distortion regions, the extraction accuracy plummeted to 62.67%; using only At that time, due to the lack of distinguishable regions, the accuracy dropped to 75.46%; while the complete model combined with and Afterwards, the accuracy recovered to 99.17%, while maintaining high visual quality and reasonable capacity, proving that the synergistic effect of the multi-mask mechanism is the key to ensuring robustness.

[0144] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0145] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A video-oriented steganography implementation method, characterized by, include: Based on a deterministic video diffusion model, intermediate latent variables are obtained after denoising processing in the latent space by performing the final time step. Based on these intermediate latent variables, classifier-free CFG scale parameter modulation is used to generate a baseline latent variable and two reference latent variables. The video generated after decoding by the baseline latent variable and the two reference latent variables is visually identical but has identifiable differences in encoding. The two reference latent variables are used as reference pairs for encoding binary information 0 and 1. Based on the aforementioned one baseline latent variable and two reference latent variables, a multi-mask mechanism is used to determine the location region in the video where a secret message can be embedded, and the secret message to be written is used to generate a first secret latent variable based on the determined location region where the secret message can be embedded. By using the first hidden latent variables, a hidden video is generated through the decoder of the deterministic video diffusion model, thus realizing video-oriented steganography.

2. The method of claim 1, wherein, The CFG scale parameters include three different CFG scale parameters as weight parameters. The weight parameters are parameters used in the calculation of noise prediction by the deterministic video diffusion model. The three different CFG scale parameters include the standard CFG scale parameters and the first and second CFG scale parameters. The process of generating one baseline latent variable and two reference latent variables includes: In the final time step of performing denoising in the latent space, a baseline latent variable is generated using the standard CFG scale parameters, and two reference latent variables are generated using the first and second CFG scale parameters, respectively.

3. The method of claim 2, wherein, The process of determining the location region in the video where a secret message can be embedded using a multi-mask mechanism includes: An invariant mask is generated based on the standard CFG scale parameters, indicating the location region least sensitive to encoding / decoding and noise; and a discriminative mask is generated based on the two reference latent variables, indicating the location region most easily decoded correctly by the receiver. A combined mask is generated based on the invariant mask and the distinguishable mask, and the combined mask indicates the location region where the secret message can be embedded.

4. The method of claim 3, wherein, The process of generating an invariant mask based on standard CFG scale parameters includes: The baseline latent variable is generated using the standard CFG scale parameters, and the video obtained by decoding the baseline latent variable is encoded back into the latent space to obtain the first latent variable; Calculate the distance between the baseline latent variable and the first latent variable at each location, and select the location with the smallest predetermined distance as 1, and mark the other locations as 0, to form the invariance mask.

5. The method of claim 4, wherein, The process of generating a discriminative mask based on the two reference latent variables includes: Calculate the distance between the two reference latent variables at each location, and select the location with the largest predetermined distance as 1, and the remaining locations as 0, to form the distinguishability mask.

6. The method of claim 5, wherein, The process of generating a combined mask based on the invariant mask and the distinguishable mask includes: The invariability mask and the distinguishability mask are subjected to a bitwise point product operation, and the result is taken as the combined mask, and the position region with a value of 1 in the combined mask is taken as the position region in which the secret message can be embedded.

7. The method according to any one of claims 1 to 6, characterized in that, The process of generating the first secret-containing latent variable comprises: For the position region in which the secret message can be embedded in the combined mask, for message bit 0 in the binary message constituted by the secret message to be written, the first reference latent variable of the two reference latent variables is written in the corresponding position, and for message bit 1 in the binary message, the second reference latent variable of the two reference latent variables is written in the corresponding position; For the position region in which the secret message cannot be embedded in the combined mask, the reference latent variable is written in the corresponding position.

8. The method of claim 7, wherein, At the receiving end of the secret-containing video, the method further comprises: The receiving end encodes the received video to obtain a second latent variable; The secret message in the video is extracted based on the first reference latent variable, the second reference latent variable and the combined mask generated by the receiving end itself.

9. The method of claim 8, wherein, The extraction process of the secret message comprises: For each position marked as 1 in the combined mask, the distance between the value of the second secret-containing latent variable at the position and the corresponding values of the first reference latent variable and the second reference latent variable is calculated respectively, and: If the distance between the second secret-containing latent variable and the first reference latent variable is closer, the extracted message bit is 0; If the distance between the second secret-containing latent variable and the second reference latent variable is closer, the extracted message bit is 1; After traversing all positions, the complete secret message is obtained.

Images